Motor Vehicle Fuel Analyzer

1. Field of the Invention

The invention relates generally to systems for identifying organic liquids, and more particularly to a field-utilizable analytical instrument using a semiconductor gas sensor array and neural network processor to identify and characterize liquid fuels.

There is an increasing demand for small, rugged and portable instruments to identify organic substances in the field. In particular, there is need for a field instrument capable of rapid, on-site analysis and identification of aviation and automotive fuels. Some of the applications for such an instrument are ensuring that an aircraft is refueled with the correct grade of fuel, checking gasoline at the point of sale for alcohol and/or octane number (ON), and verifying that fuel storage facilities contain the proper grades of fuel. All of these needs as well as other safety-related applications would be met by an instrument capable of quick determinations of fuel type and octane number in the field.

The usual test for octane number is the rigorous CFR test which is based on running the fuel through an internal combustion engine, and requires sophisticated equipment and hours of setup time. Laboratory-grade instruments normally used in conjunction with octane number measurements may involve infrared spectrum analysis, multivariate regression analysis, flame reactions, mass spectrometers, gas chromatographs, microprocessor-controlled octane analyzers, and others. The laboratory instruments often have the capability to analyze a large number of compounds at low concentrations, but generally are too cumbersome for field use and often require time-consuming sample preparation as well.

2. Description of the Prior Art

The first examples of prior art involve gas sensor arrays coupled with pattern recognition approaches to identify unknown gas samples. The first of these (P. K. Clifford, "Selective Gas Detection and Measurement System", U.S. Pat. No. 4,542,640, Issued Sep. 24, 1985), is a chemometric approach that relies on the solution of linear equations. Clifford's method is unsuitable when dealing with liquid fuels, which may have tens or hundreds of individual organic constituents. Clifford's approach also requires that the number of sensors be greater than or equal to the number of unknown gases. The second example (A. Ikegami and M. Kaneyasu, "Olfactory Detection Using Integrated Sensor", Proc. of Transducers '85, 1985 Intl. Conf. on Solid State Sensors and Actuators, Philadelphia, Pa., 136-139, 1985), uses a microcomputer to mathematically calculate the similarities in patterns of certain essentially pure substances that are not necessarily chemically similar.

Recent work by B. S. Hoffheins at the Oak Ridge National Laboratory showed that the problem of qualitative gas analysis of a few essentially pure substances can be greatly simplified through the use of semiconductor gas sensors together with pattern recognition and/or neural networks (B. S. Hoffheins, Master's Thesis, University of Tennessee, 1989). Hoffheins reported results of using a nine sensor array with a neural network to analyze the patterns of a six item group comprising isopropanol, methanol, ethanol, heptane, hexane, and hexane-2% ethanol. In addition, an integrated sensor array was used to obtain patterns of one gasoline sample and samples of the same gasoline with either methanol or ethanol added to it. Although slight differences were noted in the patterns, no attempt was made to classify the gasoline samples with a neural network.

In 1991, Gardner and Bartlett in separate papers reported results similar to Hoffheins (J. W. Gardner, "Detection of Vapours and Odours from a Multisensor Array Using Pattern Recognition Part I. Principal Component and Cluster Analysis", Sensors and Actuators B, Vol 4, 109-115, 1991), (P. N. Bartlett, "The Design of an Artificial Olfactory System, " Paper no. 253, Pittsburgh Conference, Chicago, March 1991).

Also in 1991, anonymous authors reported applying a neural network to the spectroscopic analysis of infrared data ("Determination of Fuel Properties", Research Disclosure, pp. 571-572, 1991). This last reference appears to be the first attempt to produce an instrument designed specifically for determination of octane number by indirect means, that is, without actually burning the fuel in an engine and observing knock characteristics.

It is a principal object of this invention to provide a system for rapid and convenient analysis of complex gas mixtures that are chemically similar.

Another principal object of this invention is to provide a selective analysis system for fuels, particularly those used in aviation and automobiles.

A third principal object of this invention is to provide a selective analysis system for classifying the signatures of a plurality of fuels using a fewer number of sensors than fuels.

It is also an object of this invention to provide a rapid analysis system for classifying fuels according to octane number or suitability for a particular vehicle, engine, or class of vehicles.

Another object of the invention is to provide a fuel inspection method that is more convenient than the CFR engine method.

Yet another object of this invention is to provide a fuel analysis system that eliminates tedious calibration processes and extensive numerical calculations.

