Using Infrared Imaging to Monitor and Control Welding

This invention relates to a method and apparatus for welding and/or control of welding in real time by analysis of the weld being made.

The present invention is directed to welding various materials and will be described hereinafter in conjunction with the welding of metals. It is directed to welding systems such as gas tungsten arc welding (GTAW), gas metal arc welding (GMAW), fluxed cored arc welding (FCAW), plasma arc welding (PAW), and resistance or spot welding.

The analysis, in real time, of the weld puddle has been attempted with closed circuit television, but the light emitted at the weld is extremely bright, and for various metals there may be insufficient contrast to obtain good definition of the weld puddle. Further, the video scanner must work in a hostile environment where there are sparks and spattering metal. Dirt, dust, and foreign materials may also be present, as well as stray light. Although a video system may be useful in viewing the weld puddle geometry, it still lacks the ability to determine puddle depth and quality, and to provide temperature variation data except through an exterior color analysis of the puddle.

A video system for weld puddle analysis is shown in U.S. Pat. No. 4,595,820 which discloses that video data indicative of some weld conditions can be generated by a video camera directly observing welding operations. According to the technique disclosed in this patent, the width of the weld puddle is monitored by observing local minima of light intensity associated with the edges of the weld puddle. However, this technique or approach is not able to respond effectively to trends of weld speed, current, or resulting puddle geometrics. Further, this approach is not able to recognize conditions below the surface of the weld puddle. This is clear because under video observation, reflected external light is received from the surface of the weld puddle.

Present commercially available adaptively controlled systems that use weld puddle analysis are very limited not only in analysis of the weld, but in their ability to operate in real time at commercially desired speeds, e.g., of nine to twenty-five inches per minute. Typically, current adaptively controlled welding systems using weld puddle analysis operate at a speed of about two to four inches per minute. Current systems further analyze only a few of the variables involved in influencing weld puddle geometry including joint geometry, welding process parameters, chemistry of the materials, and internal defects. In arc welding, the number of process and procedure variables involved in an adaptively controlled welding system becomes particularly large and includes such process variables as heat input, filler wire input, current flow characteristics, shielding gas flow, power supplied (a function of amperage and voltage), material, part position, torch operation, torch position, and torch travel speed. With so many variables, it is little wonder that present day adaptively controlled welders operating at commercial speeds are not able to effectively make an analysis of such a multitude of variables but instead, on occasion, rely upon no more than simple timers and electro-mechanical control arrangements, and monitor simply such parameters as amperage, voltage, and shielding gas temperature. Even the aforesaid patented system analyzes only a few of these processes and procedure parameters or variables, and, nonetheless, operates at a relatively slow weld speed. While laser and video systems for seam tracking are available, they, too, process but a limited amount of information about the seam and the track being followed. These prior art systems also do not, in real time, effectively analyze and assure proper alignment of the torch and joint, and correct effectively for deviations therefrom where the parts are mismatched in the vertical or horizontal directions, or correct for variations in the penetration of the puddle or for the joint root opening between the parts. Further, it would be desirable in real time to be able to make weld size measurements and to correlate them in real time with the mechanical properties of the weld such as tensile and shear strength. In addition to controlling the welding process, it is desirable to have a real time weld inspection system that could determine its physical properties and strength properties. Also, it would be desirable to have a welding system which records and stores information regarding all welding events for later use in the analysis of failed weld joints, analysis of welding consumables used, and welding process conditions encountered when completing a particular weld.

The achievement of a welding system having many or most of the above described features should reduce defects in welds and control weld inspection costs, as well as improve productivity. Such a weld control and weld inspection system requires considerable computing technology. But such systems traditionally require large amounts of computing power and tend to be too slow for real time process control. Accordingly, an object of the present invention is to promote the practical application of artificial intelligence technology to welding process control.

A more general object is to provide a new and improved welding apparatus or method, and to promote the development of adaptively controlled welding techniques, including welding techniques effective for producing high quality welds because of their ability to monitor and control critical processes and procedural variables in real time.

It is an object of the invention herein to adaptively conduct resistance welding, and particularly adaptive resistance spot welding. Adaptive welding techniques generally ensure the creation of better welds. While this does not eliminate the need to consider testing or inspecting the weld after its creation, it is certainly better to do it right the first time. In the case of resistance spot welds, as are particularly useful in the manufacture of automobiles, air frames, aircraft engines, and other sheet metal items, there is frequently no cost effective way to inspect the welds after the fact. Thus, they need to be made correctly the first time. Thus, adaptive control of the weld variables during welding is essential. More automatic production, without in-process inspection and responsive modification of the variables is inadequate.

It is a further object of the invention to more than merely perform seam tracking, but to conduct weld process control and real time weld inspection in the same arrangement for a plurality of weld variables and parameters. Further, it is an object to reduce defects and inspection costs in welding, to improve productivity, to monitor critical process and procedure variables, and to inspect weld defects in real time.

According to the invention, four independent processors are interconnected for serial communication to form a control system for an automated welding arrangement. Infrared images of the weld are captured by scanning with an infrared camera, by digitizing the information scanned, and by feeding the information into a first image processor and thereafter transferring the processed, compressed output of the first image processor into a second image processing expert system processor. The first processor locates the center of the weld puddle by determining the average center of brightness of the image.

Next, it determines a brightness level that corresponds to the puddle boundary, establishes the puddle boundary in view of said brightness level, and then measures the width of the puddle (perpendicular to the direction of travel) through the image center. Next, it measures, within predetermined peripheral zones of interest, including for example left, right, and front or tip zones of the puddle, the area of the weld puddle within each zone, the length of puddle perimeter within each zone, and the number of "zero crossings" or major reversals of the perimeter within the front or tip zone. The zones defined in the weld puddle for analysis include side zones straddling the center of the weld puddle, and showing the width of the weld puddle.

The tip zone of the weld puddle assists in the determination of weld puddle speed and position in real time, and the side zones permit effective puddle tracking and ascertainment of weld quality during operation. Further, the side zones provide information regarding the cooling gradient at the edge of the weld puddle, which is indicated by the sharpness of the gray scale change at the puddle edges, thereby establishing the rate of change in puddle brightness at the puddle boundary within the left and right side zones.

The information thus acquired is analyzed by the second image processing expert system processor, which then determines features relevant to the tracking of the weld puddle, and interprets selected geometric and metallurgical data characteristics of the weld puddle such as the speed and direction of the welding process, temperature, and whether the welding is proceeding normally or is in need of correction.

A third controller expert system processor further analyzes this information to produce information defining control strategies relating to power, speed, and direction. In particular, weld speed and weld current are controlled in response to the size of the weld puddle. For example, if the size of the puddle is too small, weld speed can be decreased or weld current increased. This information is fed into a fourth machine control expert system processor, which may contain special machine control programs possibly written in "C", and which actually controls the welding apparatus.

The detection of infrared information, according to the invention, is particularly useful in providing information not otherwise available because the heat gradient established during welding dissipates through the materials being welded and also through the air adjacent the weld puddle. Light reflected from the surface of the weld puddle, does not present the same wealth of interpretable heat gradient information.

According to the invention, most of the processors contain a semantic network of goals and forward and backward chaining rules as well as rule sets and data frames, which together permit data-driven, message-interruptible inference processing through the interplay of interrupting and scheduling in real time of pre-established ready, running, and suspended goal-driven inference processes. Further, the rule bases are hierarchically, transparently distributable among processors (hereinafter called "nodes").

The present invention further contemplates transparently distributing the process controller artificial intelligence components among a plurality of nodes, some of which may reside on independent hardware and all of which are able to communicate effectively with each other in real time, thereby achieving parallel processing with conventional hardware to whatever degree is necessary to achieve acceptable real time performance. This artificially intelligent controller arrangement can distribute inference processes multi-nodally. The knowledge base including, for example, frames, rules, and a semantic network representing the goals and subgoals of the process controller, and containing forward chaining rules that define processes which may be scheduled for execution, and backward chaining rules that drive the inferencing and communication processes, can thereby direct all real time operations. In particular, the semantic network permits data-driven, message-interruptible inference processing through the interplay of pre-established, scheduled processes organized into queues of ready, running, and suspended goal-driven processes whose execution is governed in real time by a time-sharing scheduler. The rule bases are thus hierarchically, as well as transparently distributed among the nodes for inference processing, and real time process control is achieved in an artificial intelligence environment.

Further, according to the invention, the process controller employs multiple software communications layers including a physical layer to communicate knowledge between nodes, a logical layer to identify the location of sought information, and a knowledge layer to communicate knowledge within a particular node. The logical layer evaluates requests for information in terms of local availability at a particular active node. If the information required for a particular node or process is missing locally, a communications sequence is initiated with the physical layer to search other nodes for the information of interest. In the preferred embodiment of the invention, the knowledge and logical layers are implemented using Lisp, and the physical layer is implemented using Pascal. The interface to the real world sensors and controllers is implemented in "C".

According to the invention, the queues are data structures that include pointers to Lisp stack groups which define inference processes. The Lisp scheduler maintains ready, running, and suspended queues of inference processes and controls the inference nominated by the running queue. Further, a context switcher and an inference scheduler (implemented respectively in Pascal and Lisp), invoke the Lisp scheduler with a command to change inference processes periodically and also in response to incoming messages from other nodes.

The Lisp scheduler has a semaphore flag which the context switcher can toggle. This flag, implemented through an interrupt, can direct the Lisp scheduler to take the running process, put it on the ready queue, and get the next process off the ready queue, or to take a process signalled by an incoming message and set it running. Thus, according to the invention, the context scheduler can inform Lisp that it is either to switch contexts, i.e., swap one of the ready processes with the running process, or that it is to process an incoming message. Accordingly, the high level Lisp scheduler can do one of two activities. Either it can swap processes, known as context switching, or it can acquire a message from the Pascal system for processing. The Pascal handles all communication with the outside world and also processes low level messages.

According to the invention, the controller arrangement further includes a consistency checker arrangement for establishing and evaluating new, proposed rules for introduction into the rule base, which ensures that the new rules will be consistent, non-redundant, and non-subsuming with regard to the established set of rules already integrated into the controller arrangement.

According to the invention, adaptive control of weld process variables is undertaken for all kinds of welds including fusion or arc welding, and resistance welding, for example. In short, the process variables are modified in response to thermographic measurements, which are based upon inspection of heat patterns or the geometry of the weld puddle and its associated heat affected zone. By thermographic measurement of the weld geometry, it is meant that the heat pattern actually observed is analyzed into isotherms, which can be established to correspond to predetermined heat levels. Further, the isotherms follow the contour of heat along each of the predetermined levels. Typically, the isotherms will be rounded, circular, or oval when measuring a proper fusion weld puddle between aligned, matched materials without a gap or excessive root opening therebetween.

This invention recognizes and takes advantage of the fact that the isotherms deviate from their normal rounded pattern and are characterized by interpretable protrusions extending from the previous, ideal rounded shape of the isotherms when there are errors or defects found in the weld puddle, which may be due to any of a number of circumstances. Such circumstances may include impurities or material problems in the weld itself. Further, they may indicate misalignment, mismatch or excessive root opening defects, as discussed herein.

The isotherms not only exhibit interpretable shape variations, according to the invention, but their very separation may be indicative of useful information. Both of these sources of information, shape and separation of isotherms, is considered interpretation of the weld puddle and heat affected zone geometry.

In the case of resistance spot welding, the weld puddle is frequently referred to as the weld nugget. Because the electrodes in resistance spot welding extend directly against the materials being spot welded, only the heat affected zone adjacent the electrodes, in the form of a smile, can be observed. This smile is called the weld nugget. For purposes of this invention, the weld puddle is considered to include its heat affected zone, and the term "weld nugget" is generally considered to have the same meaning as the weld puddle, as defind herein. The pattern of isotherms, in effect, constitutes the signature of the weld being performed. Of course, neither is directly observed, but takes on dimensional meaning and geometric shape in terms of the isotherms or heat patterns indicative thereof. These heat conditions are not visually determined herein, but only in terms of the information produced by an infrared detector system, as discussed below.

