JP2005141591A - Quality evaluation device and product evaluation system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To make it possible to judge similarity even when a distribution pattern of a frequency spectrum is deviated to a frequency axis direction. <P>SOLUTION: A learning vector quantization means 22 finds out deflection quantity obtained when a frequency spectrum already learned by a learning sample and included in a sound wave from an object is deviated to the frequency axis direction so that the degree of coincidence with the frequency spectrum of the learning sample is increased and deviates the frequency spectrum obtained from a feature sound extraction device 1 to the frequency axis direction only by the found deviation quantity. The deviated frequency spectrum is inputted to a sorting neural network 23 to sort the distribution pattern of the frequency spectrum obtained as an output from the learning vector quantization means 22 and evaluate the quality of the object. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a quality evaluation apparatus and a product evaluation system.

Conventionally, a technique for recognizing characters and speech by using a neural network as a classifier is known (see, for example, Patent Document 1).
Japanese Patent Publication No.7-113943

  The above-described patent document describes a technology that enables character recognition by learning of a neural network, and this type of technology can also be used for matching frequency spectrum distribution patterns in speech recognition. However, for example, when the quality of a device is determined based on operation sound in an electric device or a mechanical device that is operated by a motor, the distribution pattern of the frequency spectrum is shifted in the frequency axis direction as the rotational speed of the motor changes. In the case of such a shift in the frequency axis direction, even if the technique described in the above-mentioned patent document is simply diverted, the similarity of the frequency spectrum distribution pattern There is a problem that it is difficult to judge.

  The present invention has been made in view of the above reasons, and an object thereof is to provide a quality evaluation apparatus that can determine similarity even if the distribution pattern of the frequency spectrum is shifted in the frequency axis direction. In addition, an object of the present invention is to provide a product evaluation system for evaluating a product using this quality evaluation apparatus.

  The invention according to claim 1 is a quality evaluation device, wherein the frequency spectrum contained in the sound wave from the object is shifted in the frequency axis direction so that the degree of coincidence with the frequency spectrum of the learning sample is high. A learning vector quantization unit that obtains a shift amount, shifts the frequency spectrum obtained from the object in the frequency axis direction by the shift amount, and a frequency spectrum obtained as an output of the learning vector quantization unit. And a classification neural network that evaluates the quality of the object by classifying the distribution pattern.

  According to this configuration, even if the distribution pattern of the frequency spectrum included in the sound wave of the target object is shifted in the frequency axis direction with respect to the distribution pattern of the frequency spectrum of the learning sample, the shift amount is obtained and superimposed. Therefore, the similarity of the distribution pattern can be easily evaluated. As a result, for example, when the presence or absence of a motor abnormality or the type of abnormality is determined based on the operation sound of the motor, the motor rotation speed is different from the rotation speed when the learning sample is obtained, Even if the distribution pattern of the frequency spectrum is shifted in the frequency axis direction, it is possible to evaluate whether or not it is similar to the learning sample. Furthermore, since the frequency spectrum distribution pattern obtained from the object is classified by the neural network for classification, it is possible to determine the presence or absence of abnormality and the type of abnormality accurately in a short time, and regardless of human sensuality. Since the type of the characteristic sound can be discriminated from each other, the variation in results is reduced.

  According to a second aspect of the present invention, in the first aspect of the present invention, a pre-quantization unit that narrows down the characteristics of the frequency spectrum obtained from the object more coarsely than the learning vector quantization unit is provided before the learning vector quantization unit. It is characterized by having been added.

  According to this configuration, candidates having high similarity to the characteristics of the frequency spectrum obtained from the object in the frequency spectrum of the learning vector can be narrowed down by the quantization unit. The load can be reduced, and as a result, the quality can be evaluated in a short time.

In the invention of claim 3, in the invention of claim 1 or claim 2, the learning vector quantization means divides the frequency spectrum obtained from the object and the frequency spectrum of the learning sample by 1/2 ^{n in the} frequency axis direction. (N is a natural number) The candidate of the deviation amount of both frequency spectrum is calculated | required in the state compressed, and both in the predetermined range on the basis of the said candidate in the state which compressed both frequency spectrum to 1/2 ^n-1. The process of obtaining the frequency spectrum deviation amount candidates is repeated, and finally the deviation amount candidates obtained without compressing both frequency spectra are used as the deviation amount.