A further object of the invention is to provide an analysis system for complex organic substances that is not limited to simple, essentially pure compounds.

The rapid fuel analysis instrument works on the principle that an array of gas sensors will react to fuel vapors in such a way that each fuel has a distinctive signature. The signatures are then "learned" by a neural network. When an unknown sample is tested, the neural network examines the signature and determines the closest match to the known patterns. The sensor head contains ten commercial metal oxide gas sensors. The measured parameter is the resistance of the metal oxide, which decreases in the presence of organic or reducing vapors. Nine of the sensors are nominally identical, except their heaters are powered at different levels. This causes the sensors to react differently to the vapor sample present in the sensor chamber. The tenth sensor has a different combination of metal oxide and catalyst; consequently, its characteristic response can be quite different from the other sensors. The signature for a sample comprises the resistance readings taken from all ten sensors simultaneously. Signatures from the sensor array are scaled by an analog-to-digital (A/D) converter and transformed into binary bit patterns, one for each sample gas. Each pattern consists of a coded matrix, which is an appropriate size for the problem at hand. A pattern size used was 1.times.16 for each sensor, or 10.times.16 for the ten sensors. A smaller pattern size can be used if fewer sensors are needed to determine the identity of substances in a particular application. Alternatively, the size can be larger to increase the power of resolving the different signatures. A collection of these bit patterns from each of the possible substances in a problem set, called training vectors, is used to train the neural network. A neural network simulation program called ANSim.TM., developed by Science Applications International Corporation, San Diego, was used. Once the neural network "learns" the training vector set, it can identify patterns from unknown samples. Collecting the signature of an unknown sample takes about five minutes; identification of the signature is virtually instantaneous. Samples require no special preparation.

FIG. 1 is a block diagram showing an embodiment of the present invention;

FIG. 2 is a flowchart of the method of this invention;

FIG. 3 is a graph illustrating the signatures of three civil aviation fuels;

FIG. 4 is a graph illustrating the signatures of four military aviation fuels;

FIG. 5 is a graph illustrating the bit patterns of the seven aviation fuels of FIGS. 3 and 4.

FIG. 6 is a graph illustrating the signature of an 87 ON motor fuel;

FIG. 7 is a graph illustrating the signature of an 89 ON motor fuel;

FIG. 8 is a graph illustrating the signature of a 92 ON motor fuel;

FIG. 9 is a graph illustrating the signature of a 93 ON motor fuel with 10% ethanol; and

FIG. 10 is a graph illustrating the bit patterns of the seven aviation fuels of FIGS. 3 and 4 using only three sensors;

FIG. 11 is a block diagram of a simplified embodiment of the present invention to be used for screening purposes;

FIG. 12 illustrates the use of the present invention to inspect an unknown fuel supply;

FIG. 13 shows one approach for incorporating the present invention directly into a vehicle; and

FIG. 14 is a block diagram of a system wherein the pattern classes are selectable by the user.

FIG. 1 illustrates a gas sensor array 11, data acquisition system 12, and data analysis system 13 for the classification of volatile fuels according to the invention. In the figure, a chamber 14 contains the fuel 15 being tested. If desired, humid air may be passed through the chamber 14 in order to pick up fuel vapor and move it to chamber 16, which contains the gas sensor array 11. Sensing elements comprising commercially-available semiconductor gas sensors 1-10 are exposed to the vapors from the fuel 15. D.C. power is supplied to the heater of each sensor 1-10 through individual voltage divider networks 20-29 connected to a D.C. power supply 30. Each heater network 20-29 can be individually adjusted, a useful feature for tuning each sensor's sensitivity for a particular problem set. Sensor outputs are D.C. biased (31) to conform to the input requirements of a multiple channel analog-to-digital (A/D) data acquisition board 32 in a personal digital computer 33.

In the computer 33, a commercial neural network simulation program 34 called ANSim.TM. was used for classification of the fuel signatures (ANSim.TM., Science Applications International Corp., San Diego, Calif., 1987). In addition to the commercial neural network processor 34, two other computer programs are used in the invention. A bit pattern conversion program 35, written in the C language, is used to store and transform data from the A/D convertor 32 to bit patterns for use by the neural network 34. A second computer program 36, written in GW Basic, facilitates the data acquisition process in the computer 33. Copies of the two programs are included in the Appendix hereto.