FIG. 1 is a schematic drawing of an artificial intelligence welding system, according to the invention herein;

FIGS. 2A-2C are illustrations of respective weld puddle isotherm information zones detected, and the contour of weld puddles under conditions of misalignment, mismatch, and excessive root opening;

FIGS. 3A and 3B show schematic details of the complete welder arrangement which permits actual implementation of adaptive welding in real time;

FIG. 4 shows, in schematic form, one version of the process controller, according to the invention herein;

FIG. 5 is a generalized block diagram of the process controller arrangement, according to the invention herein, comprising in particular a minimal process controller arrangement, which includes a single multi-inferencing expert (MIX) module responsive as shown to a sensor and effective for controlling a selected controllable external device;

FIG. 6 is a generalized block diagram of the MIX module including artificial intelligence elements, according to the invention herein;

FIG. 7 shows the process controller, according to the invention herein, including a MIX expert controller module, which is effective for communication with an external control element;

FIG. 8 is a block diagram of the process controller arrangement, according to the invention herein, comprising a network of MIX modules;

FIG. 9 is a block diagram of a generalized MIX module architecture, according to the invention herein, with the actor and advisor features explicitly set forth;

FIG. 10 is a block diagram of the scheduler arrangement for scheduling inference processes on respective running, ready, and suspended queues;

FIG. 11 is a block diagram of a number of inference processes typically held in the ready process queue when not actually running;

FIG. 12 is a flow chart of operation of the scheduler and arrangement, according to the invention herein;

FIG. 13 is a block diagram illustrating control of the knowledge bases, according to the invention herein, by corresponding knowledge mechanisms;

FIG. 14 shows schematically the relationship between particular knowledge bases and the associated knowledge mechanisms which direct their operation;

FIG. 15 illustrates the handling of messages according to the communication technique of the invention herein;

FIG. 16 illustrates the communications scheme of the invention;

FIG. 17 illustrates overall operation of the controller, according to the invention herein;

FIG. 18 shows a three-dimensional representation of heat in the weld puddle as produced by the image analyzer;

FIG. 19 is a block diagram illustrating the physical environment of the system, including the system-level hardware software design of the preferred embodiment of the invention; and

FIG. 20 is a schematic illustration of the control arrangement which emphasizes the multi-node approach, according to the invention herein.

According to one version, as shown in FIG. 1, the welding system 13, according to the invention, includes a radiometer 15, a controller 16 including an image analyzer 16(1), and an infrared (IR) detector 17 including, for example, a twelve-inch, close-up lens with a 10.6 micron wavelength upper range lens filter which produces clear images of the weld puddle 19 under observation.

The detector 17 and the welder arrangement 20 are carried in a controllably cooperative relationship by robot 22 to insure that the weld puddle 19 produced by welder 20 is in the field of view of the detector 17 The robot 22 can position and/or tilt the detector 17 in an (x,y) plane and about several axes of rotation, as is conventionally known in robotics technology.

The detector 17 can be adapted to operate for low range detection at 3.0 microns wavelengths as well as at 10.6 microns. This enables identification of puddle defects by observation of both the cooler and the hotter portions of the weld puddle at generally the same degree of resolution. Image analysis by the image analyzer 16(1) can be performed by image processing algorithms written in the "C" language, for example as discussed below, either on an IBM AT computer or a similar device. The image analyzer 16(1) can, accordingly, be implemented as an image processing computer directly interfaced with the infrared radiometer 15, and then, as will be discussed (with reference to FIG. 8), to an image processing advisor 115i.

The detector 17, the radiometer 15, and the image analyzer 16(1) preferably include features for black and white shape data recording, multiple horizontal, and vertical line scanning, gray scale imaging and color imaging, shape recognition, geometric comparison, seam tracking and defect detection to determine actual weld process parameters and variables. The arrangement includes line scan recognition features to enable defect detection based upon metallurgical data and to permit recording of both actual defects and the location of defects.

The detector 17 and the radiometer 15 are preferably an Inframetrics Model 525 arrangement made by Inframetrics, Inc. of Bedford, Mass. This arrangement enables image production on a 30 Hz cycle basis, based upon a mercury-cadmium-telerude sensor, which is preferably liquid nitrogen cooled. The lens of the detector 17 is selected to accommodate the radiance of the weld puddle being scanned. The output of the detector 17 to the radiometer 15 is indicative of the difference in radiance "dI" between an elected reference radiation level, and the amount of radiation from the weld puddle under observation. The difference "dI" particularly equals:

where "e.sub.t " is the emittance of the weld puddle 19 surface as a function of temperature, "N.sub.t " is the radiance value of a black body at temperature "T.sub.t " at an unknown target surface, "N.sub.b " is the radiance measured directly from background, "e.sub.r " is the emittance of the reference surface, and "N.sub.r " is the reference system radiance of a black body at the selected puddle temperature "T.sub.r ".

The detector 17 produces individual frames of information, which define 256 by 256 pixel matrices. The output of the detector 17 is particularly a video signal for transmission on a parallel bus 17' to radiometer 15. The video signal produced is an 8-bit 256 by 256 pixel representation having a broad-based gray scale capability. The radiometer 15 is setable to operate at selected temperature ranges. The radiometer 15 is further programmably setable by an input signal in RS232 format. The radiometer 15 further includes built-in tables and mapping functions that convert the received detector signal along the bus 17' into video representations of the image viewed by detector 17.

The output of the radiometer 15 is connected to an image analyzer 16(1). The image analyzer 16(1) processes the data received with selected shape recognition algorithms preferably written in "C". The determination of the shape of the weld puddle and finding its outer boundary is accomplished by tracking the weld puddle 19 along its boundary, and by determining the maximum puddle width and the centroid of the puddle. This is done digitally, as discussed below. In particular, the radiometer 15 provides a video signal which is received by an image processing board on the image analyzer 16(1). This board converts the analog video signal into a digital representation as an array of bytes on the image processing computer 16(1). Each byte represents one of 256 different gray levels. The image processing array on the computer 16(1) is able to accommodate and store a complete imaginary of up to 512 by 512 pixels, 8 bits of information per pixel. The array is scanned 30 times per second permitting the radiometer 15 effectively to employ a video display (not shown) with controls that permit setting cross hairs over specific features on the weld puddle using manual dials and knobs. Line scanning, according to the invention, includes shape recognition in three-dimensions using a single camera which has an (x,y) axis, and is capable of presenting (x,y) information plus the brightness associated with particular points on the (x,y) representation plane. The two-dimensional (x,y) image has numerical values indicative of intensity. This produces a three-dimensional, displayable effect. By examining the actual images generated, certain features, such as the slope of the sides of this three-dimensional image, become apparent as a qualitative measure of how fast the sides of the weld puddle are cooling. In such a representation, defects in the weld puddle 19 show up as dark spots.

In the preferred embodiment of the invention, the precise technique whereby the image analyzer 16(1) extracts information from the digital representation of the infra-red image of the puddle 19 is as follows. First, the center of the puddle 19 is determined by locating a point at the center of brightness of the image. Assuming that the upper left corner of the image has the coordinates (o,o), with the x-axis increasing in value to the right across the top of the image and the y-axis increasing in value downwards along the right edge of the image, and if P(x,y) is the brightness of the pixel at position (x,y), the center of the weld puddle is determined as follows:

S=sum from y=0 to y=Ymax of the sums from x=0 to x=Xmax of P(x,y); R=sum from y=0 to y=Ymax of the sum from x=0 to x=Xmax of the product y times P(x,y); and C=sum from y=0 to y=Ymax for the sum from x=0 to x=Xmax of the product x times P(x,y), where P(x,y) is the pixel intensity as a function of x and y, and x and y are planar position coordinates of the puddle 19.

The coordinates of the center of brightness, i.e., XB and YB, are then

Having found the puddle center, the image analyzer 16(1) subtracts the value 40 from the brightness magnitude at this center point to give rise to a pre-established edge brightness value. The value "40" was selected by an expert metallurgist as giving a reasonable definition of the puddle edge brightness, according to the preferred embodiment of the invention. (The puddle edge is fairly well defined in the image, and naturally different hardware and software would require a different edge brightness value or method of computing the same.)

Next, the analyzer 16(1) scans from the puddle center of brightness towards the puddle edge to the left, right, and downwards, and possibly in other directions as well, thereby defining points on the puddle's edge. The x-axis width of the puddle 19 is then determined as a first puddle parameter. A rectangular box of predetermined adequate dimensions, e.g., a centered square with sides equal to the radius of the measured weld puddle 19 is then established about each of three puddle edge points, i.e., the front tip and the sides. Within each box the following puddle parameters are computed: the area of the puddle 19 within the box; the length of the puddle perimeter; and the number of times that the puddle boundary doubles back upon itself and crosses an imaginary "smooth" puddle boundary. All of this information is presented in digital, floating-point form and is readily retrievable by the image processing advisor 115i.

The weld puddle line scanning effectively enables the establishment of isotherms in the weld puddle image, as shown in FIG. 18. Establishing the isotherms is particularly accomplished by examining each line of the pixel matrix, and evaluating the data on the line pixel by pixel. At each transition across a selected reference level, a particular isotherm value is established corresponding to the selected reference level. Thus, by using line scan data and examining each of the points in the pixel matrix, line by line, a three-dimensional isotherm representation, as shown in FIG. 18, can be established from the raw line data.

Additionally, as suggested above, by examining the pixel matrix, a mathematical operation can be performed to determine the approximate center of the data clusters in the frame, which are specifically indicative of the image of the weld puddle. This is useful for focusing effectively on the centroid of the weld puddle 19 and helps tracking the weld puddle and further processing of the image in view of cognizance of its actual center. It further improves puddle tracking according to a selected boundary tracking algorithm. Image processing further includes the establishment of a theoretical box around the weld puddle. By scanning along the box edge and checking the brightness of the image along its perimeter, one may determine whether bright pixels are present, which would indicate the puddle has reached the edge of the box. If no such bright pixels are present, it is clear that the puddle has not grown outside the established box boundary. This provides a very simple measure of how big puddle 19 is, which is particularly useful in resistance welding shape recognition. One image processing algorithm looks at the leading edge portion of the weld puddle 19 and performs a length calculation of the perimeter, which is useful as a defect indication. The side zones are employed for determining the puddle width by locating the sideward edges of the weld puddle and doing a subtraction. The tip zone of the weld puddle 19 produces substantial amounts of tracking information, because the weld seam passes directly under the tip of the progressing weld puddle 19. The side zones are useful for establishing the patterns and derivatives of predetermined isotherms, the latter indicative of the cooling rates of said weld puddle 19.

Tracking is further accomplished by measuring transverse displacement of the weld puddle 19 from the seam it is intended to follow. In particular, by viewing the front edge or tip of the weld puddle 19, it can be determined in what direction the seam of the weldment is proceeding. The seam of the weldment may deviate from the general direction of the robot and the welding torch it carries, defining an offset angle. According to the invention, this causes welder 20 to turn transversely back toward the interface 19" between the members 14 being welded, to ensure their effectively being joined.

Further, mismatch errors can occur, according to which one plate 14(1) is misaligned with the other plate 14(2), establishing undesired weld puddle protrusions. This causes the weld puddle to flow from the edge of one of plates 14 into the space between the plates 14. In order to effectively counter this, the welder torch 20 is rotatably adjusted to aim directly between plates 14 from an oblique angle to ensure creation of a strong weld.

According to the thermographic detection techniques employed herein, and as suggested in the Appendix, a box is defined about the center of the weld puddle 19 to determine whether or not it has obtained a minimum acceptable size for data to be permitted to be taken by the detector 15 in order to establish a pixel array frame for use in image processing, as indicated herein. If the weld puddle 19 crosses the perimeter of the box, the pixel brightness will exceed a predetermined threshold, and the puddle 19 is deemed sufficiently large for taking data arrays for image processing. According to this technique of determining the minimum acceptable weld puddle size, resistance welding can be implemented effectively.

As before, the technique includes establishing the size of the box, and determining a pixel brightness threshold value based upon weld process parameters. The selected box is, for example, established at suitable coordinates, (e.g., 100,100) in a field of 200 by 200 pixel elements, with a pixel intensity threshold value set at 150, for example. These values are established in view of the predetermined minimum acceptable size of the weld puddle 19 and the camera imaging factors such as zoom ratio and emissivity of heat from the weld puddle 19, for example. To determine the position of the box established around the weld puddle 19, the infrared detector 15 is calibrated with respect to the axis of the welder torch. A looping process accomplishes this by continuous scanning of consecutive pixel array frames at successively increased pixel threshold values until the box contains a sufficient number of pixel values. When the weld puddle 19, accordingly seen, exceeds the bounds of the box, a signal of that event initiates calculation of the center or centroid of the weld puddle 19. Finally, this information is communicated to the image processing computer through a suitable communication interface.

FIG. 18 shows line scan information generated from an array of pixels representing infrared intensities of a weld puddle plotted as a function of position on a flat planar projection of the weld puddle. A fairly good representation of puddle intensities can be established with even a limited number of line scans as shown in FIG. 18. A more precise representation might include 500 row scans and 500 column scans, or 250,000 pixels. This would permit the development of an extremely accurate set of predetermined isotherms along the contour of the line scan information in FIG. 18. However, this kind of accuracy is not required and the desired isotherms can be produced with only several scans, as suggested in FIG. 18.