  According to this configuration, the range of the deviation amount is narrowed down while the data for obtaining the deviation amount is compressed in the frequency axis direction, and the existence range of the deviation amount is narrowed while gradually reducing the compression rate. When the deviation amount is obtained without compression, the existence range of the deviation amount is narrowed down to a small range, and the deviation amount can be obtained in a short time.

  In the invention of claim 4, in the invention of claim 1 or claim 2, the learning vector quantization means calculates the center of gravity of the region having a peak in the frequency spectrum obtained from the object and the frequency spectrum of the learning sample, respectively. The minimum value of the distance between the centroids obtained from the frequency spectrum obtained from the object and the frequency spectrum of the learning sample is used as the shift amount.

  According to this configuration, since the minimum value among the distances of the centroid positions of the frequency spectrum is used as the shift amount, the shift amount can be obtained in a short time.

  The invention of claim 5 is a product evaluation system, wherein the sound wave from the object is an operation sound of the product, and the quality evaluation apparatus according to any one of claims 1 to 4, and the quality evaluation A feature sound extraction device that is provided in a front stage of the device and extracts the frequency spectrum of the operation sound, and the learning vector quantization means determines whether the product is normal or abnormal as the quality of the object, and the abnormality at the time of abnormality The type is distinguished.

  According to this configuration, by using the frequency spectrum of the operation sound of the product, it is possible to determine whether the product is normal or abnormal and the type of abnormality at the time of abnormality.

  According to a sixth aspect of the present invention, in the fifth aspect of the present invention, the characteristic sound extraction device receives a sound input unit that detects a sound wave from the object and a signal obtained by the sound input unit as a first input signal. Wavelet transform means for separating a low-frequency component and a high-frequency component by performing wavelet transform on the signal, and when the target characteristic sound is included in the low-frequency component separated by the wavelet transform means, the low-frequency component As a next input signal to the wavelet transform means, and a monitoring means for extracting the frequency band of the input signal of the wavelet transform means when the characteristic sound is no longer included, and the frequency band extracted by the monitoring means as the passing frequency band Fill in which the input signal to the wavelet transform means when it is set and no longer includes the characteristic sound is input Characterized in that it comprises means, and vector extracting means for detecting a frequency spectrum of the signal passed through the filter means.

  According to this configuration, the process of extracting the low frequency component from the input signal to the wavelet transform unit is repeated, and the input signal to the wavelet transform unit at the time when the characteristic sound is no longer included in the low frequency component is extracted. Since only the frequency band of the input signal is extracted by the filter means, it becomes possible to extract the frequency component related to the characteristic sound. Furthermore, since the frequency spectrum is obtained for the frequency band extracted by the filter means, the type of the characteristic sound can be easily determined using the frequency spectrum from which unnecessary components not related to the characteristic sound are removed.

  According to the present invention, even if the distribution pattern of the frequency spectrum included in the sound wave of the object is shifted in the frequency axis direction with respect to the distribution pattern of the frequency spectrum of the learning sample, the shift amount is obtained and superimposed. Therefore, there is an advantage that the similarity of the distribution pattern can be easily evaluated. Furthermore, since the frequency spectrum distribution pattern obtained from the object is classified by the neural network for classification, it is possible to determine the presence or absence of abnormality and the type of abnormality accurately in a short time, and regardless of human sensuality. Since the type of characteristic sound can be discriminated, there is an advantage that variation in results is reduced.

  In the embodiment described below, an example in which the technology of the present invention is applied to a product evaluation system for evaluating the quality of a motor will be described. As shown in FIG. 2, the product evaluation system includes a feature sound extraction device 1 that extracts feature sounds related to motor quality from operation sounds generated when the motor is operated, and a feature sound extracted by the feature sound extraction device 1. And a quality evaluation device 2 that evaluates the quality of the motor based on the above.

  The feature sound extraction apparatus 1 includes a voice input unit 11 for inputting motor operation sounds. A voice signal input from the voice input unit 11 is converted into a digital signal by an A / D conversion unit 12 and is not shown in the drawing. Is stored in the storage unit. As the voice input means 11, a vibration pickup used in contact with a motor can be used in addition to a microphone.