Among the several neural network paradigms available in the ANSim.TM. system, three autocorrelation paradigms (Hopfield, Hamming and Boltzmann) were chosen because they retrieve patterns most like the input pattern presented to them. The operation of these networks also most closely fits the operation of the gas sensor array detector; that is, the stored patterns would be the possible choices to which the input pattern would be compared. Generally, the Hamming and Boltzmann networks represent improvements over the operation of the Hopfield network and they have better success with the identification problem. The final element of the invention is a display means 37 for presenting the results of the classification to the user. The ANSim.TM. program provides side-by-side graphical displays of the sample pattern being tested and the training vector or pattern it most closely resembles. The ANSim.TM. program also provides a listing of names of all the samples the network has been trained on, and indicates which sample is the closest match to the sample under test.

One sensor array used in tests to date comprised 10 discrete commercial Taguchi-type tin oxide sensors. A Taguchi-type metal oxide sensor exhibits a large change in resistance in the presence of gases to which it is sensitive. The resistance of the sensors used in this work were observed to decrease by four to five orders of magnitude when exposed to alcohol vapor, for example. Nine of the sensors (1-9) were the Figaro Engineering Model TGS 812, and were operated at power levels ranging from 100% (sensor 1) to 27% (sensor 9) of the manufacturer's recommended operating power (900 mW). Sensor 10, a Figaro Model TGS 824, was operated at 100% power. Table I gives the sensor power levels in detail.

This is only one of many sensor arrangements. In addition to using nominally identical sensors operated at different power levels (and therefore different temperatures), different models of sensors can be used in the array. Many examples of sensor arrays composed of discrete sensors and miniature integrated sensor arrays on a single substrate have been described in the literature. Clifford, for example, described the value of homogeneous semiconductor gas sensors, sensors that vary according to their composition, and sensors that vary according to their operating temperature, etc. The general rule is to ensure that each element (sensor) of the array will have different response characteristics for the problem set of gases being examined. With such a variety of sensors and operating modes, the user can have wide control over the range of sensitivities of the different sensing elements, thereby allowing the method to be optimized for a given analysis problem.

The method of operation of the invention will be more dearly described with reference to FIG. 2. FIG. 2 shows the process by which parameters for configuring an instrument for a specific application are obtained. The process consists of a series of steps, which are taken in order, and which may require multiple passes through all or part of the steps in order to refine and optimize the system for a particular instrument application.

The process begins with the selection of a sensor array (step 40). Sensors that are known to be sensitive to the problem substances should be chosen, as well as sensors that respond less strongly. Also, at this step, several sensors of one model might be included in the array, but each one operated at different temperatures as previously described, in order to study the range of possible responses. At this initial step 40, one should not be concerned with limiting the size of the sensor array. Following sensor selection, the sensor array is tested with the possible or typical species expected in the problem system and a library of signatures is developed (step 41, see also FIGS. 3 and 4). At this step, sensors that do not respond at all might be dropped from the array. Also, sensors giving redundant information could be eliminated. The next step, forming training patterns (step 42), is optional if the neural network can accept as training patterns the kind of data represented by the signatures; otherwise, some conversion routine is performed to transform the data into input to the neural network 34. In the case of the ANSim.TM. neural network software, such a step 42 was required, and Applicants' bit pattern conversion program 35 was written to transform the signatures to two-dimensional bit patterns for input to the neural network 34. Whether in the form of signatures, bit patterns, or some other suitable arrangement, training patterns consist of a set of patterns, each of which represents a particular outcome or combination of outcomes for presentation to the neural network 34.

Some neural network training systems require only one training pattern per class, others may accept several patterns offered as members of one class. In any event, the training step 43 is part of the operation of whichever neural network 34 is used. In this step, the training patterns are presented to the neural network which "learns" the patterns and thereby defines the limit of each class. The ability of the neural network 34 to classify the set of training patterns is evaluated at this step by checking its ability to classify its own training set and then by testing its ability to identify "unknown" samples as well. If the neural network does not perform adequately, one may elect to a) improve the uniqueness of the sensor array responses by changing the sensors or their operation (return to step 40), b) redesign the library of signatures and training patterns for improved performance (steps 41 and/or 42), or c) examine the neural network for improved performance. If the neural network classifies the full signatures, and instrument simplicity is also important, the sensor array should be examined to determine which sensors can be eliminated without degrading neural network performance.

It is expected that several passes through steps 40-43 will be required, during which the sensor parameters are modified until an optimized system is attained. It is realized that in general, the better the neural network is at classification, the more drift and variation it can tolerate from the sensor array patterns. In contrast, if the information from the sensor array is very reliable, and each signature is very distinctive, then the neural network does not have to be as powerful. With the neural network "trained", the only step remaining is that of analyzing an unknown sample (step 44), a step that is inherent in any neural network program, but which comprises analyzing the unknown sample against the library of signatures or training patterns as may be the case, and classifying it with respect to the storm signatures or training patterns.

