TWI451070B - Abnormal sound diagnostic equipment - Google Patents

Description

Abnormal sound diagnostic device

The present invention relates to the possibility of determining the occurrence of abnormal sounds of a machine in operation based on a time-frequency analysis of a signal collected by a microphone (hereinafter referred to simply as a microphone) or a vibration sensor. Device. In particular, an abnormal sound diagnosis device diagnoses various abnormal sounds generated during operation of a system composed of a plurality of devices.

In the first technique of the abnormal sound diagnosis device for the conventional diagnosis abnormal sound, the non-extraction is obtained from the timing frequency distribution obtained by performing time-frequency analysis processing on the vibration data (data) from the determination target object. The non-fixed vibration data at the time when the intensity of the fixed vibration is equal to or higher than the set value, and the probability of occurrence of the abnormal sound is determined based on the extracted non-fixed vibration data (Patent Document 1); the vibration of each frequency for the time-frequency analysis result When the frequency is generated, the time ratio of the data which is equal to or greater than the vibration generation determination amplitude value obtained by the product of the maximum amplitude of the frequency and the vibration generation threshold value is calculated, and the frequency of occurrence of the frequency is calculated (Patent Document 2) And calculating a region having a large intensity indicated by a contour line from the sound component of the time-series spectrum, and extracting only the spectrum column including the noise component from the region (Patent Document 3).

In addition, the second technique of the abnormal sound diagnosis device for the conventional diagnosis abnormal sound is a special treatment means for seeing a specific known abnormal phenomenon and confirming whether or not the abnormal phenomenon occurs, and is known. When the abnormal phenomenon is not generated, general noise analysis is performed, and the analysis result is compared with the normal time to detect an abnormal phenomenon that is not specific, and when an abnormal phenomenon is detected, a change is detected to detect the normal state. The processing procedure is given to a dedicated processing means (Patent Document 4).

(previous technical literature) (Patent Literature)

Patent Document 1: Japanese Patent No. 3885297

Patent Document 2: Japanese Patent No. 4373350

Patent Document 3: Japanese Patent No. 4262878

Patent Document 4: Japanese Laid-Open Patent Publication No. Hei 6-309580

The conventional first technique is configured to detect abnormalities based on various thresholds that appear to be appearance patterns of specific abnormal sound components that appear in the frequency analysis result, and therefore cannot be equally high in each individual configuration. Accuracy to diagnose problems with abnormal tone components of various frequency bandwidths and durations. Furthermore, since the intensity of the time-frequency distribution is large from the observation data in a bottom-up manner, there is a problem that it is impossible to obtain the most appropriate diagnosis result.

On the other hand, in order to diagnose an unknown abnormal phenomenon, the conventional second technique has a problem that it is necessary to register a diagnosis procedure for an unknown abnormal phenomenon in advance to a dedicated processing means.

The present invention has been made in order to solve the above problems, and an object of the present invention is to provide a detection method without registering a dedicated diagnosis procedure for a dedicated operation. The processing means can guarantee the most appropriate diagnosis result, and can equally detect the abnormal sound component having various band widths and durations included in the frequency analysis result with high precision.

The abnormal sound diagnosis device according to the present invention includes: waveform data acquisition means for acquiring waveform data of sound or vibration generated by the inspection target device; and time frequency analysis means performing time-frequency analysis on the waveform data, and obtaining One of the axes is the time axis, and the other axis is the time-frequency distribution of the frequency axis; and the region extracting means generates a plurality of regions defined by the coordinate values of the time axis and the frequency axis of the time-frequency distribution, and extracts and includes An area having a variation component different from the fixed state of the time-frequency distribution; and a determination means for performing an abnormality determination based on the time-frequency distribution included in the extraction area and outputting the abnormality.

According to the abnormal sound diagnostic apparatus of the present invention, by providing a means for extracting a region of time frequency continuously formed with respect to time and frequency in accordance with the time-frequency distribution, it is possible to perform the diagnostic procedure without registration, and the same accuracy can be achieved. Diagnosing the effects of various frequency components with bandwidth and duration appearing in the frequency analysis results.