Specific zones within the weld puddle, such as the tip zone, the width zone, and the diagonal or shoulder zones, respectively. Selected zones of interest are established for examination by referencing the selected zone relative to some known feature, for example, such as the center of the area of the puddle. Once the center of the weld puddle 19 has been determined, one can then fix selected zones of interest relative to this established center. Alternatively, by detecting certain features on the edge of the puddle 19 such as the tip of the puddle 19, the position of these special interest zones can be fixed as well. Other such edge features include, for example, protrusions or fingers on the leading edge of the puddle. As already discussed, the central zone is employed for determining puddle width.

To determine the center of the weld puddle, according to the invention, the center of mass of the puddle is determined. According to this technique, the entire puddle image is considered to be a single region, and the area and the centroid of that region are calculated. From the centroid of that area, all of the special interest puddle zones are established on an (x,y) coordinate system. Accordingly, the puddle side zones are fixed between the tip or leading edge of the weld puddle 19 and the side zones. The sizes of the special interest zone are predetermined. As noted above, the actual positioning of the zones in the (x,y) coordinate system is fixed, according to the image analysis techniques indicated. As the weld puddle image or position changes within successive pixel frames, the images on the indicated zones will bear evidence to these changes.

In spot welding operation, according to the invention, an image processing window is established over the expected position of the developing puddle or nugget to ascertain the current physical size of the nugget. When a sufficient size has been attained, the weld current is halted.

The slope, derivative, or gradient of the detected isotherms is monitored to determine the maximum weld puddle coding rate. The hotter the puddle or nugget gets, and the greater the difference between the ambient temperature and the plasma of the puddle, the greater its cooling rate will be. This cooling rate is additionally affected by other factors, such as the shielding gas temperature, the particular materials being welded, the massivity of the material, and whether or not the materials welded have been pre-heated, for example. To ensure a quality weld according to the particular metallurgical characteristics of the metals being welded, predetermined maximum and minimum cooling rates must be maintained during welding operation. In other words, the best metallurgical properties occur within a certain band of cooling rates. Thus, according to the invention, to maintain the cooling rate within desired bounds, the weld current and weld speed are modified to keep the cooling rate within bounds. If the cooling rate is too great, for example, the weld current is reduced and the weld speed is increased. According to the invention, preference is given to decreasing the weld current or increasing the weld speed alone, without modifying both parameters. Further, according to the invention, an increase in weld speed is preferred as a parameter modification because this tends to increase the throughput at a particular work station, within bounds, of course. Further, if the cooling rate is quickly to be modified, weld current is modified first because weld current is more responsive to speedy modification than weld speed.

To determine the rate of cooling, the pixels in the side zones are examined because the maximum cooling rate is expected in those regions. In particular, selected pixel rows in the side zones are examined, looking particularly at the peripheries of the side zones, namely, the transition between the molted puddle and the heat affected zone immediately adjacent the puddle, which is still solid but not metal. There is a narrow band of about 10 to 20 pixels which constitute the interface zone between the molten and the solid regions. By observing the pixel values in this portion of the pixel array, one can establish derivative of the heat in terms of the maximum and minimum values that occur in this threshold region. Further, the difference of maximum and minimum is divided by the number of pixels over which the change occurs. That establishes the proportional cooling rate desired. In particular, that establishes the transverse proportional cooling rate be along the "x" axis, while the robot and the welding progress along the "y" axis.

In addition to controlling and monitoring weld speed and weld current, the process controller tracks and controls torch position and disposition by monitoring the location of the weld puddle 19. The process controller further monitors error conditions, such as mismatch (when the surfaces being welded together are not coplanar and substantially coincident at the weld point), misalignment between the plate seam and the path followed by the welder and excessive root opening, for example, and adjusts torch position, altitude, and disposition to minimize or eliminate such error conditions. In response to detection of such mismatch, misalignment and excessive root opening conditions, weld speed or current is modified, or the torch is tilted or shifted to ensure clear establishment of an adequate weld seam between the plates to be welded. The torch angle is typically perpendicular, or normal to the surfaces being welded. However, the welding robot can tilt the welding torch sideways, forward or backward, as required, and can position the torch forward, backward, and sidewards anywhere on the (x,y) coordinate plane, as well as shift it along the z-axis to control the arc length of the torch.

The image analysis further includes an algorithm that observes the shape of protrusions at the leading edge of the weld puddle. In particular, the image analysis algorithm generates an indicative notation that the puddle is in a state of vertical mismatch between plates being welded, if a weld puddle signature indicative of mismatch has been noted. The rule set then determines the torch orientation change needed to counter the effects of mismatch. Such an orientation change, for example, might call for the torch to be tilted sideways during welding of the mismatched segment, and the position of the torch might be slightly offset as well, if necessary.

The signature of mismatch, as indicated in FIG. 2B, is a characteristic protrusion or pair of fingers at the leading edge of the weld puddle. Misalignment, as can be seen in FIG. 2A, is also indicated by a protrusion or finger 222, 222', and 222" at the leading puddle edge, depending upon the degree or amount of misalignment. In particular, misalignment occurs when the direction of travel of the torch is not parallel and on center with the weld seam. Excessive root opening, as suggested in FIG. 2C, develops as the adjacent plates being welded separate for a period and then come together again. This also shows up in terms of a rather large protrusion 222, 222', or 222" on the front edge of the weld puddle, which may or may not progress, depending upon what the actual opening does. According to the invention, welding speed is slowed, weld current is pulsed more heavily or is increased to ensure that both sides of the weldment are effectively fused.

The arc length between the torch 20(5) and the weld puddle 19 is controllable, according to the invention. An excessive arc length will be insufficient to cause a weld because not enough current will be transferred between the torch and the plate being welded to actually perform welding.

The image processing, according to the invention, includes comparing pixel images acquired against a qualified body of information on weld puddle images, which has previously been stored for comparison with actual sensed weld puddle information and/or images. The center of area or centroid of the weld puddle 19 is determined, as well as the partial perimeter of the weld puddle 19. The weld puddle cooling rate can, for example, be determined from change in the puddle width over time. According to the invention, the partial perimeter in the tip zone indicates the absence or presence of protrusion 222, 222', or 222".

Additionally, by counting zero crossings of the fingers 222, 222' or 222" and their location, the precise character of the defect can be determined. The number of zero crossings are determined by applying a smoothing filter to the zone of interest. A large number of fingers 222 indicates mismatch, for example.

Further, by observing the relative sizes of the fingers in terms of width and area, mismatch and misalignment between seam and torch can be differentiated. Additionally, as misalignment is progressive from frame to frame of pixel matrix information, the trends indicated in the image processing can suggest a tracking error due to misalignment. As one finger 222 moves farther and farther from the center of the travel direction, misalignment is indicated rather than mismatch. In misalignment, one would naturally expect to see a major protrusion 222 grow and also go off at an angle from one frame to the next. Other minor protrusions may be present as well. Thus, if angular progression of a major protrusion 222 is not seen, there is no misalignment, but instead there is another defect.

The image processing algorithms determine the raw features of the weld puddle, and actual error determination is performed in the image processing expert system, which employs a predetermined rule set to establish those determinations. All incipient weld errors look very similar. Accordingly, no effort is immediately expended to distinguish the kinds of defects. Instead, the weld defect is permitted to grow a bit under observation. As it grows, a characteristic pattern begins to emerge, including one or more protrusion 222.

In the case of mismatch, there is typically one protrusion 222 which is larger than another. It is a primary object of the inference techniques of the invention herein to notice such a trend. Weld root opening errors are, for example, indicated by the growing and shrinking of a single finger or protrusion 222. Misalignment errors are generally indicated by one or more protrusions that diverge at an angle relative to the axis of the weld seam. The direction of the weld in this instance has a progressively changing angle with the weld seam. Mismatch, on the other hand, is not a very progressive event and it is signified by at least two protrusions, one of which is noticeably smaller than the other. Misalignment is characterized by equally-sized multiple protrusions.

As set forth explicitly in the Appendices hereto, image processing computer 115i thus calculates features such as weld puddle perimeters, zero crossings through a selected area of the weld puddle, the centroid of the weld puddle 19 and various combinations of features which produce a high level description of the type of weld error, indicated by the protrusions.

To operate the invention, according to one version thereof, four IBM AT microcomputers 115 are employed, as suggested in FIG. 8, with one of them 116(1) specifically designed to do image analysis with an installed board. Connected to the auxiliary controller or image analysis node 115x is another advisor node 115i which has the image-processing rule set loaded up. Referring now to FIG. 19, the physical environment of the present invention is disclosed. The invention contemplates using a plurality of distributed processors to spread out the logical computations and thereby enable multiple processes of control to proceed together in real time. In the preferred embodiment, four AT-class IBM personal computers are utilized, as is shown in FIG. 19. A first computer is an image analyzer 16(1) which accepts the infrared image from the infrared detector 17 and the radiometer 15 and extracts, therefrom, the useful image data, which is used by the remaining logical systems to control the welding process. A second computer is used to implement the image processing advisor 115i, which accepts the image data and, by the application of rules thereto, determines the state of the welding process and whether there is any need for corrective action. A third computer is then used to implement the expert controller 115e, which then responds to the data generated by the image processing advisor 115i by determining the proper control strategies. These strategies are then fed as commands into the fourth and final processor, the controller actor 115a, which ultimately controls the welder 20.

Each of these computers is set up preferably using the industry standard MS-DOS operating system of Microsoft. Each of these computers includes an industry standard 80286 microprocessor equipped with extended memory so that logical computations can be carried out in extended memory space, while normal programs run in a 640,000 unit memory bank of each of the processors.

All four of the processors are set up and programmed in a similar manner, and so FIG. 19 highlights the internal structure of the image processing advisor 115i as representative of the internal structures of each system. Very broadly described, the image processing advisor 115i contains a physical layer 153 which, in the preferred embodiment, is implemented using Borland's Turbo Pascal programming language. This physical layer 153 interacts with the serial ports and buffers 739, 740 and 741 of the AT-Class computer using standard software utilities provided by Blaise to service data flowing into the serial ports, to buffer that data, and to present it to the Pascal programming system. In the preferred embodiment of the invention, each of the four computers contains a physical layer 153 implemented in Pascal, and, therefore, all four systems are able to communicate among themselves using this common programming physical layer. When any of the systems is initially turned on, the Blaise Computing software that services and buffers the serial ports is first made resident within the MS/DOS operating system. Then, in the case of the processors 115i, 115e, and 115a, a Lisp programming system kernel is loaded over the operating system and given program control. The Lisp kernel then sets up a variety of storage areas within the extended memory of the computers. The Lisp kernel implements a logical layer 152 within each of the three computers just mentioned. Because of its ability to "shell" into other programs, the logical layer implemented in Lisp may execute programs written in the "C" language, as is indicated at 152' in FIG. 19. In the case of the image processing advisor 115i, there are no "C" language sub-routines. But in the case of the controller/actor 115a, a plurality of "C" programs are run by the logical layer to control the welder 20.

Once the logical layer 152 gains program control and allocates its memory for its own needs, it launches the Pascal layer 153 by shelling into a Pascal program. Through the use of shared, common memory locations in low memory beginning at hexadecimal address 4F0, the logical layer 152 places pointers to various addresses, which it sets up to be shared with the physical layer 153. By accessing the locations beginning at 4F0, the physical layer 153 gains access to these address pointers and is thereby able to share certain memory locations with the logical layer 152. In particular, a blackboard region 742 is set up into which the logical layer 152 may write the values of variables, later access the values of those variables, and from which the physical layer 153 may access the values of those variables without having to interrupt the operation of the logical layer 152. In this manner, one processor may request the values of certain variables from another processor directly from the physical layer 153 without having to interrupt the logical processing proceeding in the logical layer 152.

A second shared memory location is the message queues 734, 742, and 736. A first message queue is a priority message queue 734 into which the physical layer 153 writes high priority messages that must be serviced immediately by the logical layer 152. Whenever such a message is entered into the queue 734, the physical layer 153 sets a semaphore 717', generates an inference interrupt 717 to interrupt the observations of the logical processor 152, and directs its attention to the priority matter. Other matters not requiring the interruption of the logical layer are messages placed in the non-priority queue 742. The logical layer generates output messages by placing them into an output queue 736 from which they may be read by the physical layer 153.

While the physical layer 153 is initially running, it attaches itself to the real time clock 719 interrupt of the AT computer, thereby causing itself to be placed into operation roughly every 1/18th of a second during a normal operation of the computer. Having thus attached itself to the real time clock, the physical layer 153 returns program control to the logical layer 152 which then commences its normal operations of processing logical functions of the image processing advisor 115i.

