JP2006058051A - Acoustic test method and device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a means capable of determining simply and surely a proper processing condition of a sound signal differentiated by an inspection object, in an acoustic test method using a processing means of the sound signal and a neural network (NNW). <P>SOLUTION: A digitized sound signal is processed in each step for frequency analysis, coordinate axis conversion, coordinate axis division, averaging processing, and compression processing of intensity/amplitude, and then inputted into the NNW, and learning by learing data and abnormality determination of an inspection object sound are performed. A frequency axis conversion mode or a power exponent in the compression processing is changed by using a test data set with a teacher separately from the learning data, and it is determined whether the test data set is normal or not, and a coordinate axis conversion mode wherein the correct answer rate becomes highest or the optimum condition for the power exponent in the compression processing are determined, and the abnormality determination of the inspection object sound is performed in this condition. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an acoustic inspection method using a neural network, and is particularly suitable for determining the presence / absence of an abnormality of a product by a percussion sound, and an acoustic inspection method capable of increasing the accuracy of determination by a sound signal processing means. And an inspection apparatus used therefor.

  Conventionally, as an inspection method for determining the presence / absence of abnormalities in various products and devices, the acoustic spectrum for determining the presence / absence of an abnormality by obtaining the frequency spectrum of the acoustic signal to be inspected and comparing the amplitude value of each frequency component with a threshold Inspection methods are widely used. However, this method has a problem in that it is necessary to increase the number of frequency bands to be divided in order to increase the determination accuracy, and the effort for setting an appropriate threshold becomes excessive.

  In order to solve this, an acoustic inspection method for inputting the result of the frequency analysis of the acoustic signal into a neural network (hereinafter abbreviated as NNW) and determining the presence or absence of an abnormality has been widely attempted (for example, the following). Non-Patent Document 1, Patent Document 1, etc.). In this method, sound signals in each frequency band are input to the input layer of the NNW, and the output value is compared with a threshold value to determine an abnormality. The learning data is used to learn the values of the NNW parameters. By doing so, the determination accuracy can be increased.

However, even in this method, there is a skill in how to give the input and setting conditions in the process of learning the NNW parameter values using the learning data, and considerable skill is required to efficiently learn the NNW. There was a problem. In addition, depending on the type of sound signal to be inspected, there is a problem that even if a learned NNW is used, the accuracy of abnormality determination is not necessarily improved.
RPGorman et al .: Neural Networks Vol.1, Issue 1,1988 pp.75-89 JP-A-11-241945

  The present inventors have made various studies on a method for judging abnormalities (particularly, bottle defects) by inputting acoustic test percussion sounds processed by simulating human auditory characteristics into a hierarchical NNW. It was. For processing such a sound signal, the sound signal is divided into a plurality of frequency bands, and sound pressure compression processing in each band (the output sound pressure is set to the nth power of the input sound pressure (n <1)) For example, logarithmic transformation and exponential transformation). As a result, it has been found that when the sound signal is processed and input to the NNW, the hit rate of the determination may be greatly improved as compared to inputting the original sound signal as it is.

  However, there are various inspection objects, and it is impossible to uniformly discuss what kind of processing is effective in improving the hit rate. Even in the same compression process, a difference occurs in the hit rate depending on the value of the power index n, and naturally, it is greatly different depending on the conversion process mode of the frequency axis. For example, in the case of a bottle defect inspection, the frequency axis conversion processing mode and the value of the heel index n for obtaining the highest hit rate are greatly different between beer bottle and cola bottle defects. Previously, it was necessary to search for appropriate sound signal processing conditions through trial and error depending on the type of inspection object, and there was a problem that the time required for prior examination before actually starting the hammering test was extremely large. .

Moreover, it can be said that such acoustic inspection has already entered the stage of practical application, but in order to apply it to an actual production line, it takes time to select the processing conditions of the sound signal as described above and to learn by learning data. It is required that this can be done automatically, and that anyone can operate it without requiring skill in selecting such conditions.

