US 5,426,720 AGrant
Neurocontrolled Adaptive Process Control System
Issue Date:1995-06-20
•8 Claims
•12 Drawing Sheets
Abstract
An adaptive process control system selectively controls vibrations in a given medium in real time. Unwanted vibrations present at a point being monitored in a given medium are sensed, and the system generates an appropriate offsetting vibration that is applied to the medium at a convenient location, which may be remote from the monitored point. The system includes a vibration sensor, such as one or more accelerometers, that sense both input and output vibrations present within the medium; at least one vibration generator, such as an electromagnetic shaker, that generates appropriate offsetting vibrations that are applied to the medium at one or more appropriate locations; and a neural network controller that controls the vibration generator(s) so as to force the sensed vibration at the monitored point(s) to a desired level. The adaptive vibration cancellation provided by the invention takes place in real time, and without the need to process time-consuming complex mathematical algorithms. A specific embodiment of the neural network controller includes a plurality of 4-layer neural networks configured in an adaptive filtered-x configuration.
Metadata
Assignee
- Science Applications International Corporation
Inventors
- Daniel J. Bozich
- H. Bruce MacKay
- Jay A. Eggert
- Ernest E. Muenchau
Application Information
Application Number:US 1236423
Filing Date:1993-09-17
Priority Date:1990-10-30
Art Unit:238
Classifications
IPC:
H04B 1500G21C 736
Field of Search:
39536431822;24;903;904;906508114;561
Patent Drawings (12 sheets)
Description
Background of the Invention
The present invention relates to adaptive process control systems, such as adaptive vibration control systems. More particularly, one embodiment of the invention relates to a multiple-input multiple-output (MIMO) adaptive vibration control system controlled by an artificial neural system that generates, cancels, and/or reduces vibrations in a rigid or semi-rigid (e.g., solid) or fluid (gaseous or liquid) medium.
All processes that use or produce power are susceptible to vibration. Rotating machinery-driven equipment, such as turbine generators, cooling pumps, boiler feed water pumps, forced air draft fans, condenser pumps, conveyor belts, propulsion systems, lathes, mills, grinders, and other manufacturing machines are but a few of the more critical plant systems whose performance may be affected by vibration, either directly or indirectly. Vibration transmitted to neighboring structures or machines may also be detrimental. For example, operators of heavy machinery, trucks, aircraft, or spacecraft, to name but a few, may suffer fatigue due to excessive vibration and noise. Buildings and other rigid structures may suffer permanent damage due to vibrations caused by earthquake. Delicate instrumentation, laser and optical bench systems, automatic machinery, robots, cameras, etc., may also be adversely affected by vibration. Thus, there is a need in the art for vibration reduction systems that can be readily adapted to reduce or cancel undesired vibrations in any given type of system where undesirable vibrations may be present.
Vibrations may be transmitted through solid, liquid or gaseous mediums. Most of the examples presented above relate to vibrations in a solid medium, e.g., vibrations that appear in a metal housing, bearing, or other rigid or semi-rigid structures. It is to be emphasized, however, that vibrations may also be felt and transmitted in a fluid medium (gas or liquid), and that such vibrations may also be detrimental. For example, the vibrations associated with the motor of an underwater craft are readily transmitted through the water and make detection of the location of the craft by an unfriendly source relatively simple. Further, the vibrations associated with normal speech are readily transmitted through air, and are thus available for an unfriendly eavesdropper to hear. Further, any audible sound or vibration present in a fluid medium (whether air or water) may comprise distracting background noise that needs to be eliminated before a desired operation can be carried out. Thus, there is also a need in the art for systems that reduce or cancel vibrations in a fluid medium as well as in a solid medium.
Unfortunately, effectively eliminating or reducing vibrations in a solid medium presents a formidable task While sometimes "shock absorbers" or equivalent devices, may be used to isolate a source of vibrations from a medium remote from the source, such devices are rarely effective at reducing or eliminating vibrations within a given medium when the source of the vibrations is also within the medium.
In order to reduce vibrations that cannot be eliminated through absorber-type isolation devices, it is known in the art to set up counter-vibrations using a suitable vibration source for the purpose of canceling the unwanted vibrations at a desired target point within the medium. Unfortunately, predicting the type and amplitude of the counter-vibrations that must be applied at a given point of the medium in order to effectively cancel the unwanted vibrations at a desired target point within the medium poses an extremely complex problem that is not easily solved. Moreover, the difficulty of this problem is significantly compounded when the problem must be solved in real time. Further, the complexity of the problem expands enormously if either the target point moves or if the character and nature of the unwanted vibration (source vibration) changes.
The vibration cancellation problem becomes virtually impossible to solve using classical techniques when there is more than one desired target point in the same medium that is to be quieted. Further, when a fluid medium is involved, the problem becomes even more difficult due to the non-rigid dynamics of the fluid medium itself. What is needed, therefore, is a vibration reduction system that readily reduces vibrations in real time at one or more target locations in a desired medium, regardless of whether the medium is rigid or fluid, and regardless of the types of vibrations that may be present at the target point(s), and regardless of the location within the medium of the desired target point(s).
Moreover, for some applications, there is a need for a vibration generation system, e.g., a system that generates a desired vibration at the target point that has specified characteristics. For example, the output of vibration actuators are often nonlinear, that is, they produce an output that is distorted both in amplitude and frequency as compared to an input. Therefore, there is a need in the art to linearize the actuator output such that it faithfully reproduces the desired input signal. A vibration generation system has wide applicability, for example, in shake tables and fixtures used to subject various objects and structures to a wide variety of forces, and in seismic oil exploration equipment.
The present invention advantageously addresses the above and other needs.
Summary of the Invention
The present invention provides an adaptive vibration control system suitable for controlling, generating, canceling, reducing, conditioning, or interpreting vibrations in real time. As used herein, the term "vibration(s)" refers to any mechanical or acoustical oscillation or other movement of an object or a wave that occurs in a solid, fluid or electrical medium. The term "vibration control" refers to the generation, cancellation, reduction, conditioning, interpretation, or other processing of a vibration in real time.
In accordance with one aspect of the present invention, there is provided an adaptive vibration control system that senses unwanted vibrations (source vibrations) present in a given medium, and that generates an appropriate offsetting vibration that is applied to the medium at at a convenient location. The offsetting vibration is transmitted through the medium, from a convenient application point, to a desired monitoring point where the source vibration is to be canceled. At the monitoring point, the offsetting vibration, if of the right amplitude and phase, becomes a counter-vibration that cancels the unwanted vibration. If the applied offsetting vibration does not cancel the unwanted vibration at the specified monitoring point to within an acceptable tolerance, then the applied offsetting vibration is automatically altered or adapted until the desired cancellation at the monitoring point is obtained.
Advantageously, the adaptive vibration cancellation provided by the invention takes place in real time, without the need to process complex mathematical algorithms. Vibration cancellation is achieved at any given point within the medium, for almost any type of vibration, within just a few seconds.
In accordance with another aspect of the invention, a multiple-input multiple-output (MIMO) adaptive vibration control system is provided that senses vibrations at multiple monitored locations within a given medium, and that generates multiple offsetting vibrations. The multiple offsetting vibrations are applied to the medium at different locations throughout the medium. The offsetting vibrations are simultaneously transmitted through the medium and combine at the multiple monitored locations to cancel the unwanted vibration at each monitored location. Advantageously, the MIMO adaptive vibration control system operates in real time, without the need to process complex mathematical algorithms. The monitored points whereat the vibrations are controlled may be moved within the medium as desired. Vibration cancellation or other control is achieved at all of the multiple monitored points, for almost any type of vibration, within a very short time.
The adaptive vibration control system of the present invention, whether configured to monitor and control a single point or multiple points within a given medium, and whether using one or multiple sources of offsetting vibrations, includes at least the following elements: (1) source sensing means for sensing a vibration present in the medium from one or more vibration source(s); (2) error sensing means for sensing vibrations at the desired monitoring point(s); (3) at least one vibration generating means for generating at least one offsetting vibration that is applied to the medium at least one location; and (4) control means responsive to the source and error sensing means for controlling the vibration generating means so as to force the sensed vibration at the at least one desired monitoring point to a desired level, e.g., to zero.