In all of the instrument applications cited above, one has access to a large concentration of the fuel vapors of interest, for instance, in the headspace of a storage tank. The use of high concentrations of vapor makes the process somewhat faster and more reproducible. It is, in principle, possible to use the inventive technique with lower vapor concentrations provided that the neural network has been trained with signatures of each fuel at each expected concentration, and the signature of one fuel at one concentration does not resemble that of a different fuel at another concentration. There is, in any case, a lower concentration limit below which classification will become unreliable.

Aviation fuels

The aviation fuels examined here (Table 2) include several civilian and military grades of jet fuel as well as a high-octane low lead gasoline.

Although most jet fuels are based on the kerosene fraction (150.degree.-288.degree. C. boiling range), the actual composition of a fuel is determined to a large degree by the nature of the crude oil from which it is derived. For aviation grades, the kerosene fraction contains at least 75% saturates and no more than 25% aromatics. Depending on the crude, the saturated fraction (75%) contains from 10 to 60% straight-chain paraffins with the balance composed of various cycloparaffins. In the aromatic fraction (25%), the ratio of single- to multi-ring compounds also varies considerably.

About 95% of civil aviation uses a kerosene-type fuel (Jet A and Jet A-1). Jet A-1 is similar to Jet A but has a slightly lower freezing point for long-duration international flights. A wide-cut fuel (Jet-B) is available for severe arctic use.

A wide-cut fuel (JP-4) was developed by the U.S. Air Force in order to assure the availability of large volumes of fuel during national emergencies. The U.S. Navy uses kerosene with a higher flash point (JP-5) because of the special safety requirements for use on aircraft carriers. JP-7 is a high flash point special kerosene used in advanced supersonic aircraft and is available in both regular and thermally stable versions. JP-8 is a developmental kerosene fuel similar to Jet A-1.

Aviation gasolines also vary somewhat depending on the source of the crude oil and refining method. Relevant properties are specified in ASTM specification D 910-82.

Motor fuels

Gasoline, like jet fuel, is a highly variable mixture whose chemical composition and properties can vary widely depending on crude source, refining process, seasonal factors, etc. Among the many fuel samples that were collected, four grades were examined in detail to determine if the method could detect the two most basic differences, namely octane number (ON) and the presence of alcohol. Three grades of unleaded gasoline were used representing the octane numbers 87, 89, and 92, as well as one grade of gasoline containing 10% ethanol.

10 Taguchi-Type Sensors, 7 Aviation Fuels

The uniqueness and reproducibility of the signatures of several aviation fuels can be seen by inspection of FIGS. 3 and 4. In these figures are plotted on a common scale the responses (signatures) of the ten-sensor array described above to three commercial aviation fuels and four military fuels, respectively. Each signature represents the maximum range of 5- and 10-minute readings observed in several runs (typically three runs of each fuel). The ran-to-run variations in any one signature, indicated by the shading, are small relative to the differences between one signature and another. Perhaps the most striking result is the difference between the signatures of Jet A with and without 0.1% deicer. These two fuels look and smell alike, and therefore could not be readily distinguished by an operator without some kind of analysis. The signature of aviation gasoline is substantially different from the kerosene-type fuels as one would expect. The signature of JP-4 showed some similarity to that of gasoline, as expected, since it contains significant amounts of naphtha-range light hydrocarbons. FIG. 5 shows the effect of step 42, i.e., that of transforming the signatures into bit patterns by the bit pattern conversion program 35. These bit patterns are sufficiently different and could be learned and recognized by the neural network.

10 Taguchi-Type Sensors, 4 Motor Fuels

The same ten-sensor array described in the previous example was tested with three grades of unleaded gasoline, 87 ON, 89 ON, 92 ON, and gasoline/ethanol blends from different gasoline suppliers. The signatures, FIGS. 6-9, were converted to 10.times.16-bit patterns as before. Patterns derived from signatures of Shell RU-2000 (87 ON), SR-2000 (89 ON), SU-2000 (92 ON), and gasoline containing 10% ethanol (Citgo 93 ON), respectively, were used to train the neural network. The neural network easily distinguished between straight gasoline and alcohol-containing blends. Surprisingly, the network correctly classified fuel samples having the highest and lowest ON, even through the raw signatures of these fuels look very similar to the eye (FIG. 6 and FIG. 8).