(Embodiment 1)

In the device for diagnosing the abnormal sound pressure generated by the device constituting the inspection target system, the present embodiment is installed as a software on a personal computer (hereinafter referred to as a PC) and has a software. The learning mode (mode) of the waveform at the normal time and the diagnostic mode of the waveform at the time of the test are obtained. The tester sets a microphone or a vibration sensor or the like to the inspection target machine, and connects the microphone or the vibration sensor to the USB (Universal Serial Bus) interface of the PC. The input terminal is operated in the learning mode and the diagnostic mode.

For the inspection target system, for example, the microphone is mounted on the elevator car, and the signal of the microphone is transmitted to the PC located in the machine room via the control cable, and the lifting path is diagnosed by running the car back and forth. The operation sound of each machine inside.

A specific machine is, for example, when an abnormal sound is generated from the top return pulley, and the operation sound generated by a machine other than the machine that generates the abnormal sound, such as a slide sound of a guide rail, is generated in the approaching of the abnormal sound of the guide rail. The time zone of a particular machine, for example, when the car is near the top return pulley time zone (the second half of the measurement zone when the car is ascending, or the first half of the measurement zone when the car is descending) Frequency component.

Further, when an abnormal sound is generated from the counterweight, the time-frequency component of the abnormal sound appears in the central portion of the measurement section belonging to the time zone in which the compartment and the weight are interlaced.

Furthermore, the shape of the frequency spectrum of the abnormal sound differs depending on the machine in which the abnormality occurs and the cause of the abnormality, and there are many kinds of frequency ranges. In general, as described above, when the operation sound of a plurality of machines is diagnosed according to the microphone of the scanning machine system, an abnormal sound component appears in the measurement section. The time range and frequency range are extremely complex and diverse.

In the second diagram, the horizontal axis of the time-frequency distribution when the abnormal sound of each of the elevators is generated is set as the time, the vertical axis is set as the frequency, and the intensity of the distribution of each time and each frequency is displayed in shade. The dotted line system indicates the intensity of the operation sound of the specific machine before the abnormality is generated in the microphone, and the solid line is the intensity of the operation sound of the machine which is abnormal. Furthermore, the point chain system displays the intensity of the sound that is synthesized by the operating sounds of all machines from the machine. (A) is an example in which an abnormal sound is generated from the top return pulley, and (B) is an example in which an abnormal sound is generated from the weight.

Fig. 1 is a block diagram showing the abnormal sound diagnosis apparatus according to the first embodiment of the present invention.

In Fig. 1, 1 is a measurement signal outputted by a microphone and a vibration sensor, and 2 is provided with an amplifier, a low-frequency filter circuit, and an AD converter, and the measurement signal 1 is sampled and converted into a digital position. The (digital) signal outputs the waveform acquisition unit of the waveform data 3, and the 4 series time-frequency analysis unit adds a time window to the waveform data 3, and shifts the time window in the time direction while using the fast Fourier (Fourier). The transform (hereinafter abbreviated as FFT) calculus performs time-frequency analysis on the waveform data 3, and outputs a time-frequency distribution 5 composed of spectral values showing the intensity with respect to time and frequency.

6-time memory time-frequency distribution 5 normal time-frequency distribution 6a (not shown) normal time-time frequency distribution memory; 7-time memory time-frequency distribution 5 test time-frequency distribution 7a (not shown) Time-frequency distribution memory unit during testing; 8-series prior knowledge 8a (not shown) as a table (Table) The pre-knowledge memory unit to be memorized; 9 is based on the prior knowledge 8a of the prior knowledge storage unit 8 to generate the region candidate generation unit for the predetermined predetermined region candidate 10; 11 is for the region candidate 10, and refers to the normal time. The normal time time frequency distribution 6a of the time frequency distribution memory unit 6 and the test time frequency distribution 7a of the time frequency distribution memory unit 7 during the test, the evaluation unit for calculating and outputting the condensation degree 12; and the 13 system area extraction unit are based on The condensing degree 12 is selected from the region candidate 10 and the most appropriate region candidate is selected and output as the extraction region 14; the 15-system abnormal time calculation unit refers to the normal time frequency distribution 6a and the diagnostic time frequency distribution 7a, and is included in the extraction. The time-frequency distribution of the region 14 is used to calculate the degree of abnormality indicating the probability of occurrence of the abnormal sound and is output as the abnormality 16; 17 is based on the abnormality 16 to determine the possibility of abnormal sound generation, and outputs the determination portion of the determination result 18. .

The operation will be described below with reference to the processing flowchart of Fig. 4 .