The image analyzer 16(1) includes a physical layer similar to the layer 153 in the image processing advisor 115i, but does not include a logical layer 152 implemented in Lisp. Because of the need for speed in the processing of large quantities of image information, the image analyzer 16(1) contains programs written in "C" that accept and process image information, and generate messages for the physical layer 153 to transmit back to the image process advisor about the state of the image in response to requests received from the image processing advisor 115. Naturally, it would be possible to have additional logical layer operations proceeding in the image analyzer 166, if that computer were fast enough to handle those in addition to its immediate task of condensing the large quantity of image information into the small amount that is needed to implement control in accordance with the teachings of the present invention.

The controller actor 115a is essentially the same as the image processing advisor 115i but contains many "C" programs analogous to 152' that control the welder 20. It is thus labeled as containing "Lisp, Pascal and `C`" programs.

Returning now to the image processing advisor 115i, the normal operation of this advisor proceeds as follows. The logical layer 152 establishes a semantic net 215(4) which contains the goals of the system, all of the "forward-chaining" rules which define the various processes that the system may carry out, and all of the "backward-chaining" rules which define how various system results are computed from external and internal system variables. The logical layer 152 is implemented using Golden Hill Lisp Large Memory Version 2.2, according to a preferred version of the invention. This is a conventional Lisp processor, and other suitable A-I processors could be used as well.

As it comes, this Lisp processor is capable of managing multiple operating environments and processes, and switching between them upon request. Each of these operating environments is called a "stack group" and contains a complete inference process together with all of the relevant data. The logical layer 152 is capable of maintaining a plurality of stack groups, alternatively executing them, and suspending them under program control. However, the logical layer 152 does not include any provision for real time scheduling of these tasks but only for maintaining multiple tasks and switching between them. It was, therefore, necessary to achieve the present real time process control scheduler by integrating the logical layer 152 with a physical layer 153 that is connected to a real time clock 719.

In accordance with the teachings of the present invention, the logical layer 152 maintains a plurality of inferencing queues 218 including a running queue 218(1), a ready queue 218(2), and a suspended queue 218(3). The names of stack groups are stored within these queues. The running queue contains the name of the stack group that is currently active and being processed by the logical layer 152. The ready queue 218(2) contains a linked list of the names of inference processes that are running on a "time-shared" basis, but that are currently not being processed by the logical layer 152. The suspended queue 218(3) contains the name of tasks that have been suspended pending the result of an input-output request for information from the logical layer 152 to some other machine, such as the image analyzer 16(1) or the expert controller 115e.

To achieve real time switching between inference tasks, the physical layer 153 is driven by the real time clock 719 to count the passage of time. Periodically, the physical layer 153 decides it is time to schedule the running of a new task. Accordingly, the physical layer 153 sets the semaphore 717' to indicate that it is time for a new task to be run and generates an inference interrupt 717 to interrupt the operations of the logical layer 152. The inference interrupt is actually the "control-break" interrupt of the computer, since the control-break interrupt is already used by the logical layer 132 as a means for responding to keyboard interrupts from the keyboard 368, as is conventional in software programs designed to run on AT class computers. But the logical layer 152, in response to the control-break interrupt, checks the semaphore 717'.

If the semaphore 717' is in its reset state, then the logical layer 152 knows that a keyboard interrupt has been received from the keyboard 368 and processes it in a normal manner, servicing a keyboard request from the knowledge engineer or system supervisor. But if the semaphore 717' is in a state indicating that it is time to switch tasks, then the logical layer 152 moves the process at the top of the ready queue 218(2) into the running queue 218(1) and moves the task from the running queue to the end of the ready queue 218(2). In this manner, the physical layer 153 and the logical layer 152 combine to form a real time process control inference task scheduler that rotates program control systematically through the inference tasks in the ready queue 218(2), devoting a fixed amount of time during each period to each of the tasks waiting to be completed.

When the logical layer 152 needs to communicate with other processors, it places a message in an output queue 736. The physical layer 153, when next awakened by the real time clock 719, senses the presence of this message and initiates the feeding of the message out of the serial ports and buffers 739, 740, and 741 to one of the other computers. In this manner, the logical layers 152 within the computers may pass data back and forth, and may also pass requests for the performance of various inference tasks.

In response to an incoming message from another processor at the serial ports and buffers 739, 740, and 741, the physical layer 153 places a message either into the priority message queue 734 or into the non-priority message queue 742. When a message is placed in the priority message queue 734, the physical layer 153 also adjusts the semaphore 717' to signal that a priority message has arrived, and then it initiates an inference interrupt (actually a "control-break" interrupt) in the manner that has been described. In response, the logical layer 152 checks the semaphore, and upon finding that it signals a message, retrieves the message from the priority queue 734. Non-priority messages in the queue 742 are processed by the logical layer 152 without the assistance of an interrupt or the semaphore 717'. While the logical layer 152 is processing a priority message, it leaves the semaphore 717' set, and this semaphore blocks the physical layer 153 from taking further action with the message queues until the logical layer 152 has finished and cleared semaphore 717'.

Some incoming messages simply require the return of data from the blackboard 742 to one of the other computers. In this case, the physical layer does not need to generate a message for the message queue but simply goes directly to the blackboard 742, retrieves the necessary information, and forwards it to the other machine. However, if a message contains data that is to be placed in the blackboard 742, the physical layer 153 must place the message into a message queue and allow the logical layer 152 to transfer it into the blackboard 742.

Communications between the logical layer 152 and the physical layer 153 to access the message queues, and the semaphore 717' are carried out by assembly language programs that are tied respectively to the computer interrupt 63H in the case of the message queues, and the computer interrupt 64H in the case of the semaphore. These machine language programs enter data into the message queues, remove data from the queues, read and set the semaphore 717', and do so with all interrupts locked out, so that when one of the systems 152 or 153 is accessing a message queue or the semaphore 717', the other is prevented from simultaneously accessing the same entity, thereby destroying some data. These assembly language programs are loaded by the Pascal layer 153 when layer 153 is itself initiated.

The display 367 and keyboard 368 of the Lisp processing advisor 1152 interact primarily with the logical layer 152 and are used by the knowledge engineer to monitor system performance, to add new rules, and to alter the performance of existing rules in the expert system which is controlling the welding process.

According to a preferred version and referring now to FIG. 8, the basic start-up activity of the system includes presentation of a top-level goal such as "to perform GTA welding" (GTA means "gas tungsten arc"), for example, to the root node or expert controller 115e. This goal is analyzed on node 115e by the semantic network and its interpreter. In order to perform the particular welding goal, a certain rule set and possibly other knowledge bases, such as frames containing constant data, are required to analyze that goal. These are accordingly loaded dynamically from a hard disk (directly or via Ethernet) along data lines 115d", for example, from a file transfer network 116a, when that goal is presented. From that point, the controller 115e starts generating subgoals and passing these goals to the other computers or nodes. Those nodes receive the goals, evaluate them, and determine which particular rule sets they need to perform their particular analyses, and accordingly they dynamically load the needed knowledge bases and frames. Parallel processing is thus initiated.

The detector 17 and the welder 20 are mounted on the robot 22 for transportation. The detector 17 is preferably positioned at a 30 degree angle from the surface of the workpiece 19' (FIG. 1), focusing directly at the weld puddle 19 to enable the scanning thereof. The detector 17 is further preferably positioned about 12 inches from the weld puddle 19. The weld puddle 19 accordingly remains in the same place relative to both the detector 17 and the welder 20 as the robot 22 moves to conduct welding. This effectively permits controlled, closed-loop arc welding to be accomplished with respect, for example, upon a pair of adjacent, square grooved workpiece elements 14(1) and 14(2) of AISI 1010 carbon steel, which are about 0.250" thick.

After analysis by the image analyzer 16(1), the results are input to the other elements of the controller 16, which directly governs selected weld parameters and variables, including weld current and speed, for example. Other parameters such as weld voltage, torch arc oscillation, arc pulsing, arc length, and wire feed speed are controllable as well, according to the invention. Shield gas dew point temperature, as suggested at FIG. 3A, is additionally detected, and the results are provided to the controller 16 to shut down operation if the shield gas quality diminished unacceptably. The shield gas employed is inert and is typically argon, for example. The quality of gas is established with a dew point temperature measurement of the shielding gas by a detector 16(2) through which shield gas line 20(4) (FIG. 3B) passes. Shield gas dew point temperature, as is well known, is related to the moisture content of the gas. A device for making the measurement is available from Alnor Company in its humidity detection series. The shield gas must be very cold, which prevents oxides from forming. FIGS. 3A and 3B additionally show elements of the welder arrangement 20 including a weld wire drive 16(1,2), weld wire 16(1), the shield gas dew point temperature detector 16(2), a source of inert shield gas 17(1), a welding power supply and controller 18(1), and an arc controller 18(2). The electrode 20(5) is consumable or non-consumable tungsten, depending upon the weld process employed. If it is non-consumable, as in the GTA process, no weld wire is used, as the electrode is not made of consumable weld wire. Each of these elements is controlled by the controller 16. The weld wire 16(1) is used to fill in the gap between the welded parts 14. Accordingly, the weld wire is fed in at a speed proportional to the torch and robot travel speeds. The robot speed, of course, depends also upon the gap between the parts 14 being welded. The shield gas electrode 20 further includes a cylindrical cup 20(1) which surrounds the central electrode 20(5). It is out of this cup 20(1) that the shield gas flows to prevent oxidation of the weld puddle during welding. Accordingly, there is an inlet 20(4,1) in the top side of the cup 20(1) into which the shield gas flows. Accordingly, it flows out of weld cup 20(1) at its lower open end. FIG. 2A shows the weld puddle 19 adjacent the electrode tip 20(5) and indicates the respective special interest zones 333 of weld puddle 19, which provide information required for image processing, as is discussed herein. Several weld puddle contours are shown, depending upon the degree of misalignment of the electrode tip 20(5) from the seam between the parts 14(1) and 14(2) (FIG. 1) being welded together.

FIG. 4 shows in schematic form the general scheme of one version of the process controller 16 designed in accordance with the invention herein. In particular, it illustrates that the actors 115a which control the real-world process are data driven, while the inner nodes, such as the expert controller 115e and the advisors 115ad, are goal driven (by the centrally positioned user 400, for example).

FIG. 5 shows a preferred version of the invention herein, in which the controller 16 includes a single multi-inferencing expert (MIX) hardware node 115. The hardware node 115, which might be implemented using a single AT-class computer programmed in Lisp, "C", and Pascal, may communicate with devices such as sensor 116 and/or one or more controlled devices 117. The present invention contemplates providing multiple nodes, each of which may reside on a different hardware processor, but some of which may share the same processor.

As indicated in FIG. 6 according to one version of the invention, each MIX hardware node 115 particularly comprises an inference processor 119, a dynamic priority scheduling mechanism 121, a knowledge layer 124, a logical layer 129, a knowledge base 131, a status information base 133, a physical layer 130, and a communications mechanism 136. The node 115 acts in response to information contained in a knowledge base 131, as well as in response to real time information provided by the sensors 116 (FIG. 5), for example. The inference processor 119 is loaded in a particular hardware node 115 to conduct inference processing interpreting pre-established rules, which are stored in the knowledge base 131 with other knowledge elements (FIG. 13) such as frames 215(1), objects 215(3), and semantic networks 215(4), as indicated in FIG. 6. The nodes 115 conduct inference processing under direction of a dynamic priority scheduling mechanism 121, which operates as an inference multiplexer. The logical layer 129 thus associates each inference process with a selected logical location in any one of a number of the nodes 115. The logical layer 129 thus implements a mapping operation which enables the operation of scheduled inference processes by insuring the availability of information to complete particular operating reference processes. In particular, the scheduling mechanism 121 schedules inference processes in a regular, clocked sequence from a ready queue of waiting inference processes into the running queue. Then, the inference processor 119 evaluates the selected inference process by interpreting a corresponding rule 215(2) in conjunction with selected knowledge base elements, including the semantic network 215(4), objects 215(3) and frames 215(1), for example, as will be discussed below.

As illustrated in FIG. 8, a particular node 115 can operate as an expert controller with one or more elements, such as sensor 116 or controlled element 117 through external control 117. A suitable interface layer, discussed above, is, however, required to accommodate the specialized connection thereby established. This interface enables the use of specific data transmission types as required. In particular, data transmission can follow RS422, Ethernet, MAP, or TOP format. The interface layer or mechanism could be part of physical layer 130 in FIG. 6.

FIG. 8 further shows controller 13 to include a plurality of nodes 115 distributed to form a network interconnected by data busses 115d' at respective ports 115p. The hardware nodes 115 in FIG. 8 are variously configured as expert controllers 115e, advisors 115ad, or actors 115a. Expert controller 115e communicates with auxiliary controller 115x as to sensor 17, controls 117, the network 169, and input/output (I/O) element 117i. The expert controller 115x is a selected configuration of a hardware node 115 with particular features activated to perform expert controller functions including, for example, certain actor-directed features. In particular, the hardware node 115 features which are actor-directed or actor-characterizing include interface features enabling operational and functional compatibility with the selected sensor 117 or a control 117.

