JP2001209628A - Pattern matching method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To easily find out an optimum value without requiring much labor even when a user has no experience of a neural network by calculating a feature value from an information signal such as an image, sound and vibration in order to conduct product inspection and fault diagnosis, grasping the feature value as normal distribution and learning the neural network. SOLUTION: In a learning signal storing part, a pattern boundary setter prepares a boundary signal around an initial learning input signal belonging to the same category on the basis of an initial learning signal written in an initial learning input signal memory and writes the prepared boundary signal in a learning input signal memory together with the initial learning input signal. The pattern boundary setter prepares a boundary pattern of normal distribution around plural learning input patterns belonging to the same category by using the neural network. The boundary pattern obtained by calculating a mean value and a standard deviation from plural learning input patterns belonging to the same category and changing a part or all of the learning input patterns to a value obtained by adding or subtracting the real number times of the standard deviation to/from the mean value is used.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern matching method, and more particularly, to an image, a print pattern inspection, a product inspection of a rotating machine such as an engine or a motor, or a failure diagnosis of a rotating machine or a plant. The present invention relates to a pattern matching method in which a feature amount is calculated from an information signal such as sound and vibration, the feature amount is grasped as a pattern, and the quality is determined by using a neural network.

[0002]

2. Description of the Related Art Conventionally, in an automobile engine assembly plant, after assembling the engine, the inspector listens to the rotation of the engine and hears the rotation noise. It is determined whether or not it has been done. Since this judgment is a sensory test based on the sense of sound of the operator, a lot of experience is required to judge the quality of the engine, and there is a possibility that the judgment may vary due to individual differences. There is a worry that it can lead to mistakes. Further, in an automobile engine assembly factory, a plurality of types of engines corresponding to various types of automobiles often flow mixedly on the same line, and it is not easy to distinguish the engine sounds of these various types of engines.

[0003] For this purpose, a learning operation for inputting a pattern to be identified so as to obtain a control output corresponding to various information signals and correcting the corresponding output by comparing it with a teacher pattern is repeated, and thereafter, an arbitrary operation is performed. There has been known a pattern matching apparatus which is configured to perform automatic discrimination by inputting an identification pattern. In general, such a pattern matching apparatus is configured to use a neural network configured by connecting a multiple-input one-output type element (unit) to another unit by a connecting element. However, in the pattern matching apparatus using the conventional neural network processing, when a pattern belonging to a category that has not been learned is identified, an erroneous determination that the pattern belongs to one of the learned categories is often output. This problem is generally a problem in response to unlearned data.

That is, in the conventional pattern matching apparatus, when discriminating between categories A and B as shown in FIG. 8, patterns belonging to these two categories A and B (the ones indicated by "O" or "X") are used. ) Is used for learning. As a result of this learning, a boundary line D such as a broken line shown in FIG. 8 is set. Then, it is determined which category the pattern belongs to, based on which side of the boundary line D the input pattern is. Therefore, it can be correctly determined if the data belonging to A or B is identified, but if a pattern that does not belong to these categories is input, this pattern is forcibly determined to be either category.

Incidentally, the reason for erroneous determination is that, in the case of a conventional input / output function, when a value considerably larger or smaller than the threshold value is input, a value of 1 or 0 is output regardless of the value. It is considered that since a so-called saturated type function is used, there is only an operation of simply setting a boundary between the learned categories.

As a method for solving this problem, the present applicant has proposed a pattern matching method and apparatus in Japanese Patent Application No. 07-319625. According to the publication, as shown in FIG. 7, the pattern matching device includes a pattern input unit 112, a conversion processing unit 114, an error detection unit 132, which are sequentially connected to the image receiving device 110,
A pattern output unit 126, a display device 130, a memory 128 connected to the conversion processing unit 114 and the pattern output unit 126, a learning signal holding unit 150 connected to the image receiving device 110, and a learning signal holding unit It comprises a teacher device 134 disposed between 150 and the error detection unit 132. The conversion processing unit 114 includes an input layer 116
, An intermediate layer 118, and an output layer 120. The learning signal holding unit 150 includes an initial learning input signal memory 156, an initial teacher signal memory 158, a pattern boundary setting unit 154, a learning input signal memory 151,
Further, a teacher signal memory 152 is provided.