The original signal that is the output of the A / D conversion unit 12 is input to the wavelet conversion unit 13, and the wavelet conversion unit 13 repeats further performing wavelet conversion on the low-frequency component obtained by performing wavelet conversion. This type of processing is known as multiresolution analysis. That is, the wavelet transform unit 13 separates the signal before conversion into a low frequency component f _j−1 and a high frequency component w _j−1, and uses the obtained low frequency component f _j−1 as the next input. The conversion process of separating the low frequency component f _j-2 and the high frequency component w _j-2 is repeated. That is, the original signal becomes the input f ₀ of the first conversion process, and the input f ₀ is separated into a high frequency component f ₋₁ and a low frequency component w ₋₁ . The low frequency component f _j obtained in this processing process can be expressed as shown in Equation 1.

In Equation 1, Φ (x) is a scaling function, and ψ (x) is a wavelet function. Each time the wavelet transform unit 13 performs the conversion process of Formula 1 once, the high frequency component w _j-1 is removed and only the low frequency component f _j-1 is extracted. Therefore, when the conversion process of Formula 1 is repeated a plurality of times It is considered that a state in which the characteristic sound is not included in the low-frequency component f _j−1 after the conversion occurs with respect to the input before the conversion including the characteristic sound. The low frequency component f _j before conversion at this time is
Of the low frequency components f _i (i = 0, 1, 2,...) Obtained by the conversion processing in the weblet conversion means 13, the ratio of the frequency components of the target characteristic sound to the total frequency components Can be estimated to be the highest.

Therefore, monitoring means 14 for monitoring the low frequency component f _j for each conversion process in the wavelet transform means 13 is provided, and monitoring means for checking whether or not a characteristic sound is included in the low frequency component f _j output from the wavelet transform means 13. 14 to determine. The monitoring means 14 includes, for example, a filter bank (a filter having a plurality of bandpass filters having different frequencies and dividing an input signal into components for each frequency band) or a fast Fourier transformer, and a learning vector quantization means 22 described later. A neural network having a similar determination function can be used. In other words, the neural network is made to learn various characteristic sounds and determine the presence or absence of characteristic sounds. However, in this neural network, it is only necessary to determine the presence or absence of a characteristic sound, and the type of characteristic sound is not questioned. Further, when the low frequency component f _j−1 output from the wavelet transform unit 13 does not include the characteristic sound in the monitoring unit 14, the frequency band of the low frequency component f _j input to the wavelet transform unit 13 is changed. It also has a function of extracting by FFT or the like. In addition, as for the part of the judgment whether or not the characteristic sound is included in the monitoring unit 14, a person may recognize the characteristic sound. In this case, although there is human intervention, since it is only an operation to recognize whether there is a characteristic sound and to notify the wavelet transform means 13 when the characteristic sound is not heard, there is almost no variation in judgment due to human intervention. .

As described above, by using a wavelet transform unit 13, it is possible to proportion of the frequency component of the feature sounds of interest with respect to all frequency components to extract the highest low frequency component f _i. The feature sound extraction device 1, since the aim is to extract features sounds relating to the quality of the motor, the use of low-frequency components f _i described above as the characteristic sound used for evaluation of quality, unnecessary frequency component evaluation It is considered that the accuracy of evaluation can be improved because most of the information has been removed. Therefore, by passing a low frequency component f _i obtained by the wavelet transform unit 13 to the controllable filter means 15 to pass the frequency band by the monitoring means 14, the low-frequency component ratio is high occupied frequency component of the feature tone f _i The signal of the frequency component in the vicinity of is extracted. Here, since through the low-frequency component f _i signal further filter means 15 obtained by the wavelet transformation unit 13, it is possible to remove the unnecessary component included in the output of the wavelet transform unit 13. The signal extracted by the filter unit 15 is input to the vector extraction unit 16, and the vector extraction unit 16 obtains a frequency spectrum using a technique such as FFT (Fast Fourier Transform). The frequency spectrum obtained by the vector extraction means 16 in this way becomes the output of the feature sound extraction device 1.

  The operation of the feature sound extraction device 1 will be described more specifically. It can be considered that there is some defect when the operation sound of the motor includes, for example, an “on / on” sound that is not included during normal operation. In the present embodiment, the characteristic sound extraction apparatus 1 extracts a specific frequency band including a sound “ON / ON” from the original signal input from the voice input unit 11, and pays attention only to this relatively narrow frequency band to “ON / ON”. The sound is judged from the sound. Since the original signal input from the voice input means 11 is stored in the working storage unit as described above, the waveform of the original signal in a section of 4 to 4.5 seconds, for example, as shown in FIG. Can be read out. FIG. 2B shows a waveform read out in the section of 4.2 to 4.4 seconds of the same original signal.

