JPH11241945A - Foreign sound inspecting device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a foreign sound inspecting device capable of shortening the identifying time and performing a detection excellent in precision. SOLUTION: This device is formed of a detecting means 2 for taking out signals of the sound pressure of generated sound and vibration of a rotating machine 1, a calculating means 7 for analyzing the signals taken by the detecting means 2 and finding a plurality of feature quantity data, a teacher data forming means 8 for inputting feature quantity data of the same kind as the feature quantity data of the generated sound of a model rotating machine 1 consisting of the same as the rotating machine 1 the presence of abnormality of which can be judged and forming a teacher data, and a discriminating means 9 for discriminating the presence of abnormality of the rotating machine 1 from the feature quantity data found by the calculating means 7 by a neural network constructed by learning on the basis of the teacher data inputted from the teacher data forming means 8 and the presence of abnormality of the model rotating machine 1.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an abnormal sound inspection apparatus that can identify the presence or absence of an abnormality in an object to be inspected in a short time from a plurality of feature data by using a neural network. is there.

[0002]

2. Description of the Related Art A conventional abnormal noise inspection apparatus for inspecting abnormal noise includes:
For example, the output of a vibration detector described in JP-A-58-108419 is classified into a plurality of frequency components by a filter having a plurality of frequency bands, and provided in correspondence with each frequency component. In some cases, the threshold value and the maximum peak value in each frequency component are compared for each frequency component to perform abnormality identification.

Japanese Patent Application Laid-Open No. 58-219424 discloses that a frequency spectrum is calculated by a vibration detector, and an amplitude value in a specific frequency spectrum is compared with a threshold value to determine whether the balance is poor or whether a bearing is good or bad. Inspection is underway.

Further, Japanese Patent Application Laid-Open Nos. 59-97016, 59-97017, and 59-1970 have been disclosed.
In Japanese Patent Publication No. 09831, the abnormal noise is identified by comparing the difference between the difference between the frequency spectrum converted from the vibration or noise to be inspected and the reference spectrum and the threshold value.

[0005]

In each of the conventional abnormal sound detecting devices, it takes time to set a threshold value as a reference for identification.
It took a long study time to improve the identification accuracy. Also, when comparing only specific frequency components,
There is a possibility that an abnormality outside the band to be inspected may be overlooked, which hinders improvement in accuracy.

[0006] If many frequency components are compared to solve this problem, a threshold must be set for each of the frequency components, which takes time. Further, if at least one of the frequency components exceeds the threshold value, it is determined that the frequency components are defective even if many other frequency components are determined to be good, and desired identification cannot be performed. was there.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and has as its object to provide an abnormal noise inspection apparatus capable of shortening the identification time and performing highly accurate detection.

[0008]

Means for Solving the Problems Claim 1 according to the present invention.
The abnormal noise inspection apparatus includes a detecting means for extracting one or both signals of sound pressure and / or vibration of a sound generated from an object to be inspected, and analyzing a signal extracted by the detecting means to obtain a plurality of feature amount data. Calculating means, and inputting the same kind of feature amount data as a plurality of feature amount data of a sound generated from a model object which is the same as the object to be inspected whose presence or absence of an abnormality can be determined; An identification means is provided for judging the presence or absence of an abnormality in the object to be inspected by comparing the correlation between each feature amount data whose presence or absence can be determined and the characteristic amount data of the sound generated from the object to be inspected.

Further, according to a second aspect of the present invention, there is provided the abnormal sound inspection apparatus according to the first aspect, further comprising: teacher data creating means for creating feature amount data of a sound generated from the model object as teacher data; The neural network is constructed by learning the neural network based on the teacher data input from the teacher data creating means and the presence / absence of an abnormality in the model object. Processing is performed on the network to identify the presence or absence of an abnormality in the inspection object.

According to a third aspect of the present invention, in the abnormal noise inspection apparatus according to the first or second aspect, when the teacher data is composed of feature amount data of a model object including only good products, In the above, the inspection object is identified as good or bad.

According to a fourth aspect of the present invention, there is provided the abnormal noise inspection apparatus according to the first or second aspect, wherein the teacher data is characterized in that the model data includes a good product and a plurality of types of defective products. When constituted by the quantity data, the identification means identifies good and bad types of the inspection object.