Advantageously, the control means of the present invention comprises an artificial neural network, or neurocomputer. Such network or neurocomputer includes many processing elements, each connected to many others, arranged in multiple layers. An input array, or sequence of signals (numbers) from the source sensing means, is entered into a first layer of the network in a time series. This signal propagates forward through the array. The signal(s) from the error sensing means are applied to the output layer of the network and propagate backwards through the array. Each processing element in the first layer takes a component of the input array, operates on it in parallel with the other processing elements in the layer according to a prescribed transfer function, and delivers a single output to the processing elements in a second layer. The prescribed transfer function typically involves multiplying the received signal by an adaptive coefficient, or weight. The adaptive coefficient or weight is adapted by the backward propagation of the error signal(s). Several layers of such processing elements may be used, with the processing elements in each layer operating in parallel on the outputs of a prior layer in accordance with a specified transfer function. The resulting output array, or output signal(s), represent some characteristic associated with the input array, or source signals, as modified by the backward propagating error signal(s). Because the input array and adaptive coefficients (weights) change over time, the network is able to adapt and learn. Hence, there is no need to mathematically model the medium and vibration sources in order to calculate the type of drive signal needed to cancel, or otherwise control, the sensed vibration. Rather, all that is required is to train the artificial neural network or neurocomputer, e.g., by constraining the error signal(s) to assume a desired value, such as zero, so that the adaptive coefficients are changed to produce the desired result.
In accordance with another aspect of the invention, the artificial neural network used as the control means of the invention may be realized in software and/or custom hardware configurations. The software configurations may be implemented on standard computers, such as, PC's, or on PC's augmented with special high speed processing elements, or on custom processing hardware that has been designed to achieve ultra-high processing speeds. The custom hardware configurations are generally preferred because they are able to respond sufficiently fast to provide operation of the vibration control system in real time for most mechanical and/or acoustical vibrations of interest. Custom hardware processing systems control high frequency vibrations, e.g., on the order of 20 KHz or more. The software configuration implemented on a standard computer, in contrast, is generally only suitable for controlling very low frequency vibrations, e.g., on the order of 1 Hz or less; whereas the software configuration implemented on a PC augmented with a special high speed processing element, such as is used in the examples discussed below, is generally suitable for controlling low frequency vibrations, e.g., on the order of 200 Hz or less.
Advantageously, the vibration control system of the present invention finds application in numerous fields. In addition to conventional vibration cancellation systems used, e.g., to cancel unwanted mechanical vibrations in machinery, aircraft, or other mechanical devices containing moving parts that may give rise to the unwanted vibrations, the invention may also be used to steady and protect delicate instruments, laser optical systems, cameras, or building structures from potentially damaging vibrations originating external thereto. For example, the invention may be used to cancel or reduce the vibration forces felt within a specified structure, such as a building or missile silo, caused by earthquake or other forces. Further, the invention lends itself nicely to any application where a given vibration needs to be conditioned or linearized, e.g., linearizing nonlinear actuators, or linear actuators operating in a nonlinear medium.
The invention may also be used to "quiet" a desired area, such as a noisy work area for personal comfort; or a conference room or office where confidential discussions are to occur, so that sounds originating within the area either cannot be heard (as when the system of the invention is used to cancel such sounds), or cannot be understood (as when the system is used to obfuscate such sounds) from a location outside or remote from the area. This same general approach may be used to "quiet" objects so that such cannot be heard from a remote location, such as the "hum" of machinery in air or other media.
Further, the invention may be used in a speech interpretation device, e.g., designed to aid the deaf. In such a device, the vibration control system of the invention senses the vibrations associated with spoken speech. These vibrations cause an equivalent electrical signal to be generated, i.e., a "matching" (as opposed to canceling) vibration signal, representative of the various phonetic components of the sensed speech. These phonetic components, in turn, are then combined and interpreted as language alphabet characters and converted to a visual display of the phonetic alphabetic elements. The combination of the displayed phonetic elements thus provides a user of the device with a written interpretation of the speech that has been spoken. Such an interpretation device may be worn, e.g., by a deaf person, with a display of the sensed speech appearing on the inside surface of a pair of specially configured glasses where only the deaf person can see the display or a hand held display device. With such devices, the deaf person is able to discern, with only a second or two of delay, the words that are spoken to him or her. In fact, the position of the speaker may also be discerned to aid the deaf in locating the speaker.
It is thus a feature of the present invention to provide adaptive process control systems that provide adaptive control in real time for a wide variety of applications, such as vibration control.
It is another feature of the invention to provide a vibration control system that cancels or otherwise conditions unwanted vibrations at a specified location or target point in a given medium.
It is an additional feature of the invention to provide such a vibration control system wherein the desired vibration cancellation or conditioning is accomplished by generating at least one offsetting vibration and applying the offsetting vibration to the medium at a location that may be remote from the target point within the medium.
It is a further feature of the invention to provide such a vibration control system that cancels or otherwise conditions the vibrations existing at multiple target points within the medium.
It is yet another feature of the invention to provide a vibration control system that automatically adapts to provide a desired vibration cancellation or conditioning at one or multiple target points within the medium, regardless of whether the target points remain stationary or move within the medium.
It is still an additional feature of the invention to provide vibration control systems having the above-described features that operates in real time.
It is also a feature of the invention to provide a vibration control system having features as above described that does not require mathematical modeling of the medium or applied vibrations, nor the programming and processing of complex algorithms, in order to operate.
Brief Description of the Drawings
The above and other aspects, features and advantages of the present invention will be more apparent from the following more particular description thereof, presented in conjunction with the following drawings and appendix wherein:
FIG. 1 is a block diagram of a vibration control system made in accordance with the present invention;
FIG. 2 is a diagram of a 4-layer neural network of the type utilized within the artificial neural system (ANS) of FIG. 1;
FIGS. 3A, 3B and 3C are block diagrams showing various embodiments of a single-input, single-output neural network control system used, e.g., for the ANS of FIG. 1;
FIG. 3D is a block diagram showing the preferred manner of implementing a multiple-input, multiple-output (MIMO) neural network control system used for the ANS of FIG. 1;
FIGS. 4A, 4B, 4C, 4D and 4E illustrate representative transfer functions, both amplitude and phase, between selected points of the neural network control system of FIG. 3C;
FIG. 5 shows a block diagram of another embodiment of a vibration control system in accordance with the present invention;
FIG. 6 schematically depicts a vibration control system used to secure an area from eavesdroppers;
FIG. 7 shows a vibration control system used in a speech interpretation device; and
FIG. 8 is a block diagram of a vibration table and associated controls used in conducting Examples 1 and 2, described below.
Appendix A contains the code listing used to program the neurocomputer of FIG. 8 in accordance with a MIMO operating mode of the invention as set forth in Example 2.
Detailed Description of the Invention
The following description includes the best mode presently contemplated for carrying out the invention. This description is not to be taken in a limiting sense, but is made merely for the purpose of describing the general principles of the invention. The scope of the invention should be determined with reference to the claims.
Referring first to FIG. 1, there is shown a block diagram of a vibration control system 20 made in accordance with one embodiment of the present invention. The invention includes an artificial neural system (ANS) controller 22 that is coupled to control a plurality of transducers, identified in FIG. 1 as XDUCER-1, XDUCER-2, XDUCER-3, ... XDUCER-N. The transducers, in turn, are each coupled to a given medium 24. The medium 24 is represented in FIG. 1 as a box. This generalized representation emphasizes that the medium may take several forms, either solid or fluid. Typically the medium will be some sort of structure, such as a table or the mechanical housing of a particular item of machinery. However, it is to be emphasized that the medium may take other forms, such as an entire building, a mass of air, or a mass of water.
For most types of mediums with which the present invention is used, the transducers, such as the transducer XDUCER-1, each comprise a device that converts an electrical signal, obtained from the ANS controller 22, to mechanical motion. The mechanical motion is coupled to the medium through conventional means. For example, where the medium is a rigid or semi-rigid structure, the transducer is typically realized using an electromagnetic shaker that is physically coupled to the medium, causing the medium to shake as controlled by the control signal applied to the electromagnetic shaker. Where the medium is a mass of air, the transducer may comprise an audio speaker, or an audio speaker system (comprising a plurality of speakers), that moves a diaphragm in contact with the air in order to transmit acoustic waves through the air.
As further seen in FIG. 1, a plurality of sensors, S0, S1, S2, S3, . . . Sn, are also coupled to the medium 24. The number of sensors n need not be the same as the number of transducers N, although it frequently will be. The sensors are selected to detect vibrations that occur at a particular monitoring point within the medium. One sensor, S0, senses the source vibration, e.g., caused by an external force F.sub.E or an internal force F.sub.I. The other sensors sense error vibrations at designated points throughout the medium 24. Thus, as shown in FIG. 1, sensor S1 detects vibrations occurring at monitoring point P1, sensor S2 detects vibrations occurring at monitoring point P2, sensor S3 detects vibrations occurring at monitoring point P3, and so on, with sensor Sn detecting vibrations occurring at monitoring point Pn. (The monitoring point may sometimes be referred to herein as the "target point", because it is the desired location, i.e., the target, where the vibrations are to be controlled.) The actual location of the various monitoring points P1, P2, P3, . . . Pn is controlled simply by moving the sensors S1, S2, S3, . . . Sn to a desired target location.