7 Taguchi-Type Sensors, 4 Motor Fuels

A seven-sensor array was constructed from the ten-sensor array in the previous example by eliminating sensors 8, 9, and 10. The pattern size was 7.times.16 bits. Surprisingly, results of the seven sensor configuration were not greatly different from that of the 10-sensor array described earlier. These results are unexpectedly good given the size of the pattern and the visual similarity of all the straight gasoline fuel signatures. With larger sample sizes, the performance of this instrument would be further improved.

3 Taguchi-Type Sensors, 7 Aviation Fuels

A three-sensor array was constructed by eliminating four additional sensors from Example 3 so that only sensors 1, 6, and 10 remained. The 3.times.16 bit patterns for the same aviation fuels as Example 1 are shown in FIG. 10. The data for sensors 1, 6, and 10 in FIGS. 5 and 10 are the same. Surprisingly, the neural network trained with the three-sensor data set (the 3.times.16 bit patterns) from the aviation fuel signatures performed fairly well. Examples 3 and 4 demonstrate that Clifford's requirement in U.S. Pat. No. 4,542,640 that the number of sensors exceed the number of unknowns does not apply in Applicants' neural network method when the substances of interest have distinct and reproducible signatures.

3 Taguchi-Type Sensors, 4 Motor Fuels

Because the differences in the signatures of motor fuels are more subtle, this particular three-sensor array was not able to classify automobile gasolines as to their octane level. It did, however, successfully differentiate between gasoline and gasohol.

In focusing on the development of sensor arrays and pattern recognition system for portable instruments, Applicants recognize that minimization of power consumption and simplification of the pattern recognition system are two important concerns. Most of the power consumption is by the metal-oxide type sensors, which typically consume 300-900 mW when operated at full power. Power consumption can be reduced by minimizing the number of discrete sensors in the array or by miniaturizing the array.

Incorporating the pattern recognition approach into hardware might be done several ways. A sensor testbed, such as described herein, can help identify a viable approach and define system parameters. Special-purpose devices such as neural network chips are becoming available and, if successfully implemented, would replace several general-purpose chips, thereby simplifying the overall design. An Intel chip ("80170NW Electronically Trainable Analog Neural Network," Experimental Report, Intel Corp., Santa Clara, Calif., May 1990) is a more general purpose device that can be programmed many ways. A Micro Devices chip ("MD1220 Neural Bit Slice," Data Sheet, Micro Devices Corp., Lake Mary, Fla., March 1990) could be used for a problem set of eight patterns, or up to eight chips could be connected to address a problem set of 256 patterns. Chemometric algorithms such as Clifford's, sized properly, could be coded into microprocessor devices. Also, both neural network and chemometric algorithms, sized for the problem set, could be incorporated into application-specific integrated circuits. Besides a portable battery-powered implementation of this invention using miniaturized components, the technology could also be incorporated into aircraft and other large vehicles (such as tanks) to provide a check against inadvertent misfueling. A possible large-scale use is in regulatory procedures such as spot-checking gasoline at the point of sale for improper labeling as to octane rating and alcohol content.

As an analytical instrument, a neural network and optimized sensor array using this methodology can classify an unknown sample in five to ten minutes. Compared with the rigorous "Comparative Fuel Research Engine Test for Octane Number", and "Standard Test Method for Knock Characteristics of Motor Fuels by the Motor Method," which require sophisticated equipment, running the fuel in an internal combustion engine, and hours of setup time, the neural network method of this invention is a time-saving and relatively inexpensive arrangement for regulatory screening tests in the field. Suspect samples could be taken back to a laboratory for testing by the CFR engine technique, if desired.

As will be apparent from the discussion herein, minimizing the number of sensors will minimize power consumption and data processing requirements. In some instances, a few sensors can produce a set of signatures which can be recognized by very simple circuitry without even applying the invention at all. For example, two sensors may be enough to unambiguously distinguish gasoline from gasohol as was predicted in B. S. Hoffheins' Master's Thesis. In such a case, there would be no pattern recognition as such, and no use of a neural network.

The sensor array approach can be successfully applied to complex mixtures such as fuels. Although the sensor arrays in this work were not optimized, relatively simple neural network techniques can classify aviation and motor fuels. With signatures that exhibit a high degree of variability, such as the aviation fuels, successful classification could be obtained by using very few carefully chosen sensors. With signatures that are much closer in range and behavior, such as those of the motor fuels, a greater number of sensors is needed for reliable classification. Surprisingly, a 7-sensor array performed as well as the 10-sensor array for the classification of a set of motor fuels, indicating that three of the sensors were providing redundant information.