With reference to FIG. 9, the various actor-characterizing features of a node 115 are the knowledge layer, the logical layer 152, including transmitter/receiver map features effective for establishing the location of inference operations, and the physical layer 153, which features a translator for data communications in Ethernet, RS422, MAP or TOP, as well as specific data communication mechanisms effective to implement the translation operations performed by the knowledge layer 151.

The disclosed arrangement interfaces the detector 17 and the radiometer 15 directly to the image analyzer 16(1) which, in turn, is interfaced to image processing advisor 115i. (No auxiliary controller 115x exists in the currently operative arrangement; it is shown as an additional alternative or enhancement to the invention.) The image analyzer 16(1) node contains programs which process and condense the information received from the radiometer 15 in preparation for rule base analysis one node up in the image processing advisor 115i. The auxiliary controller 115x is preferably interfaced directly with the expert controller 115e node, rather than with an advisor node, such as image processing advisor 115i. The auxiliary controller 115x can be employed as a work cell controller, for example, which provides an initial goal to the expert controller 115e. The work cell controller, for example, could conceivably be a CAD system or a robot controller. What characterizes the auxiliary controller 115x is that it requires special interfacing that takes into account the external components. The key factor in being an actor is the inclusion of specific control or sensor processing algorithms for particular associated hardware devices. This additional software, for example, might include "C" algorithms for image processing.

Each hardware node 115 additionally includes at least a single generic communication port 155p which is customized to interface with other computer nodes 115 of the controller arrangement 16. Additionally, according to the invention, each hardware node 115 is capable of managing a plurality of inference processes, in sequence or in parallel, under control of a dynamic priority scheduling mechanism 121 (FIG. 6), subject to a predetermined adaptive process control language (APCL) (See Appendix IX), which features block scoping techniques, the use of various type variables and assignments, standard and user defined data structures, and, additionally, literal and conditional control structures. The language itself implements such features of artificial intelligence as frames 215(1), objects 215(3), semantic nets 215(4), and pattern matching 217, as is suggested at FIGS. 6 and 9.

Further, as suggested in FIGS. 6 and 9, the controller 16 architecture includes a logical layer 152, which associates particular inference processes with specific nodes in the network. Thus, the logical layer prevents controller 16 from having to concern itself with the question of at which node an inference process is to be carried out. Further, by utilizing knowledge-base communications, the controller 16 ties together the nodes into a single intelligent controller. Simply stated, a distributed expert system, designed according to the invention herein, is an expert system which has its knowledge distributed among more than one node 115 but which nonetheless operates as a single process.

In operation, the various pre-established process variables including weld speed and current, for example, are determined, and any defects in the weld puddle are discovered through image processing of the sensed information in view of the contents of the knowledge base 151 (FIG. 9), subject to predefined rules. For example, if the materials 14(1) and 14(2), which are being square groove butt joined in a plasma arc process are subject to joint mismatch, there will be a substantial, definable transverse puddle shift and decreased puddle width, or the puddle 19 may have an elliptical leading edge. If the materials 14 are subject to excessive root opening, there may be present two notches on the leading edge of the puddle 19. If the materials are axially misaligned, the weld seam may be on the side of the puddle. Unacceptable axial misalignment in a fillet joint weld using the flux cored arc technique is indicated by a defined transverse puddle shift.

The width of the weld puddle 19 is continuously monitored during welding. Weld travel speed is governed by microprocessor control, as is the current supplied to the welder 20. Further, the axial or longitudinal weldline 19" or centerline of the weld joint is continuously monitored for notches, which are suggestive of failures in fusion between the plate edges. Thus, according to the invention, controlled welding of square groove butt joints is enabled without substantial root openings.

The controller 16 employs primarily Lisp source code in a typical Lisp operating environment. The hardware environment for implementation preferably includes, for example, several IBM AT computers arranged in a network of nodes 115. To achieve a general goal which has been input by the operator, backward chaining occurs in the semantic network 215(4) (FIG. 13) so that the rules 215(2) applied produce intermediate goals until a plurality of intermediate goals are established and satisfied.

The controller 16 is a single expert system which is distributed over a plurality of nodes connected as a network, without each node being a separate and distinct expert system in its own right. In other words, according to the invention herein, a single, overall expert system is distributed over several nodes or machines. Accordingly, the arrangement herein distributes the inference mechanism over several processors, rather than simply distributing the knowledge components themselves among several processors by employing the technique of knowledge-based communications, according to the invention herein. Thus, a plurality of inference mechanisms or nodes can operate concurrently to achieve selected goals by continually updating sets of current knowledge until each set includes specific goal-desired information. Recursive searching to satisfy goals is thus conducted by both forward- and backward-chaining. Further, generalized knowledge mechanisms 601, as suggested in FIG. 14, are provided to operate, for example, as rule interpreter 218(2,1) and frame mechanism 218(2,2), as indicated in FIGS. 11 and 14. These knowledge mechanisms 601 are subject to control by the scheduler 218 (FIG. 10) to implement inferencing, forward- and backward-chaining, inheritance, and other implication schemes. Information regarding instances of the weld process is stored as a database in a frame 215(1) (FIG. 13) including values and procedures used to accomplish inferencing procedures. Attached to each frame 215(1), there may be software demons, which are procedures written in Lisp and invoked during the searching or inference processing to accomplish specified tasks, such as gathering data from hardware sensors or initiating communication between particular nodes 115 of the controller 16. Further, according to the invention, the scheduler 218 (FIG. 10) and knowledge mechanisms 601 (FIG. 14) operate upon semantic networks 215(4) (FIG. 13). The semantic networks 215(4) contain the goals and the forward- and backward-chaining rules, some of which launch software demons, and which initiate Lisp inference procedures as stack frames managed by the scheduler and queues 218 (FIG. 10). The scheduler 218 is clocked, according to one version of the invention, at regular interrupt intervals to switch between inference procedures defined by the Lisp stack frames. The scheduler 218 implements object processes or inferences on a selected basis. During inferencing, specific instances of the particular welding process being conducted and particular parameter values of these selected instances are developed in response to the base materials being used, and subject to the particular constraints on the configuration of the equipment used and the specifics of the welding arrangement 13 in fact employed.

The rule interpreter 218(2,1), as suggested in FIG. 11, is one of the knowledge mechanisms 60' indicated in FIG. 9. This rule interpreter 218(2,1) matches the system goals in the semantic network 215(4) to particular sets of rules 215(2) (FIG. 8). The rules then trigger the establishment of new goals in the semantic network 215(4). When a value is needed which does not exist at a particular node, the communications network 150, as indicated in FIG. 9, including the knowledge layer 151, logical layer 152, and physical layer 153, arranges for transfer of the needed information to the node 15 which actually does need it. According to the invention, knowledge mechanisms 601 are thus paired up with corresponding knowledge bases 215, as suggested in FIG. 14.

The purpose of a knowledge mechanism 215 as shown in FIG. 14 attached to a particular knowledge base 215 is to examine the contents of its corresponding knowledge base 215 and to glean from it specific values and information necessary to the accomplishment of system goals and inferences, insofar as any relevant information exists in the particular knowledge base 215. Each knowledge mechanism 601 attached to a knowledge base 215 is further responsible for the processing of various predefined inference processes, and to return an answer resulting from inferencing back to the semantic network 215(4). When the originator of that system goal or task becomes scheduled for operation, it will accordingly note that its assigned work has been completed and that its request has been fulfilled. Frames 215(1) are used because they are particularly effective for the storage of static process knowledge relating specifically, for example, to the particulars of welding procedures. By their nature, frames 215(1) define these particulars as entity relationships which set forth and establish the processes and factors that bear upon the specifics of the welding process. Further, the frames 215(1) define a class structure as to the particular process to be performed, specifying inter alia the particular base welding materials 14 to be used, the particular current levels to be employed in a welding operation, the particular weld travel speeds to be developed during welding, as well as the history of the desired weld in terms of information that has been accumulated from previous welds. According to the invention, frames 215(1) are used for the storage of process knowledge like entities and relationships. Rules and rule sets 215(2) are employed effectively for control and decision making. Semantic nets or networks 215(4) function for matching goals to the previously determined schemes, rules, frames, and objects, and for maintaining the forward- and backward-chaining rules that implement software demons which execute Lisp processes. The semantic network, in effect, establishes a tree structured database of knowledge. The frames 215(1) have slots and values, thereby permitting the employment of tree searching mechanisms which are called inheritance mechanisms. Each frame 215(1) includes software demons which are invoked by the action of accessing particular slots in a frame knowledge base. Information can be inherited from the knowledge representations of superclasses. For example, the class structures established in the semantic network 215(4) can serve to enhance particular feature descriptions of the welding process. Welding operation requires, for example, that the weld current actually be on. This information may not be present on a particular welding frame knowledge representation scheme, but it is, nonetheless, available through the process of inheritance of information from a superclass. To accomplish one version of welding operation, one further needs to specify the particulars of the gas tungsten arc (GTA) process to be employed. Inheritance is made possible because any inference process which needs to determine certain particularities or details can access the GTA welding process definition and can acquire the facts needed, namely for example that, yes, the weld current has to be on in order to perform welding operations, and also, for example, that the shield gas must be turned on.

Rules 215(2), accordingly, become paired up with tasks to be performed and are effective to control these tasks during performance. The semantic networks 215(4) established in the knowledge base 215 are further effective for matching goals and tasks to other knowledge representation schemes present in the system. The semantic networks 215(4) provide the mechanism by which, as discussed, an event can be caused to happen simply by accessing or modifying a particular variable in the system. The semantic net 215(4) itself is essentially a tree structured data base containing simple premises for triggering forward- and backward-chaining operations. The rule syntax for semantic nets is independent of that used in the rule interpreter knowledge mechanism. However, it is the forward-chaining process which establishes use of the demon capability and enables software interrupt capabilities, according to the invention herein.

Further, at the interface to each knowledge base 215, there is a collection of functions employable primarily for searching. Also included at each data base interface is a procedure for accessing the semantic network 215(4). In addition, there are two primary software mechanisms associated with these tree structured semantic networks 215(4). The premise portion of the semantic network structure constitutes a tree structure. The branches in the semantic network 215(4), in fact, contain the actual goals and subgoals to be implemented during operation.

The input of a high level goal such as "welding" may, for example, be accomplished by user entry. If the user actually types in or enters an expression such as "start welding process" through a particular input/output device 117i (FIG. 8), this input expression is entered in the premise portion of the semantic net 215(4) (FIG. 13). This entry is effective to trigger a forward-chaining rule, which has been previously defined in the semantic network 215(4) for this purpose. This, in turn, triggers the scheduler 218 to invoke a rule interpreter 218(2,1) (FIG. 11) to examine the semantic network 215(4) for any related goals which it might be able to process. Simply stated, the assertion of a system goal onto the semantic network 215(4) can cause the scheduler 218 to direct the implementation of inference processing in proper portions of the overall knowledge representation scheme by means of forward-chaining demons. In this particular case, the appropriate action in response to the insertion of a system goal to start welding is to schedule the rule interpreter 218(2,1) for operation in the running queue 218(1) of the scheduler 218. This allows the rule interpreter 218(2,1) to take control and establish a data path between the semantic net 215(4) and the rule interpreter 218(2,1), enabling the rule interpreter 218(2,1) to accept the particular goal entered and to start work on it. Inserting a system goal onto the semantic network 215(4) automatically triggers the scheduling of appropriate software processes to achieve the goal desired. The semantic network 215(4) and mechanism 218(2,3) function as a system interpreter having data base access capabilities and, additionally, the ability to trigger selected process modifications, additions, or deletions. The notion of using a semantic net 215(4) to drive the scheduling of processes, according to the invention herein, is unique. Tying the scheduler to the semantic network establishes a unique scheme of data driven scheduling, according to the invention herein.

Object processes are maintained in operation on one of three queues, respectively, 218(1), 218(2), and 218(3) established in the scheduler 218. These queues include a suspended queue 281(3) for processes that are viable but not immediately operable. For example, if a process makes a particular request for information, that process is suspended until the answer is provided from some other node. There is, additionally, a running queue 218(1) maintained by the system process scheduler which handles only one process at a time because the Lisp processor can, of course, do no more than that. Once a particular running process has been accomplished, its processing terminates or is suspended and entered into the suspended queue 218(3). Each process is given a particular predetermined amount of time to operate on the running queue 218(1). Particular processes can finish or terminate during the assigned running period allocated. They can come to a point where they need to be suspended or they can continue running until the end of their allotted time, and, as such, since they still need to be worked on at some point, they are not suspended but are directed to the ready queue 218(2). Accordingly, the process can be returned later to the running 218(1) queue for continued, subsequent processing. Accordingly, a particular unfinished process either goes on to being performed or it is held suspended.