The sensors S0, S1, S2, S3, . . . Sn may be realized, for most types of media 24 with which the vibration control system 20 is used, with any suitable transducer that converts mechanical motion or force to an electrical signal. Numerous types of sensors may be used for this purpose. For example, where the medium comprises a rigid or semi-rigid structure of some type, the sensors S0, S1, S2, S3, . . . Sn will typically be realized using a conventional accelerometer. Where the medium comprises a fluid, such as air, the sensors S0, S1, S2, S3, . . . Sn will typically be realized using a conventional microphone.
As indicated, the sensors S0, S1, S2, S3, . . . Sn generate an electrical signal having a magnitude representative of the vibration that is sensed by each sensor. These signals are connected to the ANS controller 22. The ANS controller 22, as explained more fully below, responds to the detected vibrations by generating a control signal that is applied to each transducer. The control signal drives the transducer so as to cause the vibration sensed by each particular sensor to achieve a desired level.
The embodiment of the invention shown in FIG. 1 is particularly well suited for vibration cancellation. That is, the vibration sensed at point P0 by sensor S0 is connected to the ANS controller 22 as the source or input signal. The vibration sensed at any of the monitoring points P1, P2, P3, ... Pn within the medium 24, which vibration is manifest by the respective output signals of the sensors S1, S2, S3, ... Sn, serves directly as an error signal that is fed back to the ANS controller 22. The ANS controller, in turn, is trained or adapted, as explained below, to control the respective transducers, XDUCER-1, XDUCER-2, XDUCER-3, ... XDUCER-N, so as to drive each of the error signals to zero. Any source vibration that may be present in the medium 24, as excited, e.g., by an external force F.sub.E and/or an internal force F.sub.I, is thus sensed at the respective monitoring points, and appropriate offsetting vibrations are immediately generated so as to eliminate, or reduce, the sensed vibration. Advantageously, due to the ability of the ANS controller 22 to automatically adapt itself to reduce the vibrations at the selected error sensing points, the sensors S1, S2, S3, . . . Sn can readily be moved within the medium as often as is required or desired, and the system 20 will quickly adapt to reduce or cancel the vibrations at the new monitoring points. (It is noted that the sensor SO may also be moved within the medium 24, but typically it is positioned close to the source of the input vibration so as to present a true picture of the input vibration.)
Significantly, the vibration control system 20 shown in FIG. 1 is able to simultaneously cancel or reduce vibrations at multiple monitoring points within the medium 24 by simultaneously generating multiple offsetting vibrations that are applied to the medium at various locations. This multiple-input, multiple-output (MIMO) type of vibration control has not heretofore been possible, to applicants' knowledge. This is because of the complexity of the problem involved. That is, the vibration that occurs at each monitoring point, e.g., the point P1 is not only attributable to the source vibration (caused by any external forces F.sub.E or internal forces F.sub.I that may be present in the medium 24), but also includes components of the offsetting vibrations generated by the respective transducers, XDUCER-1, XDUCER-2, XDUCER-3, . . . XDUCER-N. The magnitude and phase of the offsetting vibrations generated by each of the controlling transducers must be controlled so that the vibration at each of the monitoring points is zero. As a correction is made at one transducer to reduce the vibration at one monitoring point, the correction is felt at all the other monitoring points as well. Thus, the ANS controller 22 must control all of the transducers so that they all cooperate as a group to reduce the vibrations at each of the selected monitoring points. There is thus a very complex interaction between all of the transducers and all of the monitoring points that is extremely difficult to describe or model using classical mathematical techniques.
A simplified version of the vibration control system shown in FIG. 1 is the case where only one transducer, one source sensor, and one error sensor is used. Such a simplified system provides sufficient vibration control where the vibration is to be controlled only in one location (at the location of the error sensor). However, other applications require vibration control at more than one point simultaneously, in which case the more generalized MIMO vibration control system shown in FIG. 1 must be used.
As is evident from the above generalized description of the invention, a key component of the vibration control system 20 is the ANS controller 22. An Artificial Neural System (ANS) is a system that performs neurocomputing. Neurocomputing involves the use of one or more neural networks--networks patterned after the structure of a human brain--to perform nonalgorithmic information processing. Such processing is an alternative form of information processing that involves the parallel processing of large amounts of small pieces of information (data). Neurocomputing is fast becoming an established discipline. See, e.g., Hecht-Nielsen, Robert, "Neurocomputing: Picking the Human Brain," IEEE Spectrum, pp. 36-41 (March 1988).
There is a great deal of literature currently available describing the operation and use of artificial neural systems, or neurocomputers. See, e.g., Rumelhard, et al., Parallel Distributed Processing, Vols. 1 and 2 (MIT Press, Cambridge, MA 1988); Artificial Neural Networks: Electronic Implementations, Nelson Morgan, Editor, (IEEE Computer Society Press, 1990); Artificial Neural Networks: Concept Learning, Joachim Diederich, Editor (IEEE Computer Society Press, 1990). Thus, no attempt will be made herein, except on a very cursory level, to explain the manner in which an ANS operates. Basically, an ANS is simply one form of a parallel distributed network (PDN) made up of processing elements (PE's) that are interconnected via information channels called interconnects. Where these PE's perform simple nonlinear functions, they may be referred to as "neurons." Each PE or neuron may have multiple inputs, but typically only one output (although this output may be connected to several different locations within the ANS). The PE's are arranged in several different layers. A first or input layer of PE's is generally characterized by each PE having an input that comprises one of the input signals to the PDN. An output layer of PE's is characterized by each PE providing one of the output signals of the PDN. The signal at the output of each PE in the input layer functions as an input signal to the PE's of the next or other layers, either in a feedforward or feedback interconnect. These interconnections may be weighted by an appropriate factor.
For simple PDN applications, two layers of PE's may be sufficient, in which case the PDN comprises only an input layer and an output layer. For more complex applications, one or more hidden layers of PE's may be required between the input and output layers. The complexity of the PDN is largely governed by the number of layers of PE's.
The general structure of a four-layer neural network, for example, is illustrated in FIG. 2. This particular four-layer neural network is the preferred network for use in the ANS controller of the present invention, although it is to be emphasized that the number of layers for the neural network of the present invention is not limited to four.
As indicated, one advantage of using an PDN or ANS is that it can adapt or self-adjust the strength of the interconnections between processing elements. Self-adaptation allows an ANS to "learn" and to eventually improve in overall system performance.
Learning results from application of a selected learning rule. Most learning rules comprise variants of the Hebbian learning rule which states that if two units (processing elements) are highly active, the strength of the interconnect between these two units should be strengthened. A variation of this learning rule is the delta rule. According to the delta rule, the amount of learning (i.e., the strength or weight of the interconnect between two processing elements) is proportional to the difference (or delta) between the actual activation achieved and a target activation provided by a teacher. A variation of the delta rule is the generalized delta rule. The generalized delta rule provides a technique for backpropagating the error from the output layer to the hidden layers so that the delta rule can be used for internal weight updates. The delta rule and the generalized delta rule are discussed at length in Chapters 8 and 11 of Rumelhart reference, cited above, and several variations or extensions of the rule exist.
The generalized delta rule, its variations and/or extensions, provide the primary learning rule utilized by the ANS of the present invention. Although other learning rules, known or yet to be developed, could of course be used with the vibration control system of the present invention, use of the generalized delta rule advantageously provides an effective way to accurately teach the ANS how to control a force transducer, so that in cooperation with other force transducers, the vibrations at desired monitoring points within a medium can be effectively controlled. Other neural network algorithms and architectures could also be applied, such as recurrent back propagation, direct weight connection from input to output, and back percolation networks.
Advantageously, the ANS controller 22 shown in FIG. 1 may be realized using a commercially available neurocomputer, such as the Sigma Workstation available from Science Applications International Corporation (SAIC), San Diego, Calif. The Sigma Workstation is a 386 AT-based computer that features SAIC's Delta II floating point processor card as an accelerated processor. The Delta II floating point processor card is also commercially available from SAIC, as is a full set of documentation describing its installation and use in a 386 AT-based computer. An additional description of the Delta II floating point processor card may be found in U.S. Patent application Ser. No. 07/191,207, filed 05/06/88. assigned to SAIC.