The purpose of the pattern matching apparatus using the neural network processing is to enable an unlearned pattern to be properly identified and output, and in particular, to add an input signal multiplied by a weight coefficient in each unit. An appropriate output signal is output not only when the value is close to the threshold value but also greatly away from the threshold value, so that the accuracy of pattern matching determination can be improved. That is, data (shown by “◎”) that gives peripheral boundaries of categories A and B as shown in FIG. 8 are synthesized from learning data, and when these data are given to the input layer, all cells ( This is based on the knowledge that the output value of the unit may be learned to be “0” or a small value. That is, the teacher signal for the input signal corresponding to the boundary pattern is set to a small value over all cells. As a result, the pattern located at the boundary between the two categories is used as the learning pattern, and as a result, the range indicated by the solid line surrounding the learning data is recognized as the range of the category to be identified. However, the misrecognition rate is greatly reduced.

For this purpose, a plurality of learning input patterns are input to the input unit group, and the output pattern of the output unit group corresponding to the input learning input pattern is compared with a predetermined teacher pattern to determine an error between the two. ,
After repeating the operation of correcting the threshold value of the unit constituting each unit group and the weight coefficient of the coupling element connecting each unit group based on this error, the learning input pattern is learned. In a pattern matching method in which an arbitrary pattern is input to the input unit group and an output pattern of the output unit group is compared with an output pattern corresponding to the learned input pattern, the plurality of learning patterns belonging to the same category A boundary pattern around the input pattern is created based on the learning input pattern, a teacher signal of an output of each unit of the output unit group corresponding to the boundary pattern is set to 0 or a small value, and the boundary pattern is After learning together with the learning input pattern, an arbitrary pattern is input to the input unit group. One in which the.

The boundary pattern may be a circle, an ellipse, or a rectangular area including all of the learning input patterns belonging to the same category, and a value outside the area may be set as the boundary pattern. Further, the boundary pattern is obtained by subtracting a predetermined value from a maximum boundary value obtained by adding a predetermined value to a maximum value of each element of a plurality of learning input patterns belonging to the same category, and a minimum value of each element. Setting to include the minimum boundary value facilitates creation of a boundary pattern.

Thus, when a boundary pattern outside the category is input to the input unit group, the output of each unit of the output unit group is learned to be "0" or a small value. As a result, the output of the output unit can be reduced to a small value even for a pattern far from the category. Then, by learning so that the output of the output unit with respect to the input of the pattern belonging to the category becomes large, and comparing the output corresponding to the input arbitrary pattern with the learned output, the pattern is classified into the category. Whether or not they belong can be easily and accurately determined, and erroneous determination can be eliminated. In addition,
The boundary patterns are synthesized based on the learning patterns belonging to each category, and this synthesizing method may be individually performed according to the identification target.

[0011]

SUMMARY OF THE INVENTION However, Japanese Patent Application No.
In the pattern matching method of -319625, the method of synthesizing the learning patterns belonging to each category is performed individually according to the identification target. However, if the number of identification targets is large, the number of steps required for synthesis is increased. And
There is a problem that it is difficult to synthesize. In addition, in the method of creating a boundary pattern, it is necessary to select an optimal value by repeating the work of verifying using a predetermined value, but this work is time-consuming and requires experience,
There is a problem that it is difficult to find the optimum value if the experience of the neural network is small.

The present invention has been made in view of the above problems, and relates to a pattern matching method. In particular, in order to perform product inspection and failure diagnosis, a feature amount is calculated from information signals such as images, sounds, and vibrations. By capturing the features as a normal distribution and learning the neural network, a pattern matching method that allows easy calculation of the optimal value without the hassle and no experience of the neural network. The purpose is to provide.

[0013]

According to the present invention, it is said that the accuracy of each component constituting a product follows a normal distribution, and that characteristic quantities such as sound and vibration generated by the product also follow a normal distribution. It has been done. In other words, by making the learning data of the target category a statistical distribution represented by a normal distribution, even if there are many categories, it does not take much effort, and the optimum value can be easily obtained even if there is no experience with neural networks. Can be Also, the chi-square distribution,
Alternatively, a statistical distribution such as a Rayleigh distribution may be used.

Accordingly, the present invention provides a method for inputting a plurality of learning input patterns to an input unit group, and comparing the output patterns of the output unit group corresponding to the input learning input patterns with a predetermined teacher pattern. And repeating the operation of correcting the threshold value of the unit constituting each unit group and the coupling coefficient of the coupling element connecting each unit group on the basis of this error. After learning a pattern, an arbitrary pattern is input to the input unit group, and an output pattern of the output unit group is compared with an output pattern corresponding to the learned input pattern. A boundary pattern is created around the plurality of learning input patterns by a statistical distribution represented by a normal distribution. Configured to, in which so as to achieve the above object.