This arrangement of the invention is clearly data driven because if a goal is provided to the system, it need not be independently scheduled--it schedules itself automatically. Thus, a system goal is a data driven thing which will trigger its own scheduling.

In view of the established three queues in the scheduler 218, there are correspondingly three states in which an inference process can find itself. For example, it can find itself in the running queue 218(1), which is a queue that can handle only one inference at a time. It can find itself in a ready queue 218(2) of pending processes which are not quite complete but which have not been suspended. These processes are still being conducted and when their turn comes up to be performed in the running queue 218(1), they will be processed for an additional predetermined period of time, thus incrementally coming closer to ultimate completion. Additionally, as noted above, there is a suspended queue 218(3) which holds processes that are in a suspended state of operation. The key aspect of being on the suspended queue 218(3) is the ability to come out of suspension once the answer to the particular request for information has been returned. This waking up of a process is accomplished by a particular software demon embedded in the semantic network 215(4). The scheduler 218 itself directs the selection at each next running process, and particular processes can become the next running process because the scheduler 218 determines it is to be the next one to be scheduled, whether it is a currently pending process, or a process which has been suspended but now awakened with new information.

Between each set of hardware nodes 115 there are a pair of communication ports 115p which require a communication schedule for each of the ports 115p. All the communication processes that are to take place need to be suitably scheduled according to a context switcher 218", as suggested in FIG. 17, which operates at a selected switching frequency in response to the control breakhandler 717 of the Lisp system. The context switcher 218", when activated, checks the semaphore 717' (FIG. 1a) by a call to interrupt 64H to determine whether to cause the inference scheduler 218 to promote the first inference process in the ready queue to the running queue or whether to initiate message processing, as has been explained.

According to the invention, in communicating messages at a low level through a number of ports 115p, both read and write messages need to be implemented for transmission. This need to schedule read and write messages requires the establishment of the context switcher 218". According to the invention, there is further established a queue 734, as suggested in FIG. 16, for maintaining the most important messages that have been received, such as those which generate inference interrupts through the control break handler 717, for example. Further, there is a need to schedule how often messages on this particular queue 734 will be handled for disposition. Accordingly, a control break interrupt, resulting from such a message, intervenes in the scheduled operation of the ready queue 218. The inference interrupt feature of such messages is typically utilized to insert new system goals at a particularly high level. An inference interrupt is indicated by a specific key in the message itself. When this key is noted, it forces the context switch 218" of the Lisp scheduler 218 to add the message directly to the running queue 219(1) for immediate processing. Other kinds of messages are generally blackboard read or write messages. These are regularly scheduled, without interruption. Inference interrupts are scheduled immediately upon receipt since they need to be performed quickly. These require lengthy processing and need to be scheduled much sooner than the particular software routine that is merely performing the reading of the messages. The system, accordingly, includes multiple reads to accommodate each port handling its own associated buffer mechanism. The port buffer 739 associated with each port 115p is large enough to permit scheduling. The invention, herein, thus permits suitable scheduling according to dynamic priority assignment techniques, as indicated.

The priority of a process can be based on a number of considerations, including the inherent priority of a particular process and how much data has been received by the port, and stored in the internal buffer which, for example, is represented as a level indicator. According to the invention, there is an intermediate level at which messages come in. When the full message arrives, the information received can be loaded into the queue 739' associated with that port, so there is a high level way to figure out what your current message level is. This is done by looking at an internal queue (not shown) which maintains messages. This limits the scheduling of processes to only the first three processes available. According to context switching, the overall current clock frequency is monitored and modifiable. In fact, the context switch 218" can be held frozen, to stay at a particular context for a selected, extendable period of time. This enables the elimination of inappropriate processes which might happen with data transmission from one port subject to error, perhaps, as a result of a machine crash, for example. It, accordingly, makes little sense to complete the transmission. Thus, the first port 115p is abandoned completely and the particular process being performed is suspended indefinitely. On the other hand, it might be desirable to temporarily freeze such a process because a great deal of attention might need to be directed toward other processes. Thus, instead of subjecting it to suspension, the process could just be delayed in its implementation.

The process scheduler software shown in Appendix VII(A) performs multi-tasking on inference processes defined by frames, rules, semantic nets, and objects. Appendix VII(A) includes several functions which are invoked by the software interrupt in view of the clock tick scheduler established in the Pascal communication layer, which is invoked every eighteenth of a second. Appendix VII(A) particularly shows Pascal code communication routines that are used in the physical layer 152 (FIG. 19), including a message handler routine which prioritizes the existing messages defining inference processes to be scheduled, and selects the highest priority process for submission via the inference interrupt mechanism to the logical layer 152.

Communication processes are performed by Pascal and Lisp code, as suggested in Appendices VII(A) and VII(B). In order to invoke a Lisp process, a control break interrupt is generated in Pascal, which forces the Lisp control break handler to determine which process to run next. Among communication processes are read and write processes which can be performed with respect to the ports 115p. Each of these ports 115p is controlled and maintained by the Pascal queue management software, as indicated in Appendix VII(B). The general manager routine maintains a queue of messages and data relevant to those messages, and at selected times, the messages on that queue are "de-queued." Additionally, the memory addresses and lengths of the messages are separately transmitted by a different routine.

Communications can be conducted on the knowledge layer 151 (FIG. 9) which resides primarily on the semantic network in Lisp. The knowledge layer 151 is a set of inference rules that form a grammar to enable translation from a particular knowledge representation scheme to a general language. Processes communicate with each other in general language. The general language might, for example, call for the production of a particular value corresponding to an attribute in the frames' knowledge representation scheme. This general language representation is translated by the knowledge layer into a specific function which actually fetches the value from a particular slot in a frame 215(1) (FIG. 13). There is, accordingly, a specific language feature in each semantic network that corresponds to the general idea of fetching something which is implemented by a backward-chaining rule that triggers a software demon in the semantic network 215(4) (FIG. 13).

The knowledge layer 151 thus comprises a set of transformation procedures that perform the translation of a generic request for data from a particular knowledge base in terms of a specific syntax which calls out information from the particular knowledge mechanism. The knowledge mechanism 215 is a small entity means by which to access a knowledge representation. It is, for example, a query handler and a knowledge base maintainer. A query is simply a request for information containing variables which need to be matched to one or more elements in the knowledge base.

The semantic network 215(4) includes portions of the knowledge layer 151 used for the translation of specific requests into a form needed by the particular process to enable a response process which makes sense as a logical connection from one process to another. The semantic network 215(4) also handles the syntax of that interprocess communication. If it turns out that the process does not reside on a particular node or machine that needs to be communicated with, then the information is communicated through the logical layer 152 down through the physical layer 153 and out through a selected communications port. Accordingly, the semantic network 215(4) encompasses both portions of the knowledge layer 151 and the logical layer 152.

The semantic network 215(4) is a discrimination tree data structure which is composed of nodes and links between the nodes. Information can be stored at these nodes in the form of goals and rules, for example. When rules, goals, or similar mechanisms are entered into the discrimination tree, it becomes a semantic network. Inferencing is a search process for either a conclusion or for particular premises. In the case of inferencing a semantic network, the discrimination tree is searched or a value is entered into the discrimination tree. Appendix IV illustrates how the discrimination tree is implemented to create links and nodes. In particular, it sets forth the definition of a predicate calculus frame weld. A predicate calculus variable is a placeholder variable which may have a value, and it is used in a quantitative sense across a goal expression. For example, the statement, "For all "X", where "X" is a weld puddle, "X" is 0.6", is a statement of the kind used in a semantic network. All structures represented in the semantic network have quantified variables, with respect to which unification is performed. Unification is a way of resolving conflicts. If there is a desire to bind variable "X" to the numbers, "1, 2 and 3", for example, and another procedure seeks to bind "X" to the numbers, "1 and 2", both procedures aim to bind "X" to "1 and 2". This is an example of the notion of unification. Appendix IV further illustrates several basic access functions that index information, effectively putting it into the discrimination tree data base, or looking up information in the database and building the necessary links in the data base, if they don't already exist. The file name index list contains some additional utility functions which create and use databases. Unindex, for example, is a way to retract information from a database such as a discrimination tree.

Appendix IV further shows a file chain which illustrates how a software demon can be implemented according to the invention. During the search process through the discrimination tree, a node may be discovered that contains a Lisp program. If a Lisp program is encountered, then it is immediately evaluated since it is a software demon. The functions, "chain" and "add" in Appendix IV, are further examples of software demons implemented according to the invention herein. The unification algorithm, in particular, contains a recursive check according to which patterns are being matched. They may themselves have variables in them which would require a complicated procedure to fully resolve the findings. This procedure requires recursion. Appendix IV further shows a procedure to rename variables in the course of search. The basic search function shown may invoke backward-chaining if the information sought is not directly discovered. The search procedure chains backwards to find additional premises to support the particular conclusion indicated. Further, there is shown a system function which shows the database and prints it. Data structure is defined as self-maintaining objects. If one sends a command to an object or sends a request to the object, the object structure has mechanisms for evaluating the request or command. One example of an object is the welding power supply. It is treated, conceptionally, as an object which can receive messages. You can, for example, send the welding power supply a message such as "set the weld current to 10 amps". The power supply definition is part of the hierarchy of equipment in the factory. Accordingly, the message which goes out to the object welding power supply may inherit some procedures. Simply stated, objects have procedures and information. The object acts on that information and also informs other procedures that are related to this object. With regard to frames, information is requested from them or information is put into them in a static fashion. Of course, other frames may be accessed to inherit information but there are no inherent procedures to be actuated. A frame is a list plus the structure with attributes. According to the invention, there are from one to twenty frames to use.

The knowledge layer 151, on the other hand, is simply a language translator. The logical layer 152 is a location translator, and the physical layer 153 enables the actual translation from software to hardware and from machine to machine. Furthermore, the semantic network 215(4) is a structure which holds goals that need to be satisfied by information in knowledge base such as a frame, for example, frame 215(1). In different parts of the frame, there is information regarding different goals to be satisfied. Thus, the semantic network 215(4) communicates with the various knowledge representations.

The semantic network 215(4) further relates entities and attributes as a knowledge base. In particular, the semantic network 215(4) uses pointers to establish the relationship between entities and attributes.

The program, which actually accesses a particular knowledge base, is called the knowledge mechanism 601. This mechanism 601 transforms user queries into particular programs which are used to search a particular knowledge base for answers. In the course of searching the database, the semantic network triggers software demons just by virtue of accessing a particular attribute or value. A query is thus automatically converted into a search technique to go through the entire frame 215(1). If the search does not produce an answer to the particular question, another demon program is invoked to find, for example, particular bounds within which the answer must fall according to information in the particular database. Demons are typically attached to a slot in the structure of frame, but may also be triggered by rules in the semantic network 215(4). The knowledge layer 151 takes a simplified natural language and translates it into a small expression that initiates a search of a particular corresponding knowledge base.

Each node of controller 16 acquires the semantic network 215(4) at initialization. The controller 16 will then create inference processes from the rule 215(2) (FIG. 13) to implement the goals in the semantic network and places the processes into the scheduled queues. As goals are presented by the user to the network 215(4), each node of controller 16 acquires the portions of the overall knowledge base needed to perform the particular goals. The nodes 115 have no specific goals within them at start-up time. It is up to the user or some other program to provide the controller 16 with selected goals to achieve.

The semantic network 215(4) is, accordingly, just a framework driven by goals. The semantic network 215(4) insures that these goals are transferred to a node which can achieve them. The rules 215(2) themselves further have information which is not included in the semantic network 215(4), and the semantic networks 215(4) establish a virtual link between the rules 215(2) and all the other knowledge representations in the controller 16. The semantic network 215(4) thus enables semantic translations between different information representations.

Appendix VII(A) contains the communication routines which are part of the knowledge layer 151. These routines are independently loaded at each of the nodes 115 during initialization. The routines specifically include a control break procedure 718, which is scheduled as specifically indicated in Appendices VII(A) and VII(B), with a low-level Pascal clock service 719, as indicated in FIG. 16. This service 719 schedules particular hardware and software services for the input and output ports 115p, as well as perform context switching and servicing of the message queues. Control break is invoked by hardware interrupt (hex)lB every eighteenth of a second. Scheduling is otherwise performed based on priorities established in the ready process queue 218(2).