As indicated above, FIG. 2 is a diagram of a 4-layer neural network of the type utilized within the ANS controller 22 of FIG. 1. A first input to the network comprises the output signal from the source sensor S0. A second input to the neural network comprises the derivative of the first input. As shown in FIG. 2, the input layer includes nominally 35 units or neurons, each having an input. The sensor output signal is applied to one half of these inputs in a time series. The derivative of the sensor output signal is applied to the other half of these inputs in a time series. That is, the sensor output signal is applied to a suitable shift register, or equivalent, that is nominally seventeen (17) stages long. Thus, at time t.sub.n, a first stage of the input signal shift register contains the sensor output signal (or a digital representation of the sensor output signal) at that time. Similarly, a second stage of the shift register contains the sensor output signal at time t.sub.n-1, one sample time prior to time t.sub.n. A third state contains the sensor output signal at time t.sub.n-2, and so on, with the last stage of the shift register containing the sensor output signal at time t.sub.n-k. The shift register is thus at least k+1 stages long. Each stage of the shift register is connected to a corresponding neuron or "unit" of the first half of the first layer of the neural network. In like manner, the derivative of the sensor output signal is applied to a suitable shift register, each stage of which is connected to a corresponding neuron or "unit" of the second half of the first layer of the neural network.
As further seen in FIG. 2, the preferred neural network also includes two intermediate layers, or "hidden" layers of neurons, identified as H1 and H2. Each of the hidden layers contains nominally ten neurons or units. (It is to be emphasized that the number of neurons in each layer presented herein is only exemplary.) A final or output layer of the network contains at least one neuron or unit. In general, the number of neurons in the output layer will be equal to the number of outputs of the neural network. All of the layers are fully interconnected, including a directly connected layer between the inputs and outputs. (Only some of the interconnections are shown in FIG. 2 for clarity.) That is, each of the 35 neurons in the input layer is connected to all of the neurons in the H1 layer. Similarly, each of the neurons in layer H1 is connected to all of the neurons in the H2 layer. Each of the neurons in layer H2 is connected to each output neuron. Further, each of the neurons in the input layer is directly connected to each output neuron.
The connections between the respective neurons are weighted with adaptive coefficients that determine the strength of the connection. These weights are determined by error signals, which error signals (represented in FIG. 2 as being applied to the bottom of the layers) propagate backwards through the network in accordance with the relation: ##EQU1## where E.sub.i represents the error for the i.sup.th neuron in the i layer, W.sub.ij represents the weights or adaptive coefficients between neurons in the i and j layers of the network. O.sub.i is the output of the i.sup.th neuron, and f'(O.sub.i) is the derivative of the output of the ith neuron.
These weights are updated automatically in accordance with the generalized delta rule described above and referenced in the literature. The backpropagated error determines the change .DELTA.W.sub.ij in the connection strengths according to the relation:
where .eta. is the predetermined learning rate.
As indicated above, the ANS controller 22 may be implemented using software, using a neurocomputer (such as the SAIC Sigma Processor), or in custom hardware (e.g., in a group of custom LSI or VLSI chips). ("LSI" stands for large scale integration; "VLSI" stands for very large scale integration. Both terms refer to the number of active components, e.g., transistors, that are included on a single silicon or other semiconductor chip used to fabricate an integrated circuit, or IC).
Except for controlling very low frequency vibrations, e.g., less than 1 Hz, a software implementation of an ANS using a conventional computer is not practical. Advantageously, when a neurocomputer is programmed to realize the neural-network configurations described below in connection with FIGS. 3A-3D, excellent vibration control for vibrations up to about 200 Hz may be achieved, as indicated in the two examples described below. This is a sufficiently wide bandwidth for most non-acoustical vibration control applications. For higher vibration frequencies, a custom LSI/VLSI chip-based design that implements the ANS configuration described below in FIG. 3 may be used. Vibration frequencies up to about 20 KHz may advantageously be controlled using such a custom LSI/VLSI design. This is a sufficiently wide bandwidth for all acoustical vibration control applications.
One of the objectives of the present vibration control system is to provide effective vibration control in "real time". This means that there can not be any significant delays between the time when a vibration is first sensed, and when an appropriate corrective signal is applied to control the vibration in a desired fashion. While it is fairly straightforward to reduce delays when the input excitation signal (i.e., the vibration force that excites the unwanted vibrations in the medium) is a periodic force, because the input force can be readily predicted due to its periodic nature, achieving a sufficiently fast response time to react to non-periodic changes occurring in the input excitation signal presents a formidable challenge, even for an adaptive ANS system. Thus, a key component of the vibration control system of the present invention is to configure a collection of neural networks, of the type shown in FIG. 2, in a configuration that provides the requisite response time. One such configuration is shown in the block diagram of FIG. 3A. Others are shown in the block diagrams of FIGS. 3B-3D.
FIG. 3A illustrates a block diagram of an embodiment of a single-input, single-output neuro-system for use, e.g., with the vibration control system 20 (FIG. 1) of the present invention. The neuro-system 30 forms part of the ANS controller 22 (FIG. 1), and is used for the simplified situation where only one transducer, e.g. XDUCER-1, is used to impart an offsetting vibration to the medium 24, and where only one sensor, e.g., S1 is used to sense vibrations at a given location, e.g. point P.sub.1, within the medium 24. The neuro-system 30 shown in FIG. 3A is patterned after an LMS algorithm utilized in classical adaptive signal processing design. See, e.g., Widrow and Stearns, Adaptive Signal Processing, ("Adaptive Control Systems"), (Prentice-Hall, 1985). However, there are some significant differences between a classical adaptive control system, and the configuration shown in FIG. 3A, principally relating to the manner in which the various elements of the neuro-system are realized using non-linear neural network archictures instead of linear LMS architectures.
As seen in FIG. 3A, the single-input, single-output neuro-system 30 comprises a neuro-controller 32 that is coupled to control a "controlled dynamic device" 34. The neuro-controller 32 is realized using a neural network of the type described above in connection with FIG. 2. The input signal to the neuro-controller 32 is a signal obtained from a suitable source sensor, e.g., the sensor SO (FIG. 1). This input signal is identified at the far left of FIG. 3A as the signal A0. This input signal A0 is converted from analog to digital format using an "adc" (analog-to-digital converter) and time delayed as required in order to provide an input signal to each input of the first neuro-controller 32 (i.e., to each input of the first layer of the neural network that comprises the neuro-controller 32). This time delay is achieved by using the input layer as a shift register giving the network a time history of the input values. Further, the input signal may be differentiated and time shifted using a suitable differentiator represented in FIG. 3A by the ".DELTA." circuit, and input shift register. The error signal is applied to the neuro-controller 32. The error signal back-propagates through the neural network that makes up the neuro-controller 32 to set the value of the "weights" used therein, which error signal is obtained either from the sensor 39 for cancellation operations or from the summing point 37 for linearization operations as described below.
The controlled dynamic device 34 comprises a suitable transducer for converting an output signal from the neuro-controller 32 to linear mechanical motion. The transducer device 34 may be, e.g., an electromagnetic shaker. The transducer device 34 is coupled to an intermediate medium 24'. A dynamic sensor 39 is also coupled to the medium 24'.
FIG. 3B illustrates a block diagram of another embodiment of a single-input, single-output neural-system 30I for use, e.g., with the vibration control system 20 (FIG. 1) of the present invention. the neural-system 30I is similar to neural system 30 (FIG. 3A) with the addition of an inverse neuro-emulator 36I which is connected in parallel with the transducer 34, medium 24; and dynamic sensor 39. It is the function of the inverse emulator to learn to emulate the inverse response of the medium to outputs from the neuro-controller 32I. The input signal is obtained from the dynamic sensor 39 output to the inverse neuro-emulator. The input signal and its derivative are time shifted, as required, using shift registers as above described, in order to provide an input to each input of the inverse neuro-emulator circuit 36I. The output of the inverse emulator 36I is compared with the input to the device 34 at a summation point 35 in order to develop an inverse emulator error signal, identified as "ee0". This inverse emulator error signal is fed back to the inverse emulator 36I in order to backpropagate through the layers of the inverse emulator 36I and change the weights as required in order to force the inverse emulator error signal to zero. When the inverse emulator error signal is zero, the inverse neuro-emulator 36I thus represents an exact imitation of the inverse transfer function of the coupled transducer 34 and medium 24.