  Therefore, the present invention provides means that can easily and reliably determine an appropriate processing condition of a sound signal that differs depending on an inspection object in a sound inspection method using a sound signal processing means and a neural network (NNW). Accordingly, an object of the present invention is to provide an acoustic inspection method and an inspection apparatus that can significantly increase the accuracy of determining whether there is an abnormality or not.

In addition, the present invention provides means for automatically determining the appropriate processing conditions for the above sound signals, thereby making it possible for any person to carry out the sound inspection without requiring special skills. Is an issue.

The acoustic inspection method of the present invention for solving the above problems is as follows.
An acoustic inspection method for inputting a sound signal subjected to frequency analysis to an input layer of a neural network learned using a learning data set, and determining whether the sound signal to be inspected is in a normal state or an abnormal state based on an output thereof,
Frequency analysis of digitized sound signal to obtain its frequency spectrum, step of converting the frequency axis of the spectrum, and dividing the converted frequency axis into the same number of frequency bands as the number N of input units of the neural network A step of averaging the intensity or amplitude of the sound signal in the divided band, respectively, and inputting the averaged sound signal to each input unit of the neural network, And determining whether or not the inspection target sound signal is normal.

  According to the above method, in an acoustic test using a neural network, the frequency axis of a sound signal is converted in accordance with human auditory characteristics and then frequency-divided and input to the neural network. It can be close to the acoustic inspection by the sensuality. Therefore, in many cases, the hit ratio of abnormality determination can be improved as compared with the case of inputting to a neural network without performing such conversion.

  In the acoustic inspection method, the step of inputting the sound signal averaged for each frequency band in the averaging step and compressing the intensity or amplitude so as to satisfy the relationship of the following equation (1): It is preferable that the signal compressed in the step is input to each input unit of the neural network.

W = C · I ⁿ (1)
Here, I: Input signal to the compression processing step W: Output signal from the compression processing step C: Proportional constant n: Power exponent set according to the processing purpose, when the input signal is strength
0 <n <0.5, 0 <n <1 when amplitude
By such compression processing, the output signal is relatively small in the range where the input signal is large, and relatively large in the range where the input signal is small. Therefore, it becomes easy to catch a small sound signal at the time of abnormality from the sound of relatively large percussion itself, and the accuracy of abnormality determination can be improved.

In the acoustic inspection method, a test data set including a plurality of normal and abnormal sound signals is used separately from the learning data set, and the frequency axis conversion mode in the frequency axis conversion step is set. A plurality of selected, in each of the selected conversion modes, learning by the learning data set and determining whether the test data set is normal,
By this determination, a conversion mode in which the correct answer rate is the highest among these conversion modes is determined, and learning based on the learning data set and determination of whether the inspection target sound signal is normal or not are performed using the determined conversion mode. Is preferred.

  In the above-described method of converting the frequency axis, which conversion mode is most effective for improving the hit rate differs depending on the sound signal to be inspected, and cannot be discussed uniformly. Conventionally, it has been necessary to determine an optimal frequency axis conversion mode by trial and error, and there has been a problem that the labor for that is excessive. On the other hand, according to the above method, the above trial can be automatically performed using a certain calculation program, and the labor of the trial is remarkably reduced. The object of the present invention can be easily achieved.

  Furthermore, in the acoustic inspection method, the frequency axis conversion mode in the frequency axis conversion step includes two types from the group consisting of frequency axis compression conversion such as logarithmic conversion, power exponent conversion, mel conversion, and bark conversion, and linear conversion. In addition to selecting the above, the power index n in the equation (1) is changed to two or more levels with numerical values including values in the range of 0 <n <0.5 when the input signal is strong and 0 <n <1 when the input signal is amplitude Then, in each combination of the conversion mode and the n value, learning by the learning data set and determination as to whether the test data set is normal or not are performed, and by this determination, the combination having the highest correct answer rate is determined. Further, using the combination of the determined conversion mode and the n value, the learning by the learning data set and the determination as to whether the inspection target sound signal is normal may be performed.