In the learning mode or the diagnostic mode, the waveform acquisition unit 2 acquires and amplifies the measurement signal 1 outputted by the microphone and the vibration sensor to perform AD conversion, thereby converting to a 16-bit linear PCM (pulse code) having a sampling frequency of 32 kHz. Modulation, pulse code modulation) The waveform data of the digital signal 3 (step S1).

The time-frequency analysis unit 4 cuts out the frame of the 1024-point time window by dividing the time window of 1024 points in the time direction by an interval of 16 ms, and cuts out the frame by each The frame performs an FFT calculation to obtain a time series y(t, f) of the frequency spectrum, and outputs it as a time frequency distribution 5 (step S2). Here, the t system indicates that the acquisition corresponds to At the time of analyzing the discrete value of the shift interval of the window offset, f is the frequency at which the discrete value of the frequency index corresponding to the result of the FFT calculation is obtained. Further, the time t and the frequency f satisfy the relationship of 0≦t≦T and 0≦f≦F, respectively. Here, T is the time width in the time direction of the time frequency distribution 5, and F is the Nyquist frequency (F=fs/2) which is 1/2 of the sampling frequency fs of the waveform data 3.

When the time frequency distribution 5 is calculated by the time frequency analysis unit 4, the abnormal sound diagnosis device determines that it is in the learning mode or the diagnosis mode (step S3).

In the learning mode, the time-frequency distribution 5 is transmitted to the normal-time time-frequency component storage unit 6 as the normal-time time-frequency component 6a (step S4). On the other hand, when the result of the determination in step S3 is the diagnosis mode, the time-frequency distribution 5 is transmitted to the diagnostic-time-time-frequency component storage unit 7 as the diagnosis-time-frequency component 7a and is memorized (step S5).

Next, the operation of the diagnosis processing in the diagnosis mode will be described.

The area candidate generating unit 9 generates the area candidate 10 based on the prior knowledge 8a (step S6). The prior knowledge 8a is used to define the knowledge of the shape of the appearance region of the time-frequency distribution of the abnormal component generated by the device constituting the system to be diagnosed, and to display the knowledge obtained by the designer of the device in analyzing the object in advance, and as the device The area candidate generated by the area candidate generating unit 9 is stored in the form of the prior knowledge storage unit 8 in a table format. In this example, the entire time interval T is divided into N for all regions of the time-frequency distribution, and all the band frequencies F are divided into m to obtain a lattice-shaped divided region, and an arbitrary lattice line is generated as an edge. The rectangular area is stored in the prior knowledge storage unit 8 as a table of prior knowledge 8a.

Fig. 3 shows an example in which A and B are used as the rectangle of the grid as the prior knowledge 8. The rectangular area A is formed into an optimum shape with respect to a time-frequency component having a short time among frequency components of a high frequency in the latter half of the time interval. Further, the rectangular region B is formed into an optimum shape with respect to a time-frequency component having a long duration in the intermediate frequency band in the time interval before the measurement time. Here, by increasing the number of divisions N and m, the boundary of the region can be expressed in more detail. However, in the first 1/6th time interval and the last 6th/6th time interval of the lattice-shaped divided region, since the operation speed of the inspection target is slower than the rated speed, the operation sound is not sufficiently generated. Therefore, it can be removed from the generation of regional candidates. In addition, as described above, in the entire region of the time-frequency distribution, it is described that the entire time interval T is N-divided, and all the band frequencies F are m-divided to obtain a lattice-shaped divided region, and an arbitrary one is generated. The grid line is an example of a rectangular area of the side. However, depending on the prior knowledge of the abnormal component with respect to the time-frequency component, the lattice-shaped divided area may be selected or not selected to generate an area of an arbitrary shape.

The evaluation unit 11 calculates the condensing degree E(R) 12 for the region candidate 10 (hereinafter, R indicates the region candidate) (step S7).

The condensing degree E(R) is set to y(t, f) when the test is performed, x (t, f) is set in the normal time, and R = [t1, t2 is set in the rectangular region. In the case of f1, f2], the degree of condensation E(R) with respect to the above can be obtained by the calculation shown in Formula 1. Here, t1, t2, f1, and f2 are the lower limit time, the upper limit time, the lower limit frequency, and the upper limit frequency of the rectangular region R, respectively. Furthermore, the candidate R system other than the rectangle can be obtained by the calculation of the general formula 2 instead of the formula 1. Here, the mark (t, f) The R* means that the sum is obtained for the combination of the discrete time t and the discrete frequency f included in the extracted region R*.