  When the original signal of FIGS. 2A and 2B is input to the wavelet transform unit 13 and the conversion process is performed an appropriate number of times, the characteristic sound “on-on” that existed before the conversion is converted into a low-frequency component after the conversion. It is considered that it will not be included. The monitoring means 14 described above determines the presence or absence of the target characteristic sound by monitoring the low-frequency component output from the wavelet transform means 13, and when the characteristic sound disappears from the low-frequency component, the low-frequency component before conversion The pass band of the filter means 15 is instructed to pass. The signal before conversion at this point includes the frequency component of the characteristic sound “ON / ON”, and has a waveform as shown in FIGS. In the signal shown in FIG. 3, since the frequency component unrelated to the characteristic sound “ON / ON” is almost removed, when the frequency spectrum is extracted for this signal, the characteristic sound “ON / ON” can be quantified. Predictable.

  Here, an example of the frequency spectrum of the original signal including the characteristic sound “ON / ON” is as shown in FIG. 4A, and an example of the frequency spectrum of the original signal not including the characteristic sound “ON / ON” is illustrated. As shown in FIG. Both show the frequency spectrum in the 30 to 1500 Hz band, and comparing the two shows that there is a difference in the frequency spectrum, but there does not appear to be a significant difference. Even if these waveforms are compared, It is difficult to clearly recognize the differences.

  On the other hand, FIG. 4B gives the original signal shown in FIG. 4A to the wavelet transforming means 13, and the low frequency component output from the wavelet transforming means 13 is passed through a frequency band of 80 to 160 Hz (FIG. 4 ( FIG. 5B shows an example of the frequency spectrum of the signal obtained through the filter means 15 (frequency band near the circle shown in FIG. 5A), and FIG. 5B shows the same processing as the original signal shown in FIG. (The pass frequency band of the filter means 15 is a frequency band in the vicinity of the circle shown in FIG. 5A). As can be seen by comparing FIG. 4B and FIG. 5B, there is a difference that can be seen at a glance in the frequency spectrum. That is, it is considered that most of the characteristic sound “ON / ON” is included in the frequency band of 80 to 160 Hz. Therefore, by comparing the two using the output of the filter means 15, it is possible to easily identify whether the signal includes the characteristic sound “ON / ON” or does not include the characteristic sound “ON / ON”. I understand.

  FIG. 6 shows an example of a signal waveform obtained by passing the signal containing the “ON / ON” characteristic sound output from the wavelet transform unit 13 through the filter unit 15 having a pass frequency band of 80 to 160 Hz. According to the figure, there is a change in amplitude in about 3 cycles per second, and it can be seen that this change in amplitude corresponds to the characteristic sound “ON / ON”. Here, it is considered that the characteristic sound “ON / ON” from the frequency component shown in FIG. 4 is caused by a defective brush of the motor, and when it is replaced with a non-defective brush, the characteristic sound “ON / ON” hardly occurs. That is, it can be inferred that the above-mentioned characteristic sound “ON / ON” is caused by a defective brush of the motor.

  Based on the knowledge described above, if the pattern of the frequency spectrum output from the feature sound extraction device 1 is classified, it can be determined whether or not the motor is normal, and the content of the abnormality when it is determined that the motor is not normal It can be seen that can be estimated. Therefore, in this embodiment, in order to classify the pattern of the frequency spectrum output from the feature sound extraction device 1, the quality evaluation device 2 provided with a neural network is used. However, if the rotational speed of the motor changes, even if the same kind of abnormality occurs, the frequency spectrum pattern shifts in the frequency axis direction. Judgment technology is required. Therefore, in the present embodiment, a learning vector quantization (hereinafter referred to as LVQ) method is adopted, features in the frequency spectrum pattern output from the feature sound extraction means 1 are extracted by the LVQ method, and the extracted features are converted into the extracted features. By applying the classification neural network 23, it is possible to classify the frequency spectrum pattern regardless of the frequency shift. Further, in adopting the LVQ method, in order to extract a rough feature of the frequency spectrum output from the feature sound extraction means 1, the rough quantization obtained by the prequantization means 21 is passed through the prequantization means 21. By giving the result to the learning vector quantization means 22, fine features of the frequency spectrum are extracted.