  By configuring in this way, it becomes possible to automatically determine the combination of the frequency axis conversion mode and the value of the power index n in the compression processing, which is most effective for improving the accuracy of abnormality determination. A more effective means for improving the rate can be provided.

An acoustic inspection apparatus of the present invention is an apparatus for carrying out the above inspection method,
Means for frequency analysis of a digitized sound signal to obtain a frequency spectrum; means for converting the frequency axis of the spectrum; means for dividing the converted frequency axis into a plurality of frequency bands; A means for averaging the signal strength or amplitude in the band, a means for compressing the averaged sound signal so as to satisfy the relationship of the above equation (1), and the compressed signal are input. An acoustic inspection apparatus comprising a neural network that determines whether learning and inspection target sound signals are normal.

  According to the present invention, in an acoustic inspection method using a sound signal processing means and a neural network (NNW), it is possible to easily and surely determine an appropriate processing condition of a sound signal that differs depending on an inspection object, thereby causing abnormalities. The accuracy of the presence / absence determination can be greatly increased compared to the conventional method.

  Further, according to the present invention, the acoustic inspection can be performed by any number of people without requiring special skills, and can contribute to the spread of the inspection.

  In the acoustic inspection method of the present invention, input sound signals are subjected to certain processing and then determined by the learned NNW. However, the test data with the teacher signal (normal or known) is used to determine the sound. After optimizing the signal processing conditions, the presence or absence of abnormality is determined for the inspection target data. Therefore, as shown in the block diagram of FIG. 1, the operation procedure is divided into two processes of optimization processing and inspection.

  In both the optimization process and the inspection process, the content of the sound signal processing is obtained by analyzing the frequency of the digitized sound signal to obtain a frequency spectrum (S1, T1), and specifying the frequency axis of this frequency spectrum. Conversion is performed in the frequency axis conversion mode (S2, T2). In the converted frequency axis, the frequency band is divided into the same number as the number of NNW neurons (N) (in principle, equal division), and a signal is transmitted in each band. This is the same in that the intensity or amplitude is averaged (S3, T3), and the averaged signal is subjected to an amplitude / intensity compression process (S4, T4) so as to satisfy the relationship of equation (1). Similarly, the processed signal is input to each neuron of the NNW and the determination is made based on the output value.

The content of the optimization process is to select the optimum value of the frequency axis conversion mode in step S2 and the power exponent (n in equation (1)) of the compression process in step S4 according to the inspection object. That is, a candidate frequency axis conversion mode (hereinafter referred to as M _k (k = 1, 2,..., K)) and a power exponent value (hereinafter referred to as n _j (j = 1, 2,..., J) are selected in advance, and a parameter that maximizes the correct answer rate when a digital test data set with teacher signal is determined by NNW It is to obtain (M _k and n _j ). However, instead of optimizing both of them independently, an optimum condition is obtained for the combination of both (M _k , n _j ).

  Before the NNW determines whether the test data set is normal or abnormal, it is necessary to optimize the weight coefficient of the NNW. For this reason, a set of digitized learning data with a teacher signal is prepared separately from the test data set. The sound signal of this learning data set is also processed in steps S1 to S4, and then input to each neuron of the NNW to learn and optimize the weight coefficient. If a signal after processing the test data set is input to the NNW learned in this way, a highly accurate determination can be made.

As described above, according to the inspection target, the optimal (M _{k, n j)} If the combination of Sadamare, n _j is the M _k and T4 steps in T2 step, fixed to these values. Further, since the NNW optimum weighting coefficient in the T5 step has already been obtained in the previous S5 step, it can be used as it is. Therefore, after entering the inspection process in this state, sampling the inspection target sound and performing A / D conversion (T0), the T1-T4 processed signal is input to the learned NNW ( T5) Whether normal or abnormal can be determined.