The boundary pattern is obtained by calculating an average value and a standard deviation from the plurality of learning input patterns belonging to the same category, and for some or all of the learning input patterns, the average value is a real number of the standard deviation. If a value obtained by adding and subtracting the double is used, the boundary pattern can be easily created.

[0016]

According to the above configuration, in order to perform product inspection and failure diagnosis, a characteristic amount is calculated from information signals such as images, sounds, and vibrations, and a regular amount is calculated around a plurality of learning input patterns belonging to the same category. Create a distribution boundary pattern. Since the normal distribution can describe the features using the average value and standard deviation, the average value and standard deviation of the feature values are calculated for the training data of the target category, so that even if there are many categories, the features can be clearly and easily defined. Can be grasped. Also, the optimum value can be obtained by using a statistical distribution such as a chi-square distribution and a Rayleigh distribution. For example, a sound generated with the rotation of the engine has a wide-band frequency component. Therefore, normality and failure can be distinguished by performing a sonagram analysis of a frequency characteristic and a temporal characteristic and extracting a characteristic amount. The sonagram divides the measurement signal into short-time waveforms and applies the maximum entropy method (ME
M) is used to perform frequency analysis, and these spectra are calculated by arranging them in the time direction. If the sonagram is used as it is as the feature value for the determination, several hundred parameters are required per unit time, but the feature value is expressed by about 5 to 30 parameters by using the reflection coefficient calculated by the maximum entropy method. Yes, the process is simple.
The determination is performed using a three-layer hierarchical neural network having a high pattern matching capability based on the extracted reflection coefficient. When a reflection coefficient for about one second of the characteristic parameter of the engine sound is given to the input layer of the learned neural network, a diagnosis result of the engine state is obtained in the output layer. The pass / fail judgment is performed for one of the cells in the output layer. If the output of the cell is larger than a set threshold value, the cell is judged to be good, and if smaller, the cell is judged to be defective. In the case of a defective product, it is also possible to estimate a defective part.

[0017]

Embodiments of the present invention will be described below with reference to the drawings. Note that the overall configuration of the pattern matching apparatus is the same as the configuration shown in FIG. 7, in the learning signal holding unit 150, the boundary pattern includes a maximum boundary value obtained by adding a predetermined value to the maximum value of each element of a plurality of learning input patterns belonging to the same category, Is set so as to include a minimum boundary value obtained by subtracting a predetermined value from the minimum value. On the other hand, in the present application, in the learning signal holding unit 150, the range of the boundary pattern is set by a normal distribution by extracting features.

Hereinafter, the embodiment will be described with an example using product inspection of an engine. FIG. 1 is a block diagram illustrating a main configuration of the pattern matching apparatus 1 according to the embodiment. FIG.
1, a sound receiving device 12 that collects sound is arranged near the engine 10 to be inspected. Sound receiving device 12
Is constituted by a microphone, collects a sound (engine sound) generated from the engine 10 when the engine 10 is driven, and converts the sound into an electric signal corresponding to the magnitude of the engine sound.
The signal is input to the signal analyzer 14 via an amplifier (not shown). Between the sound receiving device 12 and the signal analyzing device 14,
A control signal input / output device 16 is provided, and the sound receiving device 1 is provided.
The control signal input / output device 16 is disposed so as to process the engine sound collected in Step 2 in accordance with the control steps in FIG.

The signal analyzer 14 comprises a digital filter 20, an analyzer 22, a diagnostic device 24, and a data storage device 26. The diagnostic device 24 is connected to the determining unit 28 and outputs a result of the diagnosis of the engine 10 to the determining unit 24. The determination unit 28 determines the quality of the engine 10 and quantifies and outputs the determination result of the quality. As a result, the sensory test based on the sensation of the operator, which was conventionally performed, is eliminated, and it is not necessary to have much experience to determine the quality of the engine, and there is no variation in determination due to individual differences.
Inspection errors are eliminated when the operator is not tired.

The digital filter 20 has a configuration in which upper and lower limit frequency setting, attenuation setting, and the number of filter taps can be arbitrarily set. The analyzer 22 uses an analysis time, an analysis window length, a reflection coefficient order, a step size, and an acquisition start time as analysis parameters. As an analysis method using the analysis parameters, a feature amount is extracted by a reflection coefficient analysis and a sonagram analysis, and is used as an input pattern of a neural network. As an input diagnosis method, the diagnostic device 24 compares input data obtained by extracting a feature amount by a neural network with a feature amount of a learning pattern, and determines normality / abnormality by pattern recognition.