  That is, the quality extraction device 2 in this embodiment is based on the result obtained by the prequantization means 21 that roughly extracts the characteristics of the frequency spectrum output from the feature sound extraction device 1 and the prequantization means 21. Further, the learning vector quantization means 22 for extracting fine features and the classification neural network 23 for classifying the characteristics of the frequency spectrum obtained by the learning vector quantization means 22 are configured.

Both the pre-quantization means 21 and the learning vector quantization means 23 are composed of LVQ networks. However, the learning vector quantization means 23 has improved the LVQ network so that the feature of the pattern can be extracted even if there is a frequency shift. LVQ network is basically as shown in FIG. 7, the input data group _{X 1, ..., X j,} ..., an input layer Lin comprising _{X n} plurality of nodes is given (neurons) Ne, an input This is a two-layer neural network including a competitive layer Lcm having a plurality of nodes (neurons) Ne respectively coupled to each node Ne of the layer Lin. In the LVQ network, learning is used in a state in which the weighting coefficient of coupling between each node Ne of the input layer Lin and each node Ne of the competitive layer Lcm is adjusted, and the above-described input data groups X ₁ ,..., X _j,. If X _n is given, competitive layer output data group _d 1 in accordance with the weighting factor to each node Ne of Lcm, _..., d j, ..., is that _{d n} is obtained. For example, output data groups d ₁ ,..., D _{j in the} case where there is no characteristic sound “ON / ON” are obtained at the left half node Ne of the competitive layer Lcm in FIG. 7, and “ON / ON” is obtained at the right half node Ne. output data group d _{j + 1} when including the features sound _that, ..., d _n is obtained, it is possible to whether 2 classification including features sound "On'on".

The operation of the LVQ network will be described in more detail. Now, the input data group _X 1 given to each node Ne of the input layer Lin, _..., X j, ..., a distance between each node of competitive layer Lcm and _{X n} Ne (node corresponding to the output data _{d i) E i} Is defined as Equation 2. _{W ij} in Equation 2 is a weighting factor between the node Ne which corresponds to the output data _{d i} at the node Ne and the competitive layer Lcm input layer Lin providing input data _{X j.}

Input data group _X 1 of the node Ne in the competitive layer _{Lcm, ..., X j, ...} , X n that node Ne the distance _{E i} defined by the number 2 has fired the node Ne which minimizes for. In order to learn the LVQ network, the weighting factor w _ij is adjusted so as to fire a desired node Ne of the competitive layer Lcm by giving a learning sample. After performing such learning, appropriate input data group _{X 1, ..., X j,} ..., output data group _d 1 with respect to _{X n,} in accordance with the learning content patterns, ..., _{d j,} ... , can be obtained _{d n.}

By the way, as described above, since the rotational speed of the motor changes, the frequency spectrum output from the characteristic sound extraction device 1 undergoes a shift in the frequency axis direction according to the change in the rotational speed of the motor. Since the frequency spectrum output from the feature sound extraction apparatus 1 is the input data group X ₁ ,..., X _j ,..., X _n to the pre-quantization means 21 that is an LVQ network, the frequency spectrum is in the frequency axis direction. When deviation in the output data group d ₁ before置量Coca means _{21, ..., d j, ...} , d n is changed, it becomes impossible to obtain the desired result. Therefore, in the present embodiment, the distance E _i defined in the form of Equation 3 is obtained after the pre-quantization means 21 and the frequency spectrum is shifted in the frequency axis direction by the shift amount st that minimizes the distance E _i. Learning vector quantization means 22 is provided. In Equation 3, st represents the shift amount in the frequency axis direction.

Number 3 As can be seen compared to the number 2, number 3 the input data set _{X 1, ..., X j,} ..., Yes adds the shift amount st to shift in the frequency axis direction X _n, input If the frequency spectrum has the same distribution pattern by shifting the data groups X ₁ , ..., X _j , ..., X _n , the individual input data groups X ₁ , ..., X _j , ..., X _n Even if there is a deviation in the frequency axis direction, it is possible to superimpose them so as to almost coincide. The shift amount st is obtained by the learning vector quantization means 22 after learning. In other words, at the time of learning of the learning vector quantization means 22, previously to adjust the weighting coefficients w _ij is given a standard learning samples, actual input data group _{X 1, ..., X j,} ..., when giving X _n to the distance E _i a than determining the shift amount st to minimize, shift amount when the st is minimized, the output data group d ₁ that the distance E _i is _obtained, ..., d _j, ..., output data group _d 1 should seek pattern of _{d n, ..., d j,} ..., it is to regard as _{d n.}

In order to determine the deviation amount st in Equation 3 for an arbitrary input data group X ₁ ,..., X _j ,..., X _n , the deviation amount st is changed little by little within the specified range. A method for obtaining the shift amount st in which the distance E _i is minimized from the shift amount st can be considered. However, such a round robin method requires a long time until the shift amount st is obtained. Therefore, in this embodiment, the time for obtaining the shift amount st is shortened by using a genetic algorithm.