  Hereinafter, the flow of each process of optimization processing and inspection will be described in more detail. FIG. 2 is a flowchart illustrating an example of an operation procedure of the optimization process. In the figure, the input data set 1 is composed of a set of learning data with a teacher signal and a set of test data with a teacher signal. A neural network (NNW) is trained using a training data set with a teacher signal, and a learning effect and suitability (correct answer rate) of parameter selection are determined using a test data set with a teacher signal.

When the input sound signal is frequency-analyzed in step S1, a direct current component, a fundamental wave (f ₁ ) component, and a harmonic (2f ₁ , 3f ₁ ,... Kf ₁ ) component are calculated. The frequency band and the frequency f of the fundamental wave are set so that the number of components (k + 1) is sufficiently larger than the number N of input layer neurons of the NNW.

  Next, examples of the frequency axis transformation of S2 include linear transformation, logarithmic transformation, power transformation, mel transformation, bark transformation, ERB transformation, ERBs transformation, and the like. These conversions are expressed by the following equations, where f is the original frequency and F is the frequency after conversion.

Linear transformation: F = af (a is a proportional constant, but is usually 1 and so on)
Logarithmic transformation: F = allogf or F = alnf
Power conversion: F = af ⁿ (n <1)
Mel conversion: F = 2595 log ₁₀ (1 + f / 700)
bark transform: F = 13 tan ⁻¹ (0.76 f / 1000)
+3.5 tan ⁻¹ ((f / 7500) ² )
ERB conversion: F = 24.7 + 0.108f
ERBs conversion: F = 21.4 log ₁₀ (0.00437f + 1)
Of the above transformations, linear transformation, logarithmic transformation, power conversion when power n is 0.25 and 0.5, original frequency f and frequency after transformation in mel transformation and bark transformation The relationship with F is shown in FIG. The vertical axis in the figure is the relative value of F (f) (F (f) / F (20000) × 100) with respect to F value F (20000) at 100 when f = 20,000 Hz. is doing. Of the above transformations, except linear transformation and ERB transformation, transformation that relatively widens the low frequency bandwidth and relatively narrows the high frequency bandwidth (in the present invention, this is frequency axis compression). Conversion). Human hearing is a kind of frequency axis compression transformation, and it is known that the above-mentioned Mel transformation and bark transformation approximate human hearing characteristics well. Therefore, in the sense of simulating a percussion examination based on human auditory perception, it is highly likely that such frequency axis compression conversion will contribute to an improvement in the hit rate.

  Next, the division process in step S3 is to equally divide the converted frequency axis according to the number N of neurons. In addition, the averaging is to obtain an average value of the intensities of the harmonics in each divided band (a value obtained by integrating the above intensity using the converted frequency F as a variable and dividing by the bandwidth ΔF). Similar learning and test results can be obtained using the average value of the amplitude.

The amplitude / intensity compression process in step S4 is to obtain an output signal value W satisfying the relationship of the following equation, using the average value of S3 as the input signal I.
W = C · I ⁿ (1)
Where C: proportional constant n: power exponent set according to the processing purpose, when the input signal is strength
0 <n <0.5, 0 <n <1 when amplitude
Generally, when the power index n is less than 1, the output value W is relatively small in the range where the value of the input I is large (large sound), and the range where the value of I is small (small sound) is relatively large. Further, when n exceeds 1, the opposite is true, so that n = 1 can be divided into sound pressure compression and sound pressure expansion. In general, sound pressure compression is effective in improving the hit rate in the sense of reliably catching a small sound signal at the time of abnormality from the sound of a relatively large percussion itself. I in the above formula (1) is the intensity or amplitude of the spectrum, and the intensity corresponds to the square of the amplitude. Therefore, in the compression processing according to the present invention, compression processing is performed such that 0 <n <0.5 when the input signal is strong and 0 <n <1 when the input signal is amplitude.