According to a fifth aspect of the present invention, there is provided an abnormal noise inspection apparatus according to any one of the first to fourth aspects.
The calculating means is for performing time-series analysis on the signal extracted from the detecting means to obtain a plurality of feature amount data for each unit time.

According to a sixth aspect of the present invention, there is provided an abnormal noise inspection apparatus according to any one of the first to fourth aspects.
The calculating means analyzes the frequency of the signal extracted from the detecting means and obtains characteristic amount data having an amplitude for each unit frequency.

According to a seventh aspect of the present invention, there is provided an abnormal noise inspection apparatus according to any one of the first to fourth aspects.
The calculating means performs a time-frequency analysis of the signal extracted from the detecting means, and obtains characteristic amount data composed of an amplitude of a unit frequency component for each unit time.

[0015]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiment 1 Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing a configuration of an abnormal noise inspection device according to Embodiment 1 of the present invention. In the figure, reference numeral 1 denotes a rotating machine as an object to be inspected, and 2 denotes detecting means for detecting a sound generated by the rotating machine 1, for example, one or both of a sound pressure and a vibration of a sound generated by the rotating machine 1. Is formed by using a converter 3, for example, a microphone or a vibration acceleration pickup, which converts the data of FIG.

An amplifier 4 connected to the converter 3 for amplifying the signal of the converter 3 and an amplifier 4 connected to the amplifier 4 for removing noise from the signal of the amplifier 4 and converting the signal into a plurality of frequency components And an A / D converter 6 connected to the filter 5 and taking in the signal of the filter 5.

Numeral 7 is a calculating means for analyzing a signal detected by the detecting means 2 and obtaining a plurality of feature data. For example, a CPU or a DSP (digital signal processor) mounted on an electronic computer is used. . Reference numeral 8 denotes a teacher data creating unit that uses a model rotating machine as a model object composed of the same object as the object to be inspected for which the presence or absence of an abnormality can be determined, detects the sound generated from the model rotating machine, and detects the characteristic amount described above. The same type of feature data as the data is extracted and created as teacher data.

Reference numeral 9 denotes a neural network constructed by learning based on the teacher data input from the teacher data creating means 8 and the presence or absence of an abnormality in the model object. The neural network 9 identifies the feature amount data obtained by the calculating means 7. It is an identification unit that outputs an identification result of the presence or absence of an abnormality in the rotating machine 1.

The operation of the abnormal noise inspection device according to the first embodiment configured as described above will be described with reference to FIGS. First, before discrimination of the sound generated from the object to be inspected, teacher data is created from a model rotating machine composed of the same object as the object to be inspected for which the presence or absence of an abnormality can be determined. You need to keep it.

This operation will be described with reference to FIG. First, preparation is performed by operating a classified model rotating machine capable of determining the presence or absence of an abnormality (step S1 in FIG. 2). next,
A signal of one or both of the sound pressure and the vibration of the rotating machine 1 is captured by the detecting means 2 (step S in FIG. 2).
2). Next, a plurality of feature amount data is calculated from the signal using the calculating means 7 (step S3 in FIG. 2).

Next, teacher data is created from the feature data (a specific example of the feature data will be described later) using the teacher data creating means 8 (step S4 in FIG. 2). In this case, for example, the lower limit value which the feature amount data can take is 0 and the upper limit value thereof is 1, and the feature amount data is normalized,
Created as teacher data.

Next, when the number of teacher data is a desired number N (here, the desired number N is determined based on the required accuracy of identification. When the number N of teacher data increases, the accuracy of identification increases. However, it takes time to generate a large amount of teacher data, so it may be determined according to the required accuracy.) (Step S5 in FIG. 2). ), If not reached, return to step S1 in FIG. 2 again, and the rotating machine 1 as another model
And the same operation as above is repeated.

If the number of teacher data has reached a desired number N, the teacher data is transmitted to the identification means 9, and a neural network constructed by learning based on the teacher data and the presence or absence of abnormality in the model rotating machine is constructed. An identification unit 9 having a network is provided (step S6 in FIG. 2).

Next, the operation is shifted to the operation of discriminating the presence / absence of the object to be inspected, which is the original purpose. About this behavior,
This will be described with reference to FIG. First, the rotating machine 1, which is the object to be inspected, is prepared by operating (Step S7 in FIG. 3). Next, the sound generated by the rotating machine 1 is detected by the detecting means 2, and a signal of the generated sound is captured (step S8 in FIG. 3).