In the above formula, n is the number of specimens included in the spectral value of the rectangular region of the time-frequency distribution. Further, w(n) is a weighting coefficient corresponding to the number n of the specimens, for example, a p-th root of the number n of specimens (p is, for example, 2). Since the number n of specimens becomes a large value depending on the size of the region, and the weighting coefficient w(n) becomes a large value depending on the size of the region, the condensation degree E(R) is relatively small. The system becomes smaller and is used to mitigate the influence of the outliers that exist in a small area on the calculation results. Again, the function The spectral value is transformed into a nonlinearity, and in order to approximate the distribution of the transformed values to a normal distribution, a Box-Cox transform (also referred to as generalized logarithmic transformation) or a logarithmic transformation is used. The Box-Cox transformation is consistent with the logarithmic transformation when the parameter γ represented by Equation 3 is γ=0.

The area extracting unit 13 checks the relationship between each region candidate and the condensing degree E(R) with respect to each region candidate, and selects and outputs the region candidate whose condensed degree E(R) is displayed as the maximum value as the most appropriate extracted region (step S8). ). Each region candidate is set to {R1, R2, ..., Rk}, and the condensing degree of each is set to {E(R1), E(R2), ..., E(Rk)}, and the most appropriate region candidate is used. When R* is set, R* can be obtained by the calculation of Equation 4. Here, the k of the natural number is the number of regional candidates.

The abnormality calculating unit 15 calculates the abnormality based on the spectral value included in the most appropriate extracted region R* of each of the normal time frequency distribution x(t, f) and the test time frequency distribution y(t, f). (Step S9). Here, the extraction region R* is a rectangular region, and R*=[t1, t2, f1, f2], and when the degree of abnormality is a (R*), the degree of abnormality a(R*) is obtained by Equation 5 Calculus The value obtained. Here, t1, t2, f1, and f2 are as defined above.

In the above formula 5, Ψ(x) can use a nonlinear mapping function of the variable x, for example, using the above-described Box-Cox transformation or the like. g(t) is a value obtained by dividing the cumulative value of the frequency f direction of the region R* by the number of unit frequencies, that is, the average of the samples with respect to the frequency of time t, and h(f) is the region R* The cumulative value of the time t direction divided by the number of units of time, that is, the average of the samples with respect to the time of the frequency f. Furthermore, (t) and (f) is a value obtained by smoothing the time t for g(t) and smoothing the frequency f for h(f). Smoothing can be achieved by taking a moving average. In the end, I found the moving average. (t) about the maximum value of time t, and the moving average Any one of the maximum values of the maximum value of the frequency f of (f) is taken as the abnormality a (R*). You can also use the quantile that belongs to the statistic instead of the maximum value, or you can use the value of either side as the degree of abnormality. This example is shown by a ₁ (R*), a ₂ (R*), a ₃ (R*), a ₄ (R*), a ₅ (R*), and the like of the formula 6. Here, quantitile({x}, α ) shows the alpha quantile of the sequence {x}. If α is substituted into 1 , it is consistent with the maximum value max{x}. The α and β systems of Formula 6 are close to a value of 1, for example, 0.9 may be substituted.

Furthermore, in another simpler method, as shown by a ₆ (R*) of Equation 7, the degree of abnormality a (R*) can also be set as the normal time-frequency distribution of the extracted region R*. The difference between the average value of the map Ψ(x(t, f)) and the average value of the map Ψ(y(t, f)) of the time-frequency distribution of the test region R*.

The determination means 17 compares the abnormality a (R*) with the threshold value, and when the abnormality degree is equal to or greater than the threshold value, it is determined that an abnormal sound is likely to occur, and an "alarm" is output as the determination result 18 (step S10). In addition, when the degree of abnormality is equal to or less than the threshold value, it is determined that the possibility of generating an abnormal sound is low, and "normal" is output as the determination result 18.

In the above embodiment, the time-frequency analysis unit 4 is configured to output the time-frequency distribution 5 by FFT calculation. However, it is not limited to the FFT, and wavelet transform may be used.