In the NNW of step S5, learning is first performed using a set of learning data with a teacher signal, and the weight of connection between neurons is adjusted (the contents of weight adjustment will be described later). Next, it is determined whether each test data is normal or abnormal from the output value of the NNW using the adjustment result and a set of test data with a teacher signal. Since the test data set is known to be normal or abnormal, the correct answer rate is obtained from this determination. This correct answer rate is obtained when the frequency axis conversion mode is M _k and the power exponent value of the compression process is n _j . Therefore, a set of M _k , n _j , correct answer rate, and NNW weight coefficient is recorded in the memory as output 1 of step S5.

Then, a flag j of _{n j} as j-1, the flow returns to the entrance of Step S4. After the learning data set is compressed and the NNW is learned using the data, the test data set is compressed and input to the NNW to obtain the correct answer rate, and the procedure of recording the output 1 in the memory is repeated. Upon reaching the J = 0, together with the return to j = j, the flag k of _{M k} as k-1, back to the entrance of step S2, until k = 0, repeating the same procedure as described above. When both flags j and k become 0, the correct answer rate and the set of NNW weighting factors for all combinations of (M _k , n _j ) are recorded in the memory. A combination of M _k and n _{j to be} given is selected, a set of M _k and n _j at that time and a set of NNW weighting coefficients are output as output 2, and the optimization processing step is completed.

FIG. 4 is a flowchart illustrating an example of an operation procedure of the inspection process. First, when the optimization process output 2 (the optimum value of the combination of (M _k , n _j ) and the NNW weighting coefficient set at that time) is input, a trigger signal for starting the test is output as output 3 (S0 ). As a result, the consultation test is started, the signal input of the consultation test is amplified (P1), the high frequency region exceeding the audible range is cut by the filter (P2), A / D conversion is performed (T0), and this digitization is performed. The acquired signal is captured as inspection data.

For this inspection data, frequency analysis is performed in the same manner as in the above-described S1 step (T1), frequency axis conversion is performed in the frequency axis conversion mode _Mk selected as the optimum condition (T2), and the frequency axis is divided and each frequency band is The signal strength or amplitude is averaged (T3), and the signal of the average value is input, and compression processing is performed for each frequency band to set the power exponent of equation (1) to the selected n _j (T4). Then, the output is input to each unit of the input layer of the NNW that has already been learned (weight coefficient is optimized), and it is determined whether the inspection data is normal or abnormal based on the output value (T5). This normal / abnormal judgment result is output as output 4. The flag m of the number of inspections is set to m−1, the process returns to the entrance of the S0 step, and a series of operations of taking in inspection data, processing data, and determining by NNW is repeated.

  According to the above-mentioned means, learning of NNW and optimization of sound signal processing conditions can be performed automatically, and “anyone can do this without requiring special skill in acoustic inspection. It is possible to easily achieve the object of the present invention. In both the optimization process and the inspection process, both the amplitude and strength compression processes (steps S4 and T4) may be omitted. In this case, the output signals of S3 and T3 are immediately input to the NNW. It is also possible to insert the steps S4 and T4 between the frequency division process and the averaging process of steps S3 and T3. That is, the compression processing according to the equation (1) may be performed on each harmonic before averaging, and the result may be averaged. However, since this method requires a lot of time and effort to calculate equation (1), it is preferable to place steps S4 and T4 after steps S3 and T3 as shown in the examples of FIGS.

  Next, the configuration of the neural network (NNW) used in the examples described later and learning using learning data will be described. In this example, a hierarchical NNW that is effective for pattern recognition was used, and a hierarchical NNW composed of a total of three layers, an input layer, an intermediate layer, and an output layer, was used. Since there are two outputs for determining whether the inspection object is normal or abnormal, the number of neurons in the output layer is set to one. This NNW is shown in FIG. As shown in the figure, neurons are arranged in layers and are connected with different weights. The role of the neuron is to receive inputs from other neurons and output them according to the sum of their inputs. However, because the weighting changes, different outputs are produced even if the same input data is given.