Appendix VII(A) further shows how non-critical messages arrive at the low-level blackboard 742, as indicated in FIG. 16, while "important" messages are scheduled in an important message queue 734 for reading at a suitable time. Reading is not implemented immediately because that would throw off system timing. Accordingly, messages are scheduled for reading into the ready queue 218(2) if there is a message to be read. In addition, the routines in Appendix VII(B) include procedures for writing with respect to particular ports. Write port 740, moreover, involves a procedure very similar to control break, in that both procedures turn output queue 736 on and off. Appendix VII(B) further discloses how the output queue 736 is filled with messages that are to be dispatched. Output queue 736 is filled by Lisp. The write port 740 monitors the output queue 736 and sends out the messages through its associated hardware as required. Output queue 736 itself has a variable level, and repeatedly picks up information and parcels out messages.

Appendix VII(B) shows the procedures for reading incoming images through a synchronous port buffer 739. Incoming messages are processible only if a busy flag (not shown) associated with the buffer 739 is "false". On the other hand, if the busy flag indicates "true", the routine blocks the entry of another routine. A typical message arrives header first, indicating its length and identification, and providing some error checking information. Next comes the text of the message itself.

To process a message, a high priority control break is asserted by context switcher 218". Additionally shown are blackboard 742 writes and reads. A blackboard "read" simply de-queues the normal blackboard queue 742, as indicated. The write port routine selects a port for transmission and ascertains where the message is and how long it is. Thus, to invoke a write port 740, one must indicate which port is selected for transmission, the location and memory of the message to be sent, and the length of that message.

The blackboard 742 is a partial reflection of the semantic net. It is a subset of the semantic net reduced to Pascal, permitting information with another computer to be repeatedly and rapidly accessed in Pascal without having to go to Lisp. Accordingly, some information is simply reflected down into Pascal from the semantic network in order to avoid the lengthy process of transferring it directly from Lisp. The kind of information involved includes triplets, variables, attributes, and values. Triplets are three values in a set, namely value, name, and attribute. This is a technique for short-circuiting the control break message handler to access values in the machine without going through extra software layers. The blackboard 742 is essentially a status board, as indicated in FIG. 9.

Appendices VII(A) and VII(B), there is shown a queue manager 739' for the read port 741. This queue manager 739' holds both control and non-control break messages. Thus, there are two kinds of queues: control break 734 and non-control break 742. Further, a determination is additionally made whether it is safe to intervene without interrupting anything in operation. Every port on the machine is checked for operation. The output queue 736 is de-queued of any messages which are pending for dispatch and its contents are sent to the write port 740.

In introducing new knowledge into the knowledge base, a check is made to insure that the new knowledge does not conflict with other knowledge already in the knowledge base. Toward this end, a consistency checker, as set forth in Appendix V, isolates any knowledge which is redundant or which weakens the integrity of the knowledge base. Appendix V further shows the rule base 215(2) for the consistency checker. These rules check both redundancy and syntax, which is essentially the format of the information. The syntax of rule base 215(2) is particularly defined in rules 6, 7, and 8. These rules are called to determine precisely what the syntax of the incoming information actually is. The remainder of the rules perform actual logical consistency checking, including both simple and complex redundancy. A simple redundant rule, for example, involves the information that is repeated exactly. However, some forms of redundancy are more difficult to spot.

Processes are placed on the suspended queue 218(3), as suggested in FIG. 12, for example, in order to enable the generation and transmission of a message to another node. The event which returns the suspended process to the ready queue 218(2) is the arrival of an answer to a requested message. This concept of ready, running, and suspended status for inferencing processes is unique as applied to the performance of distributed process control.

According to the communications scheme set forth in FIG. 15, a message is received according to block 940 at a particular port, such as port 115p, for example, and is stored at a low-level in a selected software mechanism, according to block 950, which forwards the information received without modification to the physical layer 153, as indicated at block 960. The physical layer 153 is a low-level message processing device which parses a message into its individual components and makes a preliminary determination as to its context, according to block 508. Establishing context means, as suggested at block 508, determining the type of message being received, the sender, and the origin of the message. It may also include a destination identifier which specifies the specific computer node 115 to receive the particular information.

Based on the context of the message, most information need not interrupt any higher level process. The logical layer 152 of the arrangement herein, in fact, maintains an organizational structure which insures that most processing is maintained at the lowest possible level. This layer 152 has enough context knowledge to understand whether a message received is related to a particular local blackboard 741, to another node, or to machine 115. Such a blackboard-type message simply maintains information locally at a local process blackboard 741 for reference as needed by other local processes. Blackboard messages, according to the invention, include read-type messages which find a particular value on the blackboard 741 answering a request for transmission back out to the originator of the message.

As already noted, the weld process begins with the acquisition of infrared data regarding weld puddle features and geometry. Another type of message is employed, which is interrupt driven and effects significant changes to the process currently in force or operation. Interrupt messages are, for example, employed to start a new process or to delete a scheduled process. Thus, a high level request is performed to implement the interrupt process. As suggested in FIG. 21, once IR information regarding weld puddle features has been attained, selected "C" algorithms 711, preferably written in "C" language, perform low-level processing of the sensory data to extract desired weld features. At the Lisp level, which is high-level and knowledge-based, these features are analyzed as suggested at block 709 to achieve a current model of the weld puddle in terms of a high-level formulation. This enables finding the maximum and minimum axis or width of the puddle.

A software process on one node or machine 115 can acquire information from another node or machine, by translation of a request for the information through the three layers of the communication link. In a frame system, to determine the value of a particular parameter, a predetermined function is performed which searches through the frame database for substantiation of that value. If no value for the particular parameter is found in any slot of the frame searched and the absence of value is recognized, a message indicative of that absence is returned.

The inference multiplexing or scheduler mechanism 218 operates three queues, as represented in FIG. 11. These queues include first a running queue 218(1) which handles only a single process on the CPU at a particular time. Second, there is also a ready queue 218(2) which handles a plurality of processes ready for execution. Finally, there is a suspended queue 218(3) which holds processes which have been suspended. The ready queue 218(2) holds processes which have all the necessary resources to run immediately but are simply waiting for their slice of time to run. Processes can be transferred between queues by a context or process switcher, which acts as a scheduler 218 to perform transitions between the queues.

According to the invention, the ready queue 218(2) holds a number of processes including, but not limited to, the rule interpreter 218(2,1), the frames mechanism 218(2,2), the semantic network mechanism, and two types of message handlers, respectively 218(2,4) and 218(2,5). These processes are on the ready queue 218(21) at all times so that during successive time slices, successive ones of these processes come into play and portions of these processes are executed until they lose their allocated time slice. According to the invention, when a particular running process suddenly needs additional resources which it does not already have, it becomes suspended or blocked, and is transferred into the suspended queue 218(3). Processes on the suspended queue 218(3) wait for the occurrence of a specific event which enables the process to return to the ready queue 218(2) for multiplexing onto the running queue 218(1) in its regularly allocated turn, as suggested in FIG. 12. As shown in Appendices VII(A) and VII(B), which contains the software for queue handling, a queue is a data structure having a list and a position indicator for the beginning and end of that list. In particular, Appendix VII(A) shows the software establishing the data structure of the ready queue, the running queue, and the suspended queue. An object is a portion of the knowledge base which contains an abstraction of knowledge.

The rule set 215(2) includes, according to a preferred revision of the invention, a statement initially setting certain variables to "nil". This is accomplished by the Lisp expression:

Simply stated, this establishes that the time change DT is initially zero, that the weld current change DELTA-I is initially zero, and that the speed change DELTA-S is initially zero. Further, certain tasks are defined by the expression:

In other words, it is the goal of a specific task or inference process to control puddle width for a selected robot and power supply. The expression, "static robot", is a variable expression which holds the name of the robot device on which the weld process is being performed.

Additionally, to set the "FL" mode of operation, a mode selection expression SELMODE FL is employed. Further, CSET is initialized at nil value. Beyond this, a prerequisite is established according to the statement:

Additionally, entry is set in accordance with the statement:

Finally, inputs are defined as parameters width, current, speed, and shield gas temperature.

As suggested in Appendix IX, a first rule of the rule set 215(2), employed according to a preferred version of the invention, evaluates whether the shield gas temperature is greater than the dew point temperature. More particularly, Appendices III and IX are codes for the expert knowledge representation scheme including the rule interpreter of the expert systems of the invention herein. More particularly, the first rule examines the current value of the shield gas dew point temperature and compares it to a minimum acceptable dew point temperature. In case the shielding gas dew point temperature is greater than this minimum acceptable value, the welding process aborts. More particularly, the inference process interpreting the first rule is aborted, and a new "above dew point" process is initiated. This new process shuts off the welder. The second rule evaluates whether the parameter puddle width is less than the constant WMINP, i.e. minimum puddle width. If this condition is met, the too small puddle routine is called. The second rule, accordingly, compares the current value of the puddle width to minimum and maximum permissible puddle widths. Further, the current is compared to a predetermined minimum and, in case it is too small, the inference process implementing the rule is aborted and a new process called "too small puddle" is initiated. According to the third rule, the current puddle width is compared to a maximum permissible value, and in case the puddle is too wide, a process named "too wide puddle" is invoked. More particularly, the third rule evaluates whether the parameter puddle width is greater than the constant WMAXP, i.e., the minimum puddle width. If so, "too wide puddle" is called. The fourth rule compares the actual value of weld current to a predetermined desired current level and further compares measured puddle width with a predetermined puddle width constant. Further, the trend of the change in puddle width is observed. Simply stated, the previous value and current values of puddle width are compared with a maximum difference value to indicate the rate of change in puddle width. Further, the current value of the weld current is compared to a desired change that we wish to make in the current and, if there is a particular difference level, the weld current is modified by this value "di". According to the fifth rule, the weld current is compared to a predetermined constant current set point Further, the puddle width is compared with a predetermined puddle width constant, indicative of the maximum acceptable puddle width. Finally, the welder speed change is compared to a maximum speed change constant. If the weld current is less than 95% of a predetermined maximum current level, the puddle width is greater than the maximum allowed puddle width within a particular tolerance zone or, if the speed change is greater than a predetermined maximum speed change, then a predetermined inference process is initiated which attempts to change the speed of the robot by a predetermined value "ds" and also to change the weld power supply current by a predetermined value "di". The control system monitors weld speed and current directly in order to maintain a constant puddle width. According to the ninth rule, weld speed is controlled as indicated. The tenth and eleventh rules, respectively, mirror the eighth and ninth. Accordingly, by conducting a suitable trend analysis, an intelligent determination is made as to how much to change the weld speed or current by using knowledge of the trend of the puddle growth. Further, the sixteenth and seventeenth rules perform a comparison of the minimum controllable puddle width trend with respect to the existing puddle width. In particular, the trend of the last value is compared to the trend of the derivative of the puddle width change and the result is compared to an established constant called D-High, which relates to the greatest rate at which the puddle width can change. Based on these comparisons, a change is made in the weld speed. Further rules, as suggested in the Appendix, evaluate whether the weld current is below 0.95 of a constant "IQ" and whether puddle width is below constant WMAXC. If so, the routine "too low current" is invoked. Further, if the weld current is less than 0.95 of "IQ" and width is greater than WMAXC, and global delta-S is defined as delta-S high, then changes in weld speed current are invoked. If the weld current is greater than 1.05 of "IQ", and puddle width is greater than WMINC, then the process "too high current" is invoked.

During inferencing operation the rules indicated above are scanned from beginning to end to determine whether the antecedents of any of the rules are satisfied. If the antecedent of a rule is satisfied, its consequence is triggered and the search for a next rule to trigger continues.

The welder arrangement 13, according to the invention, herein further employs a process controller 16 operating according to an Advisable Process Control Language (APCL), as set forth syntactically in Appendix X, and which is an adaptation of Lisp.

Appendix IX is an extended formal definition of the entire process controller architecture. It shows, recursively, the underlying idea of the process controller 16, including advisors 115ad linked to other advisors 115ad and eventually linked to an expert controller, which in turn is linked to actors 115a which are connected to sensors 116 and/or controllers. Further, the expert controller 115e may be directly connected to sensors 116 and/or to custom interfaces, such as, for example, auxiliary controller 115. "XADVAC" is defined as an expert controller combined with any number of advisors or actors. An expert controller, moreover, is an advisor combined with a physical layer. An advisor 115ad multiple inference expert system has a knowledge layer 124, a logical layer 129, a physical layer 130, and may have actor component custom features, such as an interface to a sensor.

An actor 115a includes an application layer of programs written specifically to interface to a sensor 116, controllable device 117, or actuator. The actuators must also contain knowledge, logical, and physical layers, respectively 124, 129 and 130, so that they can be components of the entire control structure 16. An application layer is a program combined with some application specific knowledge, such as an image processing information, for example. A physical layer 130 is a program with specific knowledge of how to transmit messages according to a particular message protocol. A logical layer 129 is a program which also contains logical destination knowledge. A knowledge layer 124 is a program layer containing a translation scheme between knowledge representation schemes. A program is obviously a connected series of statements to solve a problem. The total of all programming languages in our system is "APCL", "C", "Pascal" or "Fortran" and assembly language.