Further, for the rectangular region of the pre-memory 8a stored in the prior knowledge storage unit 8, the lower limit tmin may be set for the difference t2-t1 between the upper limit time t2 and the lower limit time t1. That is, t2-t1≧tmin is limited to the rectangular area and stored in the prior knowledge memory unit 8.

Further, similarly, the lower limit fmin may be set to the difference f2-f1 between the upper limit frequency f2 and the lower limit frequency f1. That is, t2-t1≧tmin is limited to the rectangular area and stored in the prior knowledge memory unit 8.

Furthermore, in addition to the analytic function, the nonlinear function may also have a nonlinear function by approximating the polyline.

As described above, according to the present invention, by providing a means for extracting a region in which the time frequency is continuously formed for the time frequency in accordance with the time-frequency distribution, it is possible to perform the diagnostic procedure without registration, and the diagnosis can be performed with high precision. The result of the frequency analysis has various effects of abnormal tone components with bandwidth and duration.

Further, the means for generating the region candidate by the use region, the means for evaluating the degree of goodness (condensation) of the candidate for the region to be generated, and the means for selecting the region having the greatest degree of goodness (condensation) There is no need to use various thresholds that are specialized into the appearance pattern of the specific abnormal sound component, that is, among the candidates of all the regions generated by the region candidate generating means, the most appropriate region for extracting the best evaluation value is extracted. The role. In this way, it is possible to perform the diagnostic procedure without registration, and it can be equally high. Accuracy diagnosis occurs in the frequency analysis results with a variety of abnormal tone components with bandwidth and duration.

Further, in terms of the degree of goodness (condensation degree) of the region candidate, the number of corresponding specimens is added as a weight for the variation amount of the time-frequency distribution from the normal time, whereby the smaller the number of specimens, the smaller the number of specimens The smaller the weight, the larger the weight of the specimen, the larger the weight. Therefore, if the variation is the same, the number of specimens in the extraction area selected by the club is as large as possible (the area of the equivalence area is larger). The effect, and assuming that the amount of variation is large, the degree of goodness (condensation) of the region where the number of specimens is small (the area of the equivalence region is small) becomes small. Thereby, since the balance of the amount of variation and the number of specimens is preferably a region where the balance is preferably large, the effect of improving the accuracy of the diagnosis is enhanced.

Furthermore, in terms of the characteristic parameters of the distribution compared with the normal time, in the case of using the average of the specimen, it is possible to obtain a meaningful result when the distribution of the specimen follows the normal distribution, but since the actual spectral value is non-negative Asymmetric distribution, so that the distribution is close to the normal distribution by nonlinear transformation, and meaningful comparison can be made even if the specimen is averaged. Thereby, the evaluation of the degree of goodness (condensation degree) of the region can be appropriately performed, and as a result, since the possibility of the abnormal sound can be determined based on the appropriately extracted region, there is an effect of improving the accuracy of the diagnosis.

Further, the parameter for obtaining the condensing degree has a characteristic (compression characteristic) which is nonlinear with respect to the number of specimens, and is a number corresponding to the number of specimens included in the region candidate, whereby the prevention region can be exhibited even if the variation is small. The number of specimens (equivalent to the area) has become extremely powerful. Take this out In the region, it is possible to extract a balanced region having a large variation and a large number of specimens, and as a result, it is effective to improve the diagnostic accuracy based on the judgment results performed as described above.

Further, by limiting the shape of the generated region candidate to a rectangle, in general, it is effective to prevent an area of a shape other than a rectangle from being unexpectedly extracted from being presumably impossible. Thereby, in terms of the extracted area, an appropriate area can be extracted, and as a result, the effect of improving the diagnostic accuracy based on the determination result of the area can be improved.

Similarly, by using the prior knowledge of the time-frequency distribution of the variation component to define the shape of the region, it is effective to not accidentally extract an area that does not exist in the prior knowledge. Thereby, in terms of the extraction area, an appropriate area can be extracted, and as a result, there is an effect of improving the diagnostic accuracy according to the determination result performed as described above.

Similarly, by using the prior knowledge of the operational state of the machine to define the shape of the area, it is effective to not accidentally extract an area that does not exist in the prior knowledge. Thereby, in terms of the extraction area, an appropriate area can be extracted, and as a result, there is an effect of improving the diagnostic accuracy according to the determination result performed as described above.

(industrial availability)

The abnormal sound diagnosis device of the present invention can be used as a detection device that detects a position of an abnormal state by using a system device configured by combining a plurality of devices, for example, an elevator.