The "weights" used in the neuro-controller 32I are obtained from the inverse neuro-emulator 36I. This weight transfer is represented in FIG. 3B by a dotted line 33.
As seen in FIG. 3B, the error signal used to set the values of the weights used in the neuro-controller 32I is an error signal, identified by the gain block control "cse0", that may be obtained from one of two sources. When the neural-system 30I is operating in a linearization mode, the error signal "cse0" comprises the difference between the signal sensed by the sensor 39 and the input signal A0. This difference signal is generated at summing node 37. When the neural-system 30I is operating in a cancellation mode, the gain control "ctgt" is set to zero, thereby causing the error signal "cse0" to simply be the signal sensed by the sensor 39. That is, in a cancellation mode, any vibrations sensed at the sensor point represent undesired vibrations, and thus the sensor signal itself provides the error signal.
Thus, it is seen that the neural system 30I embodiment shown in FIG. 3B, discussed above, is a refinement of the neural system 30 (FIG. 3A).
In all of FIGS. 3A-3D, it is noted that the blocks marked "hpf" refer to a "high pass filter." The ovals marked "adc" refer to an analog-to-digital converter; and the ovals marked "dac" refer to a digital-to-analog converter. In some figures, e.g., FIGS. 3C and 3D, the adc's and dac's have been omitted for clarity. The other small blocks in the figures refer to gain blocks, i.e., multiplication factors used to adjust the magnitude of the various signals.
FIG. 3C illustrates a block diagram of a preferred single-input, single-output neuro-system 30x for use, e.g., with the vibration control system 20 (FIG. 1) of the present invention. The neuro-system 30x forms part of the ANS controller 22 (FIG. 1), and is also used for the simplified situation where only one transducer, e.g. XDUCER-1, is used to impart an offsetting vibration to the medium 24, and where only one sensor, e.g., S1 is used to sense vibrations at a given location, e.g. point P.sub.1, within the medium 24. The neuro-system 30 shown in FIG. 3C is patterned after a "filtered-X" LMS algorithm utilized in classical adaptive signal processing design. See, e.g., Widrow and Stearns, supra, Chap. 11 ("Adaptive Control Systems"), pp. 285-297. However, there are some significant differences between a classical "filtered-X" adaptive control system, and the configuration shown in FIG. 3C. These differences relate mainly to the manner in which the various elements of the neuro-system are realized using pairs of non-linear neural network architectures instead of linear LMS architectures.
As shown in FIG. 3C, the single-input, single-output neuro-system 30x includes a first neuro-controller 32x that is coupled to control a "controlled dynamic device" 34. The first neuro-controller 32x is realized using a neural network of the type described above in connection with FIG. 2. The input signal to the first neuro-controller 32x is a signal obtained from a suitable source sensor, e.g., the sensor SO (FIG. 1). This input signal is identified at the far left of FIG. 3C as the signal A0. The sensor outputs are typically analog values that are digitized at an appropriate sample rate for input to the digital network. The derivative, or first difference, of the input signal A0 is calculated, as represented by the ".DELTA." box 92. Both signals are then shifted into the input layer of the network which is arranged as a shift register to give the network a time history of the input values. An error signal "cseo" is applied to a second neuro-controller 38x, as explained below. This error signal back-propagates through the neural network that makes up the neuro-controller 38x to set the value of the "weights" used therein.
As with the embodiments in FIGS. 3A and 3B, the controlled dynamic device 34 comprises a suitable transducer for converting an output signal from the first neuro-controller 32x to mechanical motion. The transducer device 34 may be, e.g., an electromagnetic shaker. The transducer device 34 is coupled to an intermediate medium 24'. A dynamic sensor 39 is also coupled to the medium 24'.
Connected in parallel with the transducer device 34 is a first neuro-emulator 36x. It is the function of the emulator 36x to emulate or imitate the response of the coupled device (transducer) 34 and medium 24'. When performing this function, the input signal to the first neuro-emulator 36x is the input signal A0 to the system 30, obtained via an "emulator training bypass" route around the neuro-controller 32. The first difference of the input signal is formed, and both signals are then shifted into the input shift register of the first neuro-emulator circuit 36x. The output of the emolator 36x is compared with the output of the device 34 at a summing point 35x in order to develop an emulator error signal. This error signal may be multiplied, as desired, by a gain factor identified as "ee0". This emulator error signal is fed back to the emulator 36x in order to back propogate through the layers of the neuro-emulator 36x and change the weights thereof as required in order to force the emulator error signal to zero. When the emulator error signal is zero, the first neuro-emolator 36x thus represents an exact imitation of the coupled device 34 and the response of the medium 24'.
As indicated above, the "weights" used in the first neuro-controller 32x are obtained from a second neuro-controller 38x. This "weight" transfer is represented in FIG. 3 by a dotted line 33x. That is, by copying the neuro-controller weights (or "adaptive coefficients) from the second neuro-controller 38x to the first neuro-controller 32x, the second neuro-controller 38x actually controls the first neuro-controller 32x. The second neuro-controller 38x receives its inputs from a second neuro-emulator 40x (i.e., the input signal to the second neuro-controller 38x is the output signal from the second neuro-emulator 40x, differentiated and time shifted as required to fill all the inputs to the second neuro-controller 38x).
As seen in FIG. 3C, the error signal used to set the values of the weights used in the neuro-controller 38x is an error signal, identified by the gain block control "cse0", that may be obtained from one of two sources. When the neural-system 30x is operating in a linearization mode, the error signal "cse0" comprises the difference between the signal sensed by the sensor 39 and the input signal A0. This difference signal is generated at summing node 37. When the neural-system 30x is operating in a cancellation mode, the error signal "cse0" simply comprises the signal sensed by the sensor 39. That is, in a cancellation mode, any vibrations sensed at the sensor point represent undesired vibrations, and thus the sensor signal itself provides the error signal.
It is the function of the second neuro-emulator 40x to condition the input signal to the second neuro-controller 38x. The second neuro-emulator 40x uses the input signal A0, as its input signal, differentiated and time shifted as required in order to fill all the inputs to the first layer of the neuro-controller 40x. The "weights" of the second neuro-emulator 40x are copied form the first neuro-emulator 36x. This copying is represented by the dotted line 41.
In accordance with the present invention, the first neuro-emulator 36x and the second neuro-emulator 40X are each realized with a neural network of the type shown in FIG. 2. These two neural networks operate as a pair, with the emulator 36x serving as a "trainer" neural network, and the emulator 40x serving as the "executer" neural network. As soon as the "trainer" network 36x is trained, its weights are transferred to the "executer" network 40x. Advantageously, at substantially the same time that the "training" network 36x is training, the "executer" network 40x is executing based on the most recent weights transferred thereto from the training network. In this way, the neuro-system 30x is able to adapt very quickly to any changes that may occur in the input signal A0.
Similarly, the first neuro-controller 32x and the second neuro-controller 38x may each be realized with a neural network of the type shown in FIG. 2. These two neural networks also operate as a pair, with the controller 38x functioning as a "trainer" neural network, and the controller 32x functioning as an "executer" neural network. Once trained, the weights associated with the controller 38x are transferred to the controller 32x. Advantageously, as with the emulator pair of neural networks described above, at substantially the same time that the controller 38x is being trained, the controller 32x is executing control of the transducer device 34 based on the most recent weights transferred thereto from the controller 38x.
FIGS. 4A-4E illustrate representative transfer functions, showing both amplitude and phase, between selected points of the neuro-control system 30x of FIG. 3C.
For example, as seen in FIG. 4A, the transfer function between points (0) and (6) of FIG. 3C is depicted. This is basically the transfer function of the controller 32x. Note the wide variations in both phase and amplitude as the frequency of the input signal is swept from 10 Hz to 140 Hz.
FIG. 4B shows the transfer function between points (6) and (2) of FIG. 3. This is basically the transfer function of the coupled device 34 and medium 24'. Like FIG. 4A, there are wide variations in both phase and amplitude as the frequency is swept between 10-140 Hz.
However, note that the variations in phase and amplitude shown in FIG. 4B between 10-140 Hz are substantially equal and opposite to those shown in FIG. 4A. Thus, when the overall transfer function is made, between points (0) and (2), as shown in FIG. 4C, it is seen that, under the conditions shown, the output is able to track the input with very little delay or distortion.
FIG. 4D shows the transfer function from points (0) to (9), and thus shows how well the emulator, 40x is able to follow or imitate the coupled device 34 and medium 24' during the training of neurocontroller 38x. The transfer function in FIG. 4E compares the output of the emulator 36x with the output of the coupled device 34 and medium 24' as sensed by sensor 39 and shows how well the emulator 36x is able to follow or imitate the coupled device 34 and medium 24' in the frequency bakd from 10-140 Hz..