The control signal input / output device 16 sequentially outputs commands of the control steps shown in FIG. In step 1, a diagnosis start task is output to start diagnosis. In step 2, output the vehicle type recognition task,
The type of vehicle to be measured is determined. In step 3, an engine sound data collection task is output to collect engine sound data. Step 4 outputs the filter task,
The collected engine sound data is transmitted through band filters having different frequency bands to perform frequency analysis. In step 5, the reflection coefficient calculation task is output, and the characteristic amount, that is, the specific coefficient value of the reflection coefficient is obtained from the engine sound subjected to the frequency analysis.
(Extraction pattern) is calculated and obtained. In step 5, a diagnosis task is output, and it is diagnosed whether it is within a predetermined range of the normal distribution obtained from the learning data. The result of the diagnosis is output to the determination unit 24.

FIGS. 3 and 4 show examples of reflection coefficient values, which are characteristic amounts obtained as a result of measurement in accordance with the above. 3 and 4, the X axis represents the type of engine sound, the Y axis represents time, and
The Z axis is indicated by the reflection coefficient value. 3 and 4 show an example in which the engine sound is abnormal. FIG. 3 shows the case where the type of the engine sound is rattling, and FIG. 4 shows the case where the type of the engine sound is gurgling.

In the engine 10, the relationship between the type of the engine sound and the phenomenon that generates the engine sound is known as shown in Table 1.

[Table 1]

As described above, the quality of the engine 10 can be determined by calculating the reflection coefficient value, which is the characteristic amount.
Next, a method for obtaining a reflection coefficient value as a feature amount will be described. First, the reflection coefficient, which is a feature quantity, can be obtained by Expression 1.

(Equation 1) Here, each symbol is defined as follows. C _k : autocorrelation function of engine sound lag kΔt k: sampling point Δt: sampling time interval γ _mk : reflection coefficient P _m : average output from m + 1 point prediction error filter

The following recurrence relational expression 2 holds between the reflection coefficients γ _mk .

(Equation 2) The reflection coefficient γ _mk is calculated as a coefficient that minimizes the average output P _m of the forward prediction error and the backward error. The Burg method is one of the methods for calculating the reflection coefficient. In this method, not only γ _mm and P _m but also the auto-correlation function C _m are sequentially determined by a recurrence relation equation starting from m = 1.

Conventionally, in a neural network, a sigmoid function is often applied as an output function. The sigmoid function has a function to sort a pattern based on a given feature value.However, when performing pattern matching, it is better to adopt a limited output function that determines whether the feature value is closer to a threshold value than the sorting type. good. This apparatus employs the limited output function shown in Equation 3.

(Equation 3)

Here, x: input value θ: threshold value c: constant n: positive even number The result is shown in FIG.

As a learning method of a neutral network having a three-layer structure including an input layer, an intermediate layer, and an output layer, an error back propagation method is used. When the pattern of the feature amount is input to the input layer, the error between the localized output function and the teacher signal at the output layer of the neutral network is expressed by the following equations (4) and (5).
Is calculated by

(Equation 4)

(Equation 5)

Here, T _k : teacher signal O _k : output value of unit k in the output layer l: number of units in the middle layer m: number of units in the output layer γ _k : threshold value of unit k in the output layer c: This is the constant of the limited output function. The coupling coefficient that minimizes the error by learning is obtained by the following equation (6).

(Equation 6)

Using the steepest descent method, an updated value ΔV _kj of the coupling coefficient V _kj is obtained by the following equation (7).

(Equation 7) Here, α: constant H _j : output of the unit j of the intermediate layer a _k : calculated by Expression 8.

(Equation 8)

Similarly, the coupling coefficient W from the input layer to the intermediate layer is
update value [Delta] W _ji of _ji is determined by Equation 9 follows.

(Equation 9) Here, I _i : the output value of the unit i of the input layer θ _j : the threshold value of the unit j of the intermediate layer b _j : It is obtained by Expression 10. β: constant p: number of units in the input layer

(Equation 10)

The coupling coefficient between the output layer and the intermediate layer is learned according to the number of input patterns. Further, by using a category boundary setting method described later for discriminating between a normal pattern and an abnormal pattern, the erroneous determination rate for unlearned data that does not belong to a category can be significantly reduced.

[0033]

FIG. 6 shows an example of a product inspection of the engine 10 in which the reflection coefficient value, which is the above characteristic amount, is measured. FIG.
It shows the distribution of the characteristic amounts of the collected rotation sounds of the engine 10 in the same category, and the reflection coefficient γ _mk is used as the characteristic amounts. The reflection coefficient γ _mk is one of the feature amounts used in audio signal processing, and has a feature that audio frequency information can be stored in a compressed form. FIG. 6 shows the distribution of the amount for a specific order of the reflection coefficient. The horizontal axis in FIG. 6 represents a class value (feature amount). In this diagram, for example, a class value of 0.25 indicates that the section is 0.245 or more and less than 0.255. The vertical axis represents the frequency, which represents the frequency of the feature amount entering a certain section. In FIG. 6, for comparison, the frequency of the normal distribution (indicated by-■ in the figure) obtained using the average value and the standard error calculated from the feature amount is shown. From this figure, it can be seen that the distribution of the characteristic amount (− ◆ −) of the engine rotation sound follows the normal distribution.