1‧‧‧Measurement signal

2‧‧‧ Waveform acquisition department

3‧‧‧ Waveform data

4‧‧‧Time Frequency Analysis Department

5‧‧‧Time frequency distribution

6‧‧‧Normal time frequency distribution memory

7‧‧‧Time frequency distribution memory during testing

8‧‧‧Pre-existing knowledge memory

9‧‧‧Regional Alternate Generation Department

10‧‧‧Regional alternates

11‧‧‧Evaluation Department

12‧‧‧Condensation

13‧‧‧Regional extraction department

14‧‧‧Extracted area

15‧‧‧Anomaly calculation department

16‧‧‧ anomalies

17‧‧‧Decision Department

18‧‧‧Results

S1 to S10‧‧‧ steps

Figure 1 is a functional block showing the abnormal sound diagnosis device of the present invention. Make up the picture.

Fig. 2 is a characteristic diagram showing an example of time-frequency distribution when scanning sounds from a plurality of machines.

Figure 3 is an explanatory diagram of prior knowledge of the region of time-frequency distribution.

Fig. 4 is a flow chart showing the processing of the first embodiment of the present invention.

1‧‧‧Measurement signal

2‧‧‧ Waveform acquisition department

3‧‧‧ Waveform data

4‧‧‧Time Frequency Analysis Department

5‧‧‧Time frequency distribution

6‧‧‧Normal time frequency distribution memory

7‧‧‧Time frequency distribution memory during testing

8‧‧‧Pre-existing knowledge memory

9‧‧‧Regional Alternate Generation Department

10‧‧‧Regional alternates

11‧‧‧Evaluation Department

12‧‧‧Condensation

13‧‧‧Regional extraction department

14‧‧‧Extracted area

15‧‧‧Anomaly calculation department

16‧‧‧ anomalies

17‧‧‧Decision Department

18‧‧‧Results

Claims (6)

An abnormal sound diagnosis device includes: a waveform data acquisition means for acquiring waveform data of sound or vibration generated by an inspection target device; and a time frequency analysis means for performing time and frequency analysis on the waveform data, and obtaining one of the waveform data The axis is the time axis, and the other axis is the time-frequency distribution of the frequency axis; the region extracting means generates a plurality of regions defined by the coordinate values of the time axis and the frequency axis of the time-frequency distribution, and extracts and includes And a determination means for determining and outputting an abnormality based on a time-frequency distribution included in the extraction region, wherein the region extraction means includes: an area candidate generation unit; Extracting a region including a variation component different from the fixed state of the time frequency distribution as a region candidate; and an evaluation unit determining the condensation degree according to a relationship between a time frequency distribution included in the region candidate and a time-frequency distribution at a normal time And configured to have the above-mentioned area with a large degree of condensation Complement output as the extracted area. The abnormal sound diagnosis apparatus according to claim 1, wherein the evaluation unit performs nonlinear transformation on a time-frequency distribution included in the region candidate, and performs a time-frequency distribution through nonlinear transformation. The characteristic parameter, the characteristic parameter of the time-frequency distribution in the normal state of nonlinear transformation, and the calculation of the number corresponding to the number of specimens included in the region candidate are used to determine the degree of condensation. The abnormal sound diagnosis apparatus according to claim 2, wherein the evaluation unit is configured to obtain the nonlinear transformation of the condensing degree, and to use a transformation function having nonlinear characteristics with respect to the intensity. The abnormal sound diagnostic apparatus according to the second aspect of the invention, wherein the evaluation unit is configured to obtain a number corresponding to the number of specimens included in the region candidate, and to set a nonlinearity for the number of specimens. The function of the feature is applied to the number of specimens. The abnormal sound diagnosis device according to claim 3, wherein the evaluation unit is configured to obtain a number corresponding to the number of specimens included in the region candidate, and to set a nonlinearity for the number of specimens. The function of the feature is applied to the number of specimens. The abnormal sound diagnosis device according to any one of the first to fifth aspects of the present invention, wherein the prior knowledge storage unit is configured to analyze the inspection target machine beforehand to specify an abnormality generated from the machine. The prior knowledge of the shape of the appearance region of the time component of the sound component is stored as a table; and the region extracting means is generated based on the rectangular region candidate stored in the table of the prior knowledge memory.