As described above, the learning vector quantization means 22 shifts the input data group X ₁ ,..., X _j ,..., X _n in the frequency axis direction and searches for the appropriate shift amount st to find the distance E _i. output data group _d 1 but which _minimizes, ..., d j, ..., from seeking _{d n,} the input data group _{X 1, ..., X j,} ..., the frequency axis corresponding to the rotation speed of the motor to _{X n} even if occurring the direction of shift, if consistent distribution pattern of the frequency spectrum is substantially the output data group _{d 1, ..., d j,} ..., distribution pattern of d _n also to substantially coincide. In other words, the input data group _{X 1, ..., X j,} ..., and the frequency spectrum given, a shift amount st as degree of coincidence between the frequency spectrum given as a learning sample is the highest determined as X _n That's right.

The output of the learning vector quantization means 22 is given to the learned classification neural network 23, and based on the frequency spectrum output from the feature sound extraction apparatus 1, it is determined whether the motor is normal or abnormal. If so, the contents of the abnormality are also determined. That is, the classification neural network 23 can determine that the brush is abnormal from the frequency spectrum corresponding to the characteristic sound “ON / ON” as in the above-described example. Moreover, the trained classification network 23, the output data group d ₁ of the learning vector quantizer _22, ..., d j, _..., since as input data set to d _n, the frequency axis due to the difference in rotational speed of the motor The influence of the direction shift is eliminated, and as a result, the content of the abnormality can be determined regardless of the rotational speed of the motor.

  In the learning of the neural network for classification 23, learning samples of good products and defective products are given. By using the classification neural network 23 thus learned in combination with the pre-quantization means 21 and the learning vector quantization means 22, stable determination can be made without causing sensory variations when determining the quality of the motor. It is possible to obtain the determined result. Further, by providing the pre-quantization means 21 in the preceding stage of the learning vector quantization means 22, the characteristics of the frequency spectrum output from the feature sound extraction device 1, the pre-quantization means 21 and the learning vector quantization Of the features of the frequency spectrum of the learning sample used for learning by the means 22, those having a relatively high similarity can be narrowed down, and the processing load on the learning vector quantization means 22 can be reduced.

By the way, in the configuration example described above, the learning vector quantization means 22 uses a genetic algorithm to optimize the shift amount st. However, the genetic algorithm is more appropriate than the brute force method. Although the time required to find st can be shortened, it still takes a relatively long time to determine an appropriate shift amount st. When it is necessary to perform manufacturing and inspection in parallel as in the manufacturing line, if it takes time to determine the shift amount st, there is a problem that the tact time of the manufacturing process becomes long. Thus, hereinafter, a technique that can shorten the time for obtaining an appropriate shift amount st as compared with the genetic algorithm will be described. In the following description referred to as a codebook vector set of learning samples used for learning in the learning already learning vector quantization means 22, an input data group _{X 1, ..., X j,} ..., is referred to as the inspection data X _n. In addition, as a technique for shortening the time for obtaining the shift amount st, a method of compressing data and a method of limiting a region for searching for the shift amount st are shown.

In order to shorten the search time for the shift amount st by compressing the data, as shown in FIG. 8, first, the codebook vector and the test data are respectively divided into 1/2 ⁿ (n is an appropriate natural number) in the frequency axis direction. ) (S1, S2). If the data is compressed to 1/2 ⁿ in this way, the shift amount st is also compressed to 1/2 ⁿ , so that the shift amount st can be obtained with a relatively small number of times (S3). However, the offset of not only those that the distance E _i to a minimum as st, obtaining the shift amount st of the distance E less a predetermined number of more of _i (j-number) (S4). Here, the shift amount st obtained by compressing the data to 1/2 ⁿ is expressed as st (n). Further, when compressing data, an average value or maximum value of a spectrum in a 2 ⁿ Hz section is used as a representative value of the section.