The NNW learning is to change the weighting so that a target output can be obtained. Back propagation (error back propagation method) was used for learning. In this learning method, first, the number of neurons is I and that of the intermediate layer is J, and the weight from the i-th neuron of the input layer to the j-th neuron of the intermediate layer is V _ji and the j-th neuron of the intermediate layer is used. Assuming that the weight of the output layer neuron is W _j , the output X _i of the input layer neuron transmits the value to the next layer as it is, and the output Y _j of the intermediate layer neuron _j is given by the following equation (2):
The output Z of the output layer neuron is given by the following equation (3).
When the teacher signal is D, the error E between the output of the output layer is given by ^{E = (D-Z) 2} /2. In the NNW learning, the weight is corrected so that the error E becomes small. The current value of the weight from the intermediate layer to the output layer is W _j (t), and the value after the weight correction is W _{j. Assuming} (t + 1), ΔW _j when W _j (t + 1) = W _j (t) + ΔW _j is given by the following equation (4).
Here, η (η> 0) is a learning coefficient, and the degree of learning can be changed by adjusting this value.

On the other hand, assuming that the current value of the weight from the input layer to the intermediate layer is V _ji (t) and the value after the weight correction is V _ji (t + 1), V _ji (t + 1) = V _ji (t) + ΔV _ji ΔV _{ji at the} time is given by the following equation (5).
In this way, the correction of W _j and V _ji is repeated. If there are N sets of learning data, this weight correction is continued until the following expression (6) is satisfied.
In the above expression, ε (ε> 0) is a learning end condition.

The backpropagation algorithm changes the weighting in a direction that reduces the error between the output and the teacher signal. The correction of the weighting is in the direction opposite to the flow of data when calculating the output from the input. Done. This NNW learning can be automatically performed by a predetermined calculation program by previously setting η and ε empirically in advance. The “learned NNW” here is a value obtained by using the values of W _j and V _ji when the relationship of the expression (6) is satisfied as a weighting coefficient of the NNW.

  Using two types of cola bottles and beer bottles as samples, processing the sound signal and making a determination with NNW according to the flow shown in FIG. 3, changing the frequency axis conversion mode and the amplitude / sound pressure compression index, the correct answer rate We investigated how it changed. As supervised data, both cola bottles and beer bottles are prepared for 400 normal bottles and defective bottles, and the sound signals are digitized. Half of the normal bottles and defective bottles are used as a learning data set. A test data set was used. The number of NNW neurons is the input layer 100, the intermediate layer 5, and the output layer 1.

As frequency axis conversion modes, linear, logarithmic, mel, bark, ERB, and ERBs conversions were examined. Also, the input of the amplitude / sound pressure compression process (step S4) is the amplitude (strength of the power of 1/2), and for the cola bottle, the cocoon index n = 0.4 (with compression process) and n = 1 (compression process) In the case of beer bottles, only the case where the coffee index n = 0.4 (with compression treatment) was examined. Regarding the cola bottle, both the case where the frequency band is the entire audible range (20 Hz to 20 kHz) and the case where the frequency band is limited to a range where a relatively strong spectrum is recognized (1 kHz to 15 kHz) were examined. Table 1 summarizes the results of determining whether the test data set is normal or not by the NNW learned with the learning setter set under the above conditions, and obtaining the correct answer rate.
From the results for the cola bottles in Table 1, it is known that the correct answer rate is significantly higher when n = 0.4 than when n = 1 (amplitude as it is). In addition, when n = 0.4 and the frequency axis conversion mode is linear and ERB (similar to linear), the correct answer rate is high. Therefore, in the case of a cola bottle, it is known that it is preferable to perform an inspection in which the frequency axis subjected to frequency analysis is equally divided and averaged without being converted and the amplitude raised to the power of 0.4 is input to the NNW. In addition, the frequency range was not limited, and the correct answer rate was higher when the analysis was made over the entire audible range.