Referring next to FIG. 3D, a block diagram is shown of the preferred manner of realizing a multiple-input, multiple-output (MIMO) neuro-system 30M for use, e.g., with the vibration control system 20 (FIG. 1) of the present invention. The neuro-system 30M forms part of the ANS controller 22 (FIG. 1), and is used for the general situation where multiple transducers, e.g. XDUCER-1, XDUCER-2, XDUCER-3, .... XDUCER-N, are used to impart respective offsetting vibrations to the medium 24, and where only multiple sensors, e.g., S1, S2, S3, . . . Sn are used to sense vibrations at respective locations, e.g. points P.sub.1, P.sub.2, P.sub.3, P.sub.n within the medium 24.
For most purposes, the block diagram of FIG. 3D parallels the block diagram of FIG. 3C, and includes the same or equivalent elements as are shown in FIG. 3C, except that the gain blocks, e.g., ee0, cse0, etc., are omitted for clarity. (It is noted that the same reference characters are used in FIGS. 3A-3D to identify the same or equivalent elements, except that different suffixes are used for different embodiments. In FIG. 3A, no suffix is used. In FIG. 3B, the suffix "I" is used. In FIG. 3C, the suffix "X" is used, and in FIG. 3D, the suffix "M" is used ) The suffix "M" for example, signifies that although the element shown in FIG. 3D performs the same function as a corresponding element in FIGS. 3A-3C, the element itself has been modified in order to accommodate multiple inputs and/or outputs. Thus, as shown in FIG. 3D, a first neuro-controller 32M is coupled to control a plurality of "controlled dynamic devices" 34M. (These controlled dynamic devices comprise the plurality of transducers described above in connection with FIG. 1, i.e., XDUCER-1, XDUCER-2, XDUCER-3, . . . XDUCER-N.) The first neuro-controller 32M is realized using a neural network of the type described above in connection with FIG. 2, except that it has multiple outputs. Such multiple outputs are readily realized using a neural network of the type shown in FIG. 2 by simply including additional neurons in the output layer.
The input signal to the first neuro-controller 32M is a signal obtained by summing the signals obtained from multiple inputs, identified as A01, A02, . . . AON. A summing network 45 performs this summing function. This summed input signal is differentiated and time delayed as required in order to provide input signals to each input of the first neuro-controller 32M (i.e., to each input of the first layer of the neural network that comprises the neuro-controller 32M). The sensor outputs are typically analog values that are digitized at an appropriate sample rate for input to the digital network. The derivative, or first difference, of the input signal A0 is calculated, as represented by the ".DELTA." box. Both signals are then shifted into the input layer of the network which is arranged as a shift register to give the network a time history of the input values. The error signal applied to the first neuro-controller 32M, i.e., the error signal that back-propagates through the neural network to set the values of the "weights" used therein, is obtained from a second neuro-controller 38M.
Connected in parallel with the transducer devices 34M is a first neuro-emulator 36M. It is the function of the emulator 36M to emulate or imitate the coupled devices 34M and media 24'. Each medium 24M may represent a hybrid medium, e.g., solid or fluid or hybrid combinations. When initially performing this function, the input signal to the first neuro-emulator 36M is the summed input signals A01, A02, . . . A0N obtained via an "initial emulator training only" route around the neuro-controller 32M. Subsequent to initial training, the multiple outputs from the neuro-controller 32M are summed, at summer 49, as required to provide this input signal. This input signal is differentiated and time shifted, as required, using a shift circuit as above described, in order to provide inputs to each input of the first neuro-emulator circuit 36M. The neuro-emulator 36M includes multiple outputs. Each of these outputs are compared with respective outputs from the sensor devices 39M at summing point 35M in order to develop a set of emulator error signals, identified as "Emulator Error 1", . . . "Emulator Error N". This set of emulator error signals is fed back to the emulator 36M in order to back propagate through the layers of the neuro-emulator 36M and change the weights thereof as required in order to force each of the emulator error signals to zero. When each of the emulator error signals is zero, the first neuro-emulator 36M thus represents an exact imitation of what the coupled devices 34M and media 24' must be if the error signals are to be zero, i.e., if the vibrations sensed by the dynamic sensors 39M are at a desired level.
As with the single-input, single-output control system shown in FIG. 3C, the "weights" used in the first neuro-controller 32M of the MIMO control system 30M shown in FIG. 3D are obtained from a second neuro-controller 38M. By copying the neuro-controller weights (or adaptive coefficients) from the second neuro-controller 38M to the first neuro-controller 32M, the second neuro-controller 38M actually controls the first neuro-controller 32M. The second neuro-controller 38M receives its inputs from a second neuro-emulator 40M (i.e., the input signals to the second neuro-controller 38M are the summed output signal and its derivative from the second neuro-emulator 40M, time shifted as required to fill all the inputs to the second neuro-controller 38M).
As seen in FIG. 3D, the error signals used to set the values of the weights used in the second neuro-controller 38M are a set of error signals, identified as "Controller Error 1, . . . Controller Error N", that may be obtained from one of two sources. When the neuro-system 30M is operating in a linearization mode, the set of controller error signals comprises the difference between the set of signals sensed by the sensors 39M and the set of input signals A01, A01, ... AON. This set of difference signals is generated at summing nodes 37M. When the neuro-system 30M is operating in a cancellation mode, the set of controller error signals simply comprises the set of signals sensed by the sensors 39M.
It is noted that the number of sensors 39M need not be the same as the number of input signals or the number of transducers 34M. If the number of sensors 39M is not the same as the number of input signals, then a "sensor fusion" circuit (i.e., a summation circuit) is used as required to combine selected sensor and/or input signals so that an appropriate difference signal can be generated.
It is the function of the second neuro-emulator 40M to condition the second neuro-controller 38M. The second neuro-emulator 40M uses the summed set of input signals, from summer 45, as ont of its input signals, and the differential of this input signal as another input signal. Both of these input signals are time shifted; as required, in order to fill all the inputs to the first layer of the neuro-controller 38M. The "weights" of the second neuro-emulator 40M are copied from the first neuro-emulator 36M.
As with the single-input, single-output embodiment shown in FIG. 3C, the first neuro-emulator 36M and the second neuro-emulator 40M in the MIMO embodiment of FIG. 3D are each realized with a neural network of the type shown in FIG. 2, modified to include multiple neurons in the output layer. These two neural networks operate as a pair, with the emulator 36M serving as a "trainer" emulator neural network, and the emulator 40M serving as the "executer" emulator neural network. As soon as the "trainer" network 36M is trained, its weights are transferred to the "executer" network 40M. Advantageously, at substantially the same time that the "training" network 36M is training, the "executer" network 40M is executing based on the most recent weights transferred thereto from the training network. In this way, the control system 30M is able to adapt very quickly to any changes that may occur in the set of input signals A01, A02, . . . AON.
Similarly, the first neuro-controller 32M and the second neuro-controller 38M may each be realized with a neural network of the type shown in FIG. 2, modified to include multiple outputs. These two neural networks also operate as a pair, with the controller 38M functioning as a "trainer" neural network, and the controller 32M functioning as an "executer" neural network. Once trained, the weights associated with the controller 38M are transferred to the controller 32M. Advantageously, as with the emulator pair of neural networks described above, at substantially the same time that the controller 38M is being trained, the controller 32M is executing control of the transducer device 34M based on the most recent weights transferred thereto from the controller 38M.
Referring next to FIG. 5, a block diagram is shown of an alternative embodiment 20' of a vibration control system in accordance with the present invention. This block diagram is the same as that shown in FIG. 1 (and hence the same reference numerals are used to identify like parts) except that the sensor output signals obtained from the sensors S1, S2, S3, . . . Sn, are compared with appropriate reference signals 42 at summing nodes 43a, 43b, 43c, . . . 43n prior to being returned to the ANS controller 22. Thus, the signals returned to the ANS controller 22 comprise error signals representing the difference between the sensed output vibration, and a desired reference vibration. The ANS controller responds to these error signals so as to drive this error signal to zero. Thus, the output vibrations sensed at the various monitoring points are forced to the reference values. If the reference values are zero, then the system 20' of FIG. 5 operates as a vibration cancellation system the same as the system 20 of FIG. 1. If the reference values are not zero, however, then the system 20' operates to control the vibrations and force them to a desired value. For example, a reference signal may take the form of a sine wave or triangular wave, in which case the ANS controller 22 forces the medium 24 to vibrate in such a way that each monitored point also assumes the desired sine or triangular wave. Such a situation may be desirable, e.g., in a shake table, where particular points in the medium (the structure under test) are to be subjected to specified types of vibration for testing purposes. In such an instance, the vibration control system thus operates as a vibration generator, with the output vibration faithfully following the input reference signal.