Next, the compression rate is reduced and the data is compressed to ½ ⁽ⁿ⁻¹⁾ , and the deviation amount st is within the range of 2 · st (n) ± 1, which is the vicinity of the deviation amount st (n). (n-1) predetermined from the smaller of the distance _{E i} number (j-number) obtaining (S5 to S9). Further, the same processing is repeated until n = 0 (S10, S11).

As described above, the range in which the deviation amount st exists gradually converges by repeating the process of obtaining the deviation amount st by reducing the compression rate in the vicinity of the deviation amount st obtained by increasing the compression rate. The deviation amount st (0) that minimizes the distance E _i can be obtained from the j deviation amounts st (0) obtained in the state where n = 0, that is, in the uncompressed state. The deviation amount st (0) is used as the deviation amount st of the frequency spectrum (S12).

When the deviation amount st is determined as described above, the time required for the search is 1 / ²²ⁿ as compared with the case where the deviation amount st is obtained by the brute force method. In other words, by compressing the data, the area for searching for the shift amount is narrowed, and j candidates for the shift amount st are obtained, thereby maintaining the reliability of the range of candidates for the shift amount st. Therefore, the shift amount st can be obtained with a small number of times with high reliability. In addition, in consideration of the fact that the peak is broad in the input frequency spectrum, a candidate for the deviation amount st is searched for in a neighboring region having a certain width. It becomes possible to ask.

  By the way, in order to shorten the search time for the shift amount st by limiting the search area for the shift amount st, first, as shown in FIG. 9, the codebook vector and the inspection data are binarized with an appropriate threshold value. (S1, S2). Also, regions extracted by binarization that are close to each other (those whose distance between peaks is within a threshold value) are collectively treated as one region. Next, the position of the center of gravity for each obtained area is obtained. For example, as shown in FIG. 10A, when there is a region exceeding the threshold Th, the centroids G1 and G2 of the region are obtained. In addition, the areas D1 and D2 on the right side of FIG. 10A are divided into two areas where the threshold value Th is exceeded. However, since the distance between the peaks of both areas D1 and D2 is short, they are treated as one area. On the other hand, a region that does not exceed the threshold Th, such as the region on the right side of FIG.

Next, as shown in FIG. 10C, a deviation amount st for matching the centroids G1 and G2 obtained from the codebook vector and the inspection data is obtained (S5), and this deviation amount st is set as an initial value. A distance E _i is obtained in the vicinity of the value shift amount st (S6). Usually, since a plurality of centroids are obtained from the code book vector and the inspection data, the same amount of the obtained centroids is combined and the same calculation is performed (S3, S4, S7 to S10), and the deviation amount st candidates. And a shift amount st that minimizes the distance E _i is employed (S11). Since this method also searches by narrowing down the range of the deviation amount st, the deviation amount st can be obtained with high reliability in a relatively short time.

  In the above-described example, an LVQ network is used for the pre-quantization means 21, but a competitive neural network without learning (so-called SOM type network) can be used as the pre-quantization means 21. It is. In any case, the pre-quantization means 21 can remove unnecessary components from the distribution pattern of the frequency spectrum obtained from the feature sound extraction device 1, and the processing load on the learning vector quantization means 22 is reduced. I can do it. By adding such a pre-quantization means 21, the learning vector quantization means 22 is compared with the case where the frequency spectrum output from the feature sound extraction device 1 is directly input to the learning vector quantization means 22. Can significantly reduce the load on the network, resulting in a reduction in processing time.

  Of the characteristic sound extraction apparatus 1 described above, the part excluding the voice input means 11 and the A / D conversion means 12 and the quality evaluation apparatus 2 are realized by executing an appropriate program on a personal computer. In addition, a dedicated device using a microprocessor is configured instead of a personal computer, and the parts such as the wavelet transform unit 13 and the learning vector quantization unit 22 that require a long processing time in software processing are configured by dedicated hardware. Thus, the processing time may be shortened.