One safety-related application of the detecting system of this invention may include a plurality of sensor arrays, each in its own sensor head and dispersed in separate locations about an environment. A central computer would be responsive to signals derived from all the sensors of all the heads to determine the identities of mixtures present at each of the sensor head locations. It will be understood that fuels such as gasoline, diesel fuel, aviation fuel, or gasohol are considered to be within the scope of this invention. Also, the liquid fuels classification may be based on octane number, octane number, or alcohol content.

Using the method of FIG. 2, optimization of the sensor array and pattern recognition methods can be performed iteratively to develop and verify an application-specific system. As the set of patterns becomes more distinctive, it is likely that less resolution is needed by the pattern recognition portion; alternatively, the better the pattern recognition section, the more ambiguous the sensor output can be. With appropriate testing and characterization of the sensor arrays, and a suitable database of signatures, one can fully optimize a miniature field instrument for any given analysis and identification problem.

There are many instances in which the analysis of an unknown fuel must address only one question, viz., whether or not that fuel can be used safely in a particular vehicle. For example, highly mobile military forces wishing to exploit a captured fuel supply must determine if the fuel is suitable for a helicopter, armored vehicle, etc. For these applications, the device and method can be further simplified as shown in FIG. 11. The device 50 comprises a sensor array 11, a data acquisition and conversion system 51 which may or may not contain an analog-to-digital convertor, a pattern classifier 52, and an output system 53 capable of providing a go/no-go type of display (e.g., red and green LEDs or the like). In this case, therefore, the pattern classifier places the sensor array output into one of only two classes, viz., fuels of the desired type and fuels not of the desired type. It will be obvious to those skilled in the art that the pattern classifier can be made more or less selective depending on the demands of a particular application. The go/no-go output eliminates the need for operator judgment or interpretation.

FIG. 12 shows an embodiment of the present invention in which a technician inspects a tank of captured fuel in support of a helicopter awaiting refueling. In this embodiment, the device is completely portable and self-contained.

The present invention can also be incorporated directly into a vehicle, preferably as an integral part of the fueling system. An embodiment of this device is shown in FIG. 13. In the system 70, a sensor array 11 is disposed in the chamber 71 such that the array 11 is exposed to fuel vapor but not to direct contact with the liquid fuel in the tank 72 or filler neck 73. Valves 74 and 75 are disposed to selectively isolate the filler neck 73 from the tank 72 and drain 76. A, small sample of fuel to be tested 77 is introduced into the neck 73 after closing valves 74 and 75. Vapor from the fuel sample 77 accumulates in the chamber 71 allowing analysis of the fuel according to the method of the present invention. If the fuel is determined not to be acceptable, valve 74 is opened to discharge the sample through drain 76. If the fuel is determined to be acceptable, valve 75 is opened and fueling can safely commence.

Another embodiment of the present invention incorporates features of both the general instrument, FIGS. 1 and 2, and the vehicle-specific instrument, FIGS. 11 and 12. In this embodiment, FIG. 14, the pattern classifier can be switched among various preprogrammed settings, each of which corresponds to a particular engine or vehicle or a specific type of fuel. The instrument then provides a go/no-go response for the selected classification "window".

"Straight" gasolines, from a variety of sources and with a variety of octane values, have similar signatures as typified by FIGS. 6-8, which are quite different from those of gasoline-alcohol blends as typified by FIG. 9. It would therefore be straight forward to classify commercial motor fuels with respect to the presence or absence of alcohol and provide a go/no-go indication thereof.

Motor fuels with octane values from 87 to 92, with or without alcohol, have signatures that are quite different from that of 100 octane low lead aviation gasoline. All types of jet fuel also have signatures much different from that of gas. A specific pattern classifier set to recognize only 100 octane low-level fuel could therefore protect a piston-engine light aircraft from misfueling with either jet fuel or motor gasoline.

The signatures of Jet A (commercial) and JP-5 (military) are very similar, and different from other aviation fuels. These two fuels can be safely interchanged because of their similar composition and properties and can, at the same time, be distinguished from fuels that cannot safely be substituted for them. An analogous problem is identifying aviation fuels which not only have appropriate de-icers included therein, but also where the de-icer has properly remained in solution. The present invention also solves this problem. ##SPC1##