The process controller 16 can be formally defined in a symbolic fashion in terms of its components. According to this syntax, the process controller 16 for directing welding operation can be defined to include an expert controller and zero or more occurrences of advisor or actor elements, respectively 115ad ad 115a. Further, each expert controller 115e can be defined as the combination of an advisor 115ad and a physical layer. Additionally, the advisor 115ad itself can be defined as a multi-inferencing expert (MIX) system together with a knowledge layer 124, a logical layer 129, a physical layer 130, and zero or more occurrences of an actor 115a. An actor 115a, in turn, is defined as an application layer together with a knowledge layer 129, a logical layer 129, and a physical layer 130. Further, the application layer is defined as a layer together with application specific knowledge. Similarly, the physical layer 130 is defined as a layer together with data transmission specific knowledge. A logical layer 129 is defined as a layer together with a logical destination knowledge. A knowledge layer is defined as a layer together with a knowledge translator. Additionally, a layer is defined as a program together with a programming language. A program, in turn, is defined as a connected series of statements to solve a problem. A programming language, in turn, is defined as either the advisable programming control language (APCL), "C", Pascal, the assembler or the like.

The advisable process control language (APCL) can be defined in terms of a formal grammar. For example, a rule set in the advisable process control language is defined as a rule set label and zero or more occurrences of a static declaration, global declaration, local declaration, task expressions, inputs, mask expressions, prerequisites, or conflict set together with a selection mode and zero or more occurrences of rules. A rule set label is an identifier. A static declaration is the word "static" with zero or more occurrences of variable declarations. A global declaration is the word "global" with zero or more viable declarations. A local declaration is the word "locals" with zero or more instances of a rule label with variable declarations. A selection mode is the word "Selmode" with a key word such as "FL", "FD", or "TARD", for example. A prerequisite is the word "Prerequisite" with a Boolean expression. A task expression is the word "task" with zero or more Lisp expressions. A rule itself is a rule label with an "if" expression with zero or more occurrences of "Then" expressions. A rule label is an identifier. If and Then expressions are Boolean expressions.

A variable declaration is an identifier with zero or more occurrences of an initial value. Additionally, Appendix I is the software code for automatically loading files at power-up in order of loading; Appendix II contains the software code for the utility Lisp functions; Appendix III contains the software code for the frame knowledge representation system, according to the invention herein; Appendix IV is the software code for the semantic net knowledge representation system; Appendix V is the code for the expert knowledge representation scheme including the rule interpreter of the expert systems of the invention herein; Appendix VI shows the software code for consistency checking, as discussed above; Appendix VII(A) is the Lisp/Pascal communication and process scheduler software interface; Appendix VII(B) is the Pascal communication and scheduling interface for serial ports, hardware clock, and Lisp; Appendix VIII is the "C" image processing code including the algorithm for resistance weldbox size, according to the invention herein; Appendix IX shows the example knowledge basis for frames, the expert system, and the consistency checker; Appendix X shows the extended Backus-Naur formal definition of expert advisable adaptive controller, according to the invention herein; and Appendix XI is the index to the Appendices.

Finally, FIG. 20 shows the multi-nodal controller 16, according to the invention herein. This controller 16 is effectively both data-driven by information received in real time from sensors 116 and/or controlled elements 177, and goal driven by goals entered in the semantic networks 215(4) of the various nodes 115 of the controller 16. Several nodes 115 may reside on a single machine. The knowledge bases 215 including rule 215(2) are transparently distributed throughout the various nodes 115. Thus, inferencing processes conducted by the controller 16 are effectively multi-tasked at many levels. For example, inferencing can be conducted at the level of the root node or expert ccntroller 115e at any of the various levels of particular actor nodes 115a, which are connected either directly to root node 115e or through advisor node 115ad thereto. Further, inferencing can be conducted at any of the advisor nodes 115ad indicated, irrespective of level.

By way of additional detail, the invention herein includes real time welding of adjacent materials with a weld puddle formed between the materials by bringing the welding apparatus into position to form the weld puddle at the interface of the adjacent materials, thermographically scanning the weld puddle and its adjacent heat affected zone, and producing image signals representative of a thermographic image of the weld, transferring the thermographic puddle image signals for comparison with reference puddle images to evaluate variations, and adjusting the welding apparatus to change the weld puddle because of deviations noted between the detected and reference puddle images. The process includes adjusting the electrical power and position of the welding apparatus torch. It further includes determining whether adjustment of torch current will modify puddle the shape to a predetermined range, and adjusting the torch travel speed when adjusting the current is not sufficient. The process includes determining whether a predetermined amperage change, such as ten percent or less, will adjust the width of the puddle to within a predetermined puddle range and, if not, then determining the speed adjustment necessary for the torch to change the width of the puddle sufficiently to cause it to fall within the predetermined range, and then changing the speed of torch travel accordingly. The procedure includes storing process and procedure variables in real time for later analysis of welds made at a later point in time. The process further includes inspecting the weld puddle for defects in real time. The process further includes adjusting voltage or amperage, lateral adjustments of welding torch position, speed of torch travel, and angular adjustment to a track angle. The procedure further includes taking isotherm line scans of the weld puddle and relating the isotherms in real time to weld size measurements. The procedure includes real time analysis of isotherm scans of weld size as a fraction of weld current and tensile shear versus weld current. The procedure further includes real time determination of width and depth of the weld puddle fusion zone. The procedure further includes determining shield gas temperature in the torch and aborting welding operation when the gas temperature is too high. The process further includes analyzing the forward tip of the image of the weld puddle for projections indicating deviation of weld path from seam or interface direction and adjusting the path of the tip to more closely track the weld seam.

The invention herein is further directed toward extracting data indicative of irregularities from the image signals, including the width of the weld puddle as, for example, measured perpendicular to the path of travel through the center of brightness of the puddle to points on the puddle perimeter of lower brightness than the brightness at the center of brightness. The data extracted includes the rate of heat flow out of the puddle as, for example, measured by determining the rate of change in puddle brightness at points on the puddle perimeter, particularly the maximum rate of change. The data further includes the rate of heat flow out of the puddle determined by taking the reciprocal of the distance between two points at predetermined, respective higher and lower brightness levels located on either side of the puddle perimeter. The data extracted further includes the rate of change in puddle brightness at its perimeter at a point defined by the perimeter and a line perpendicular to the direction of travel passing through the center of brightness. The data includes the number of times the actual puddle perimeter crosses over a theoretical smooth puddle perimeter within a given zone, the zone, for example, enclosing the leading edge of the puddle, and specifically, for example, enclosing a portion of the puddle perimeter to the left or right of the puddle center. The data, for example, includes measurements of the area of the puddle within a zone overlapping its perimeter, and, for example, measuring the length of the puddle perimeter. The data may further include measurement of the length of the puddle perimeter within a zone that overlaps its perimeter.

The above description of the invention, taken together with the detailed listings set forth in the appendices, sets forth the structure of the invention. The inventive concepts which underlie the realization of a real time, multitasking, distributed process control system using artificial intelligence techniques, can be illustrated by presenting the following over-simplified example of how the invention can be used to implement one simple process-control function. In actual use, the invention would usually be used to implement far more complex functions than the simple one described below. However, a simple example, howsoever over-simplified, will illustrate how real time control of a process can be realized in practice, and the system design principles illustrated may be readily extended to far more complex systems containing many rule sets that define highly elaborate process control strategies.

Our task is a simple one: start the welding process, and then shut it down if the puddle width becomes too large or too small. The machine configuration is in accordance with FIGS. 8 and 19, both of which reveal a detector 17 and radiometer 15 feeding infrared visual signals into an image analyzer 16(1) programmed in "C" and in Pascal. The "C" programs in the analyzer 16(1) compute the weld puddle width, in the manner explained elsewhere, and present this width on a blackboard 742 from which it may be retrieved by the Pascal physical layer 153 within the image analyzer 16(1) (see these elements within the advisor 115 in FIG. 19). The image analyzer 16(1) is serially connected to an image processing advisor 115i, and the advisor 115 is serially connected to a controller actor 115a (FIG. 19). For simplicity, we leave out of this example the expert controller 115e, since our control action (starting or stopping the welding process) is fairly trivial. The controller actor 115a contains programs written in "C" and callable from Lisp that can start and stop the welding power supply and robot, initialize the image scanning equipment, and in general, run the real time process under the control of Lisp in the controller actor 115a.

First constants and parameters used in the image analysis system must be defined. For this purpose, a frame named GTA1 (for "Gas Tungsten Arc 1") is created. This frame might define, in general, such things as the desired welding current and speed, the optimum puddle width, the maximum and minimus puddle width, and it might also define a "deadband" about the optimum puddle width--if the actual width falls within the "deadband" width, then the width is assumed to be acceptable. For example:

of these, only the minimum and maximum allowable puddle widths will be used in this simplified example. This frame, stored in the logical layer of any one node 115, would automatically be made available to the adjacent nodes 115 by the physical layer 153, which would route any requests from the logical layer 152 towards the machine that contained this frame in the preferred embodiment, each node 115 is an AT-class processor, and such data can only be retrieved from an adjoining node; but a more general implementation of the invention could have plural nodes share a single processor and could permit a node to retrieve data from a nonadjoining node. The physical layers 153 automatically advise each other of what information they may access, so any given physical layer 153 knows which way (that is, over which of the two possible serial ports) to route any request for information. Accordingly, the logical layer 152 does not have to know where any given piece of information resides. However, a frame containing constant data might be duplicated within the logical layers 152 of several nodes to avoid the necessity of Lisp-to-Lisp communication.

Next, the parameters of the process, such as puddle-width, current, and speed, are defined as system parameters in the semantic network, and also into the consistency checker list of variables, so that their use in newly-entered rules will be permitted by the consistency checking code within the logical layers. These are entered into the blackboard, and their existence is again propagated by the physical layer 153 to adjacent machines.

Next, any software demons that need to reside in frames are created and entered. For example, if the puddle width were a frame value, the frame would contain a software demon (a Lisp procedure) that would fetch the value of the puddle width from the physical layer whenever it is required. But in our example, only static variables are contained in the frames, and puddle width is an entity in the semantic network that is permissable and may or may not be defined.

The forward- and backward-chaining rules now need to be entered into the semantic network. We need a backward-chaining rule to find the puddle width, whenever it is undefined:

Finally, the rule set that governs system initialization is defined:

When the inference task name "initiate image processor" arises, this single rule always executes (since it is defined to be "t" or "true"), and it calls upon Lisp to execute the named subroutine or program, which is written in "C".

Next, the image processing advisor needs to be set up. We duplicate the frame GTA-1 here, since the frame contains constant data. Then we create three rule sets:

This rule defines the actions to be taken in response to the "startup" goal. Both rules are executed unconditionally once. The first rule performs the functions indicated, while the second rule un-schedules and terminates the "shutdown system" inference task, if it is active.

This rule is executed repeatedly, aborting only when the "abort" command is executed, after which it initiates the "shutdown system" task. This happens when the variable parameter "puddle-width" is above WMAXP, the maximum width, or below WMINP, the parameter minimum width.

When the system is to be shut down, this ruleset initiates the transmission of messages to shut down the robot and power supply.

Next, demons need to be added to the semantic network of the expert controller 115e:

This first forward-chaining demon translates any command to "initialize" something into a Lisp command to send a message "something" over the network.

This forward chaining demon converts a command to "start" something (assumedly an inference process) into a Lisp procedure that causes the named task to be executed immediately, preempting any other scheduled task that may be running.

This backward-chaining rule in the semantic network initiates a search for the "puddle width" parameter, using the "find" procedure set forth below:

This transforms any "find (query)" request into a message sent onto the physical layer seeking the query value, and then the inference task is suspended ("suspend-this-query") until a reply is received. The message contains a unique id code. The message, itself, is a request that the node receiving the message send back a return message containing the query and the id code. This rule also creates a "wake up demon" having the id code that will unsuspend the inference task when the message returns with the query. When unsuspended, the inference task executes the "retrieve query" function.

To initiate the process control task, one simply provides the goal "start welding process" to the expert controller.

The process controller actor 115a, in general, contains frames defining the process control parameters and demons defining actions to be taken relating to these parameters. A simple frame relating symbols to actions that must be taken when those symbols appear suffices for many control actions. Alternatively, rule bases may be provided to define actions. In this example, two rule bases are provided:

This ruleset defines messages that are translated by a frame into specific commands:

So the command sent to port 1 is "halt" and that sent to port 2 is "off". That completes the establishment of the process.

According to the invention, spot welding can be accomplished by shutting-off weld current when the size of the puddle reaches a predetermined limit.

The information above is directed only toward representative ways of carrying out the invention herein, according to the best mode. The invention itself, however, is broader in scope, as suggested in the claims which follow. ##SPC1##