FIG. 6 schematically depicts a vibration control system used to secure an area from eavesdroppers. The area to be secured is represented by the area within the square 50. This area may comprise an office or conference room. Four speakers (or speaker systems, or equivalent), 52a, 52b, 52c, and 52d are placed at appropriate locations around the area 50. As shown in FIG. 6, the speakers are located in the corners of the area 50. Adjacent each speaker is at least one microphone. For example, two microphones 54a and 54b are shown adjacent speaker 52a. Similarly, two microphones are placed adjacent the speakers 52b, 52c and 52d. It is noted that the ratio of two microphones for each speaker is only exemplary. More or less speakers and/or microphones may be used. Each microphone is coupled through an analog-to-digital (A/D) converter 56 to an ANS controller 58. Further, each speaker 52a, 52b, 52c, 52d is coupled to the ANS controller 58 through a digital-to-analog (A/D) converter 60 and respective amplifiers 62a, 62b, 62c, and 62d.
In operation, the microphones function as the sensors of the vibration control system, and the speakers function as the transducers. The medium is the air in the area 50. Any audible sounds, such as spoken words, that occur within the area 50 are sensed by each of the microphones, e.g., microphones 52a and 52b. These sounds thus become the "vibrations" that are canceled by the system at the monitoring points. Thus, upon sensing the sounds at the monitoring points, the ANS controller 58, operating in the manner as described above, generates appropriate control signals to drive the speakers 52a, 52b, 52c, and 52d, also positioned around the area 50, to cancel the sounds. Thus, an observer near the monitoring points is not able to detect any sound, and the area 50 is secure from eavesdroppers.
FIG. 7 shows the block diagram of a speech interpretation device 70 that utilizes the vibration control principles of the present invention. Solving complex pattern-recognition problems implicit in understanding continuous speech has long been recognized as one of the strong points of a neural network. See, e.g., Hecht-Nielsen, IEEE Spectrum, above. Thus, the device depicted in FIG. 7 utilizes this inherent advantage of an neural network to provide an interpretation device 70 that may be used by the hearing impaired to better understand speech.
As seen in FIG. 7, the interpretation device 70 includes a microphone system 72, which system includes associated amplifiers and A/D converters to sense, amplify and condition speech sounds into digital signals. The digital signals from the microphone system 72 serve as the input to a chip-based neural network 74. The output of the neural network 74 is a sequence of signals representing the phonetic sounds present in the speech sounds. These phonetic signals are processed in conversion circuitry 76, which circuitry converts these signals to appropriate signals suitable for driving a suitable alphanumeric display 78. The display 78 displays sequences of letters representative of the phonetic sounds that have been heard, conditioned by the neural network so as to represent an accurate sequence of the sounds actually spoken.
Thus, when a person 71 speaks a given sequence of words, these words are processed by the neural network and converted into a sequence of letters representative of the words. The sequence of letters are then displayed on the display 78 such that a user 73 of the device can see the words or phonetic sounds that are spoken, and thereby better understand the speech. Preferably, the device 70, including the display 78, is carried by the user 76, similar to the manner a hearing aid is carried by an individual. The display 78 may be built-in to a pair of glasses worn by the user 73, with the sequence of detected words being displayed on an inside surface of the glasses where only the user 73 can see it.
FIG. 8 shows a block diagram of a vibration table 80 and associated controls used in conducting Examples 1 and 2, described below.
Appendix A contains the code listing used to program the neurocomputer of FIG. 8 in accordance with a MIMO operating mode of the invention used below in Example 2. The code listing has been prepared in the programming language "C". It includes many comment statements and is believed to be self-explanatory when considered in view of the commercially available Sigma neurocomputer available from SAIC and the manuals accompanying such neurocomputer.
Example 1
In order to demonstrate an application of neural networks to real time adaptive control, a neural network based controller was used to cancel vibration in a complex structure. The top of the vibration table 80 (FIG. 8) represented the complex vibrating structure. It comprised a flexible frame made from unistrut beams that were bolted together, having approximate dimensions of three feet by four feet. The construction of the frame made its vibration response non-linear and not easily modeled. Several containers of colored liquid, such as the container 82 containing colored liquid 84, were placed around the structure 80 to reveal the vertical vibration of the structure at various points. The frame or table 80 was mounted on rubber pads (not shown) at four points to isolate its vibration from appropriate supporting structure.
Attached to the center of the frame 80 was an electromagnetic shaker 86. The shaker 86 served as the source of vibrations to the frame 80. It thus simulated a large piece of machinery that might be causing problem vibrations in the frame or structure 80. Attached to one corner of the table frame 80 was another electromagnetic shaker 88. This shaker 88 was controlled by a neural network (neurocomputer) 92, through an amplifier 94 and a D/A converter 96, to cancel vibration at a given monitoring location on the table frame 80. An accelerometer 90 was used to sense the vibrations at the location where vibration cancellation was desired. Its output signal became the error signal applied to the neural network 92.
When the control signal to the corner shaker 88 was turned on, the vibration near the accelerometer 90 was canceled, as shown by a calm surface on the colored liquid 84.
The adaptive control system thus demonstrated was implemented on SAIC's Sigma work station, a 386 AT based computer having the Delta II floating point processor card installed therein. The Delta card has been specifically designed for high speed neural network computation, making real time control with networks possible. The Sigma system shown in FIG. 8 also contained a data acquisition card to interface the 386 AT computer to analog inputs from the accelerometers, such as accelerometer 90, and to the analog outputs to the shakers.
The block diagram of FIG. 8 helps to visualize the cancellation system and its interconnections. In FIG. 8, the flexible frame table top 80 is shown, along with the source shaker 86. Suitable isolation pads were used to isolate the table top from all other components of the system. The shaker 86 was driven by a wave form generator 98 in an open loop through a suitable amplifier 100.
In this example, a 20 Hz sine wave was used to drive the source shaker 86. The neural network 92, as implemented on the Delta II in the AT computer, required two inputs from the system. One input was from an accelerometer 102 on a reaction mass 104 of the source shaker 86. The other input came from the accelerometer 90 at the location where vibration cancellation was desired, and acted as the error signal for the network. The output of the network controlled the cancellation shaker 88 at the corner of the frame 80.
The block diagram of the neural network used in this example showed the back propagation paradigm with four layers, as shown in FIG. 2. The inputs from the source accelerometer 102 were sequentially loaded into a shift register 35 stages long. This formed the input layer. There were two hidden layers of 10 elements each and one element in the output layer. These layers were fully interconnected with some 540 weights, including bias weights.
The signal from the source accelerometer 102 fed forward through the time series input layer to form the output of the neural network. The signal from the error accelerometer 90 becomes the error term for the back propagation algorithm. This error was back propagated to adjust the weights of the network interconnections. This allowed the network to learn and adapt to changes in the control conditions. The entire process of signal acquisition, network feed forward, weight update, and final output to the shaker took less than a millisecond. This high speed processing capability allowed true wave form cancellation in real time.
To demonstrate the adaptive capabilities of the controller used in this Example, the error accelerometer 90 was first located in the corner of the frame 80 and cancellation control was turned on. The liquid surface smoothed as vibration was canceled at this point. Next, the error accelerometer 90 was moved to the middle of the side of the frame 90 near point P.sub.A. The controller adapted its output to the new situation, calming the bottle of liquid (not shown) near point P.sub.A, while the corner bottle 84 shook violently. Significantly, vibration cancellation was accomplished at a point remote from the cancellation shaker 88.
It was also shown that it was possible to cancel the motion of the center frame directly over the vibration source 86. That is, the accelerometer 90 was moved to point P.sub.B. The corner shaker 88 was able to flex the frame 80 sufficiently to balance the force from the source shaker 86. The liquid surface of a bottle (not shown) at point P.sub.B was calmed in just a few seconds.
The above adapting to a new location was true for the case where the weights (used within the neural network) had already been trained and only needed to be varied incrementally for the different positions. Training the weights from scratch took longer.
Example 2
This Example demonstrates vibration cancellation at multiple locations simultaneously. The same basic test configuration was used as is shown in FIG. 8, except that the frame table 80 was modified by installing a cancellation shaker at each corner with the source shaker 86 installed near the center as in Example 1 (FIG. 9). There were four error accelerometers, one at each corner of the frame 80, and one source accelerometer 102. All of the accelerometers were connected to the neuro-controller 92, which in turn provided four outputs to the four cancellation shakers.