  On the other hand, in the case of beer bottles, when the frequency axis conversion is mel or bark, the correct answer rate is clearly higher than linear (no frequency axis conversion). Therefore, in this case, it is preferable that the frequency axis is subjected to mel conversion or bark conversion, and the amplitude multiplied by 0.4 is input to the NNW.

  As can be seen from this example, the appropriate conditions of sound signal processing (frequency axis conversion mode and compression processing power index) for increasing the hitting test accuracy are greatly different depending on the inspection object. However, according to the present invention, it is possible to automatically find out the optimum processing conditions (without requiring special skill) according to the object to be inspected, and to perform the inspection. It can certainly be increased.

It is a block diagram which shows the example of the operation procedure of the acoustic test | inspection method of this invention. It is a flowchart which shows the example of the operation procedure of the optimization process in this invention. It is a figure which shows the relationship between the original frequency and the frequency after conversion in various frequency axis conversion. It is a flowchart which shows the example of the operation procedure of the test | inspection process in this invention. It is a figure which shows the structure of the neural network used in the Example of this invention.

Claims (5)

An acoustic inspection method for inputting a sound signal subjected to frequency analysis to an input layer of a neural network learned using a learning data set, and determining whether the sound signal to be inspected is in a normal state or an abnormal state based on an output thereof,
A frequency analysis of the digitized sound signal to obtain a frequency spectrum thereof, a step of converting the frequency axis of the spectrum, and the converted frequency axis into the same frequency band as the number N of input units of the neural network. A step of dividing, a step of averaging the intensities or amplitudes of the sound signals in the divided bands, respectively, and inputting the averaged sound signal to each input unit of the neural network, according to the learning data set And a step of determining whether or not the sound signal to be inspected is normal. A step of inputting a sound signal averaged for each frequency band in the averaging step and compressing the intensity or amplitude so as to satisfy the relationship of the following equation (1) is provided. 2. The acoustic inspection method according to claim 1, wherein the received signal is input to each input unit of the neural network.
W = C · I ⁿ (1)
Here, I: Input signal to the compression processing step W: Output signal from the compression processing step C: Proportional constant n: Power exponent set according to the processing purpose, when the input signal is strength
0 <n <0.5, 0 <n <1 when amplitude   Separately from the learning data set, using a test data set consisting of a plurality of normal and abnormal sound signals, selecting a plurality of frequency axis conversion modes in the frequency axis conversion step, and selecting each conversion It is determined whether the learning and the test data set are normal in the mode, and by this determination, a conversion mode in which the correct answer rate is the highest among these conversion modes is determined, and using the determined conversion mode, 3. The acoustic inspection method according to claim 1, wherein the learning and the inspection target sound signal are determined to be normal.   As the frequency axis conversion mode in the frequency axis conversion step, two or more types are selected from the group consisting of frequency axis compression conversion such as logarithmic conversion, power exponent conversion, mel conversion, and bark conversion, and linear conversion, and (1) By changing the power index n in the equation into two or more values including values in the range of 0 <n <0.5 when the input signal is strength and 0 <n <1 when the amplitude is amplitude, In each combination, it is determined whether or not the learning and the test data set are normal, and by this determination, the combination having the highest correct answer rate is determined, and the combination of the predetermined conversion mode and n value is determined. The acoustic inspection method according to claim 3, wherein the determination is performed to determine whether the learning and inspection target sound signal is normal.   5. An apparatus for carrying out the acoustic inspection method according to claim 1, wherein means for obtaining a frequency spectrum by frequency-analyzing a digitized sound signal and converting the frequency axis of the spectrum. Means, means for dividing the transformed frequency axis into a plurality of frequency bands, means for averaging the signal intensity or amplitude in the divided bands, and the averaged sound signal (1 And a neural network for inputting the compressed sound signal and determining whether the sound signal to be examined and the sound signal to be inspected are normal. Acoustic inspection device.