Here, as a method of creating the boundary data,
First, an average value m of feature amounts and a standard deviation σ are calculated from learning data included in a certain category. Data obtained by changing a part or all of the learning data to m ± aσ (a is a certain real number) is defined as boundary data. The thus created boundary data is added to the learning data to perform learning of the neutral network.

As to how to determine the real value a, the knowledge of the normal distribution can be referred to. That is, if it is known that a certain quantity follows a normal distribution having an average value m1 and a standard deviation σ1, the quantity La is m1 ± 3.
The probability of entering the range of σ1 is 99.7%, which means that almost all amounts fall within this range. From this, it is possible to construct a neutral network having a desirable pattern recognition ability by setting the real value a to a value of about 5 to 7 in consideration of the fact that the output value of the neutral network changes continuously.

As described above, in the pattern matching method in which an arbitrary pattern is input to the input unit group and the output pattern of the output unit group is compared with the output pattern corresponding to the learned input pattern, By using a neural network to create a normal distribution boundary pattern around the plurality of learning input patterns belonging to, the generation of boundary data becomes easy.

[0037]

As described above, according to the present invention,
The creation of the boundary data is facilitated, and the time and effort required for the creation can be reduced. Further, since boundary data is created using specific numbers such as the average value and the standard deviation of learning data included in the category, even an operator who is not a neural network can easily create boundary data. Further, for example, by introducing it into an engine inspection process, inspection standards that have not been quantified until now become clear, and inspection efficiency and quality control can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram of a pattern matching device according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of a flow of a control signal according to the embodiment of the present invention.

FIG. 3 is a diagram illustrating an example of a reflection coefficient of an engine noise (slap noise) according to the embodiment of the present invention.

FIG. 4 is a diagram illustrating an example of a reflection coefficient of engine abnormal noise (ruffled noise) according to the embodiment of the present invention.

FIG. 5 is a diagram illustrating a result of a calculation example of a normal distribution according to the embodiment of the present invention.

FIG. 6 is a diagram illustrating an example of engine product inspection measured using a reflection coefficient value that is a feature value according to an embodiment of the present invention.

FIG. 7 is a block diagram of a conventional pattern matching device.

FIG. 8 is a diagram illustrating a boundary pattern of conventional pattern matching.

[Explanation of symbols]

10 ... engine, 12 ... sound receiving device, 14 ...
Signal analysis device, 16 Control signal input / output device, 20
…… Digital filter, 22 …… Analyzer, 24 ……
Diagnostic device 26 Data storage device 28 Judgment unit 110 Image receiving device 112 Pattern input unit 114 Conversion processing unit 116 Input layer 118
... Intermediate layer, 120 output layer, 126 pattern output unit, 128 memory, 130 display device, 132 error detection unit, 134 teacher device,
150: learning signal holding unit, 154: pattern boundary setting unit

Continued on the front page (72) Inventor Hiroyuki Tachibana 3-1-1, Tamano, Tamano-shi, Okayama Mitsui Engineering & Shipbuilding Co., Ltd. Tamano Works F-term (reference) 5B057 AA12 CH04 DA03 DC32 DC40 5L096 BA16 HA07 HA11 KA04

Claims (2)

[Claims] An input unit group is input with a plurality of learning input patterns, and an output pattern of an output unit group corresponding to the input learning input pattern is compared with a predetermined teacher pattern to determine an error between the two. On the basis of this error, the learning input pattern was learned by repeating the operation of correcting the threshold value of the unit constituting each unit group and the coupling coefficient of the coupling element connecting each unit group. Thereafter, in a pattern matching method in which an arbitrary pattern is input to the input unit group and an output pattern of the output unit group is compared with an output pattern corresponding to the learned input pattern, the plurality of learning patterns belonging to the same category A pattern matching method characterized in that a boundary pattern is created around an input pattern for use by a statistical distribution. 2. The method according to claim 1, wherein an average value and a standard deviation are calculated from the plurality of learning input patterns belonging to the same category. 2. The pattern matching method according to claim 1, wherein a value obtained by adding and subtracting the doubling value is used.