Table 1 provides a list of the equipment used in Example 2. Table 2 shows the operating specifications of the vibration cancellation system thus realized.
The multiple-input, multiple-output, neurocontroller described in Tables 1 and 2 learned and canceled vibrations at the four locations simultaneously. Moving the four error accelerometers to the center, above the source shaker near point P.sub.B, caused the system to adapt such that the cancellation shakers coordinated to cancel the vibration at the remote center location. The neural-controlled vibration cancellation process was exactly the same as the single shaker example (Example 1) described above, simply expanded to multiple cancellation locations.
Adaptive cancellation is, of course, just one of the applications of real time neural network control made possible by the neurocomputer (Sigma work station) 92 used in the above examples. Its Delta II floating point processing board, with 12 million bytes of memory and 22 million floating point operations per second of processing speed, opens up new horizons in the field of adaptive control.
As described above, it is thus seen that the present invention provides vibration control systems adapted to control vibrations in real time for a wide variety of applications. One embodiment, for example, provides a vibration control system that cancels or otherwise conditions unwanted vibrations at a specified location or target point in a given medium. Another embodiment provides such vibration cancellation or conditioning at a site remote from a cancellation shaker (transducer). Yet another embodiment provides simultaneous vibration cancellation or conditioning at multiple monitoring points within the medium.
It is also seen from the above description that the present invention provides a vibration control system that automatically adapts to provide a desired vibration cancellation or conditioning at one or multiple target points within the medium, regardless of whether the target points remain stationary or move within the medium.
It is further seen that the invention provides the desired vibration control in real time, without the need to mathematically model the system or medium, and without having to program or carry out time-consuming complex algorithms.
It is further seen that the invention can be generalized to provide effective adaptive control to a wide range of diverses processes.
While the invention herein disclosed has been described by means of specific embodiments and applications thereof, numerous modifications and variations could be made thereto by those skilled in the art without departing from the scope of the invention set forth in the claims.
Claims
What is claimed is:
1. Adaptive vibration control apparatus comprising: (1) source sensing means for sensing source vibrations in a given medium; (2) error sensing means for sensing vibrations at a specified monitoring point in the medium whereat vibrations are to be controlled; (3) vibration generating means for generating at least one offsetting vibration that is applied to the medium at at least one location; and (4) controller means responsive to the source vibrations sensed by the source sensing means and to the vibrations at the specified monitoring point sensed by the error sensing means for controlling the vibration generating means in real time so as to force the vibrations sensed by the error sensing means below a first level, said controller means comprising an artificial neural network, said artificial neural network having: an input layer of a plurality of neurons to which time sequences of an input signal derived from said source sensing means is applied said input signals propagating forward through said network; an output layer of at least one neuron providing an output signal for controlling said vibration generating means; at least one intermediate layer of a plurality of neurons coupled to said input and output layers; said input, output and at least one intermediate layers being responsive to at least one error signal derived from said error sensing means so as to adapt the amount of coupling that occurs between said input, intermediated said output layers as said input signal propagates therethrough.
2. The adaptive vibration control apparatus as set forth in claim 1 further including means for generating the derivative of said input signal and applying time sequences of said derivative to a portion of said input layer.
3. The adaptive vibration control apparatus as set forth in claim 2 wherein said artificial neural network further includes at least one set of a plurality of direct connections between said input layer and said output layer.
4. The adaptive vibration control apparatus as set forth in claim 3 wherein said artificial neural network includes two intermediate layers, each intermediate layer having at least eight neurons.
5. The adaptive vibration control apparatus as set forth in claim 4 wherein said first level below which said sensed vibration is forced by said controller means comprises a vibration level approximately equal to zero, whereby said vibration control apparatus effectively cancels the sensed vibration at the specified monitoring point.
6. The adaptive vibration control apparatus as set forth in claim 4 wherein said first level below which said sensed vibration is forced by said controller means comprises a time-varying reference value that changes as a function of time, whereby said vibration control apparatus functions as a vibration generator that generates a vibration that follows said time-varying reference value.
7. The adaptive vibration control apparatus as set forth in claim 1 wherein said controller means includes: means for emulating said vibration generating means using a first artificial neural network; a second artificial neural network coupled to said first artificial network, said second artificial network being adapted to control said vibrating generating means; wherein each of said first and second artificial neural networks comprises: an input layer of a plurality of neurons to which time sequences of an input signal is applied, said input signal propagating forward through said network, an output layer of at least one neuron providing an output signal at least one intermediate layer of a plurality of neurons coupled to said input and output layers; wherein said input, output and at least one intermediate layers, are responsive to at least one error signal so as to adapt the amount of coupling that occurs between said input, intermediate, and output layers as said input signal propagates therethrough; and further wherein the input signal applied to the first artificial neural network is derived from the error sensing means, the input signal applied to the second artificial neural network is derived from the source sensing means, the output signal of the second artificial neural network drives said vibration generating means, the output signal from the first artificial neural network is compared with the output signal from the second artificial neural network to generate a first error signal that is applied to the first artificial neural network, and the input signal applied to the first artificial neural network is compared with the input signal applied to the second artificial neural network to generate a second error signal that is applied to the second artificial neural network.
8. The adaptive vibration control apparatus as set forth in claim 1 further including a first artificial neural network (36x) that emulates said vibration generating means (34); a second artificial neural network (32x) that controls said vibrating generating means; a third artificial neural network (40x) coupled to said first artificial neural network; and a fourth artificial neural network (38x) coupled to said second artificial neural network; wherein each of said first, second, third and fourth artificial neural networks comprises: an input layer of a plurality of neurons to which time sequences of an input signal is applied, said input signal propagating forward through said network, an output layer of at least one neuron providing an output signal, at least one intermediate layer of a plurality of neurons coupled to said input and output layers; weighting means for defining the amount of coupling that occurs between said input, intermediate, and output layers as said input signal propagates therethrough; and further wherein the second artificial neural network receives its input signal from the source sensing means (A0), and directs its output signal to the vibration generating means, and wherein the first artificial neural network receives its input signal from the output signal of the second artificial neural network and compares its output signal with an sensed error signal derived from said error sensing means (39) to produce a first error signal, said first error signal being applied to the weighting means of the first artificial neural network; and wherein the third artificial neural network receives its input signal from the source sensing means, and directs its output signal to the fourth artificial neural network; and wherein the fourth artificial neural network receives its input signal from the output signal of the third artificial neural network, and wherein a second error signal is applied to the weighting means of the fourth artificial neural network, the second error signal being derived from the source sensing means and error sensing means; wherein the weighting means of the second artificial neural network track the weighting means of the fourth artificial neural network; and the weighting means of the third artificial neural network track the weighting means of the first artificial neural network.
Patent Citations (13)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US3829750 | 1974-08-01 | Centner et al. | |
| US4209776 | 1980-06-01 | Frederick | |
| US4450530 | 1984-05-01 | Llinas et al. | |
| US4823053 | 1989-04-01 | McCracken et al. | |
| US4857912 | 1984-08-01 | Everett, Jr. et al. | |
| US4999534 | 1991-03-01 | Andrianos | |
| US5023045 | 1991-06-01 | Wantanabe et al. | |
| US5092343 | 1992-05-01 | Spitzer et al. | |
| US5093792 | 1992-03-01 | Taki et al. | |
| US5138924 | 1992-08-01 | Ohya et al. | |
| US5146541 | 1992-09-01 | Speidel | |
| US5159660 | 1992-10-01 | Lu et al. | |
| US5233540 | 1993-08-01 | Andersson et al. |
Non-Patent Literature (6)
- Wang et al, "Self-Adaptive Neural Architectures for Control Applications", International Joint Conf. of Neural Network, Jun. 17-21, 1990.
- Bleuler et al, "Nonlinear Neural Network Control with Application Example", Inter. Neural Network Conf. (INNC), Jul. 9-13, 1990.
- Psaltis et al, "A Multilayered Neural Network Conroller"; 1988 IEEE.
- Hecht-Nielsen, Robert, "Neurocomputing: Picking the Human Brain", IEEE Spectrum, pp. 36-41 (Mar. 1988).
- Widrow, et al., Adaptive Signal Processing, pp. 285-297 (Prentice-Hall, 1985).
- Graf, et al., "VLSI Implementation of a Neural Network Model", Computer, pp. 41-49 (Computer Society of IEEE, Mar. 1988).