  In the above-described embodiment, an example in which the quality of the motor is determined based on the sound generated when the motor rotates is shown. However, if the electric device or the mechanical device generates sound or vibration in addition to the device using the motor, the present embodiment can be used. The technology of the invention can be applied, and pass / fail inspection and failure analysis can be performed. In addition, when performing pass / fail inspection of objects such as elevated roads, tunnels, buildings, etc., a separate device for applying vibrations and sound waves to the objects is provided, and the sound waves and vibrations transmitted through the objects are analyzed by the device of the present invention. do it. In the above example, the feature sound extraction apparatus 1 performs wavelet transformation. However, after removing noise from the signal obtained by the voice input means 11 without performing wavelet transformation by a filter, the vector extraction means 16 obtains the frequency spectrum. You may make it ask. In this case, as compared with the case where the wavelet transform means 13 is provided, the amount of processing in the quality evaluation apparatus 2 increases and the time until the determination result is obtained becomes longer, but the quality can be evaluated.

It is a block diagram which shows embodiment. It is a figure which shows the example of the waveform of an original signal in the same as the above. It is a figure which shows the example of the waveform after wavelet transformation in the same as the above. It is a figure which shows the example of the frequency spectrum of an original signal in the same as the above. It is a figure which shows the example of the frequency spectrum of the signal which passed the filter means in the same as the above. It is a figure which shows the example of the waveform of the signal which passed the filter means in the same as the above. It is a figure which shows the structural example of the LVQ network used for the same as the above. It is operation | movement explanatory drawing which shows an example of the procedure which calculates | requires a displacement amount in the same as the above. It is operation | movement explanatory drawing which shows the other example of the procedure which calculates | requires variation | change_quantity in the same as the above. It is a figure which shows the concept in the procedure shown in FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Feature sound extraction apparatus 2 Quality evaluation apparatus 11 Voice input means 13 Wavelet transformation means 14 Monitoring means 15 Filter means 16 Vector extraction means 21 Prequantization means 22 Learning vector quantization means 23 Classification neural network

Claims (6)

  The frequency spectrum obtained from the object is obtained by obtaining the amount of deviation when the frequency spectrum included in the sound wave from the object is shifted in the frequency axis direction so that the degree of coincidence with the frequency spectrum of the learning sample is high. The learning vector quantization means that has been shifted in the direction of the frequency axis by the amount of deviation and the frequency spectrum distribution pattern obtained as the output of the learning vector quantization means are classified to evaluate the quality of the object. A quality evaluation apparatus comprising a neural network for classification.   2. The quality evaluation according to claim 1, wherein a pre-quantization unit that narrows down the characteristics of the frequency spectrum obtained from the object more roughly than the learning vector quantization unit is added before the learning vector quantization unit. apparatus. The learning vector quantization means compresses the frequency spectrum obtained from the object and the frequency spectrum of the learning sample to ½ ⁿ (n is a natural number) in the frequency axis direction, and the amount of deviation of both frequency spectra. Next, a process for obtaining candidates, and then obtaining candidates for the deviation amounts of both frequency spectra within a predetermined range based on the candidates in a state where both frequency spectra are compressed to ½ ⁿ⁻¹ is repeated. The quality evaluation apparatus according to claim 1, wherein a candidate for a deviation amount obtained without compressing a frequency spectrum is used as the deviation amount.   The learning vector quantization means obtains centroids of regions having peaks for the frequency spectrum obtained from the object and the frequency spectrum of the learning sample, respectively, and from the frequency spectrum obtained from the object and the frequency spectrum of the learning sample, respectively. The quality evaluation apparatus according to claim 1, wherein a minimum value among the obtained distances between the centers of gravity is used as the shift amount.   The sound wave from the object is an operation sound of a product, and the quality evaluation apparatus according to any one of claims 1 to 4 and a frequency spectrum of the operation sound provided in a front stage of the quality evaluation apparatus are extracted. A product evaluation system, wherein the learning vector quantization means discriminates whether the product is normal or abnormal and the type of abnormality at the time of abnormality as the quality of the object.   The feature sound extraction device includes a sound input unit that detects sound waves from the object, and a low-frequency component and a high-frequency component by performing wavelet transform on the input signal using the signal obtained by the sound input unit as a first input signal. And when the target characteristic sound is included in the low frequency component separated by the wavelet conversion means, the low frequency component is given to the wavelet conversion means as the next input signal and the characteristic sound A monitoring means for extracting the frequency band of the input signal of the wavelet transform means at the time when the sound is no longer included, and the wavelet at the time when the frequency band extracted by the monitoring means is set as a passing frequency band and the characteristic sound is no longer included Filter means to which the input signal to the conversion means is input, and the frequency of the signal that has passed through the filter means Product Evaluation system according to claim 5, characterized in that it comprises a vector extraction means for detecting the spectrum.

Images