Next, the calculating means 7 calculates a plurality of feature data of this signal (step S9 in FIG. 3). At this time, the feature amount data is normalized by setting the lower limit value that the feature amount data can take as 0 and the upper limit value as 1. Next, using the normalized feature amount data, the identification means 9 determines the correlation between the generated sound and the trained teacher data using the neural network, and identifies the generated sound (FIG. Step S10). Since the correlation is determined in this way, it is not necessary to set the threshold value corresponding to each feature amount data.

Next, it is determined whether the teacher data is composed of a plurality of data or a single data (step S11 in FIG. 3). Here, the classification refers to the classification of the good, bad, or bad type of the rotating machine 1 or the type of the abnormal part due to the badness, and the identification means 9 is obtained in advance using the model rotating machine. It corresponds to the number of classes of teacher data.

For example, when the teacher data of the model rotating machine is composed of one class of data of only good data having no abnormality, it is determined that the class is 1. If the teacher data of the model rotating machine is composed of two types of data including good data having no abnormality and bad data, the classification is determined to be 2. Further, there may be a case where the data includes a plurality of types of defective data or a case where the data includes data of the type of an abnormal portion due to a defect.

Next, when the number of classifications is determined to be 1, that is, here, as described above, data of only non-defective products is
In the case of learning by a neural network, an evaluation point for good is obtained by this neural network. (The evaluation point of the neural network here means that the smaller the score, the closer to the classification. Shall be represented). Then, based on the evaluation points, the rotating machine 1 is identified as good or not good (here, it is set to be determined to be defective) (step S1 in FIG. 3).
2).

As an example of this evaluation, for example, a threshold value of an evaluation point determined to be good is set, and if the obtained evaluation point is larger than the threshold value, it is determined that the evaluation point is bad, and if the evaluation point is equal to or less than the threshold value. In such a case, it may be determined to be good.

When the number of classifications is determined to be 2 or more,
In other words, here, as described above, when data such as good, bad, and the type of the defect, and the type of the abnormal part due to the defect are learned by the neural network, the evaluation score for each classification is given by the neural network. Will be obtained. Then, among the evaluation points of these classifications, the classification of the minimum evaluation point is identified (step S in FIG. 3).
13). Then, the type of the rotary machine 1 is identified as good, defective, or defective, or the type of abnormal part due to the defect is identified.

Here, an explanation will be given of the analysis of the actual signal generated by the rotating machine 1 by the detection means 2 and the characteristic amount data of the signal. First, here, feature amount data is obtained by wavelet analysis as one method of time-frequency analysis. FIG. 4 shows data of non-defective products, and FIG. 5 shows data of defective products. 4 (a) and 5 (a) show signals of the detecting means 2 and wavelet analysis of these signals to obtain respective feature data.
(B) and FIG. 5 (b).

The graphs of FIG. 4A and FIG.
The horizontal axis indicates time, and the vertical axis indicates voltage. FIG.
The graphs of FIG. 5B and FIG. 5B show time on the horizontal axis and frequency on the vertical axis, and it can be confirmed that data by wavelet analysis is obtained in a matrix.
The level of the amplitude of the unit frequency component for each unit time obtained in the form of a matrix is indicated by the shading of the color. The white color has a higher level, and the black color has a lower level.

Then, the calculation means 7 normalizes the data obtained in this way and calculates it as feature data. Then, the identification means 9 makes all of the matrix-like feature amount data obtained in this way as a determination target. In this case, in practice, the data shown in FIG. 4B is learned as a classification of good data, and the data shown in FIG. Judgment of the correlation between sound data and these data, judgment that the sound is similar to the data that is closer, judgment of good or bad sound, and identification of abnormalities in the inspected object Will be done.

According to the abnormal noise inspection apparatus of the first embodiment formed as described above, the presence / absence of an abnormality of the inspection object is determined from a plurality of feature data by setting a threshold value for each feature data. This can be performed without setting, and the desired determination can be performed in a short time. In addition, it is not necessary to set a threshold value only with data of non-defective products, so that it is possible to determine whether or not the inspection target has an abnormality. Also, depending on the type of teacher data, it is possible to identify the type of defect and the type of abnormal part due to the defect.

In addition, in the conventional method, it is necessary to set a threshold value for each feature amount data in order to determine a large number of feature amount data such as a matrix, and it takes a long time to set. Although it was not possible in practice, it is not necessary to set a threshold value for each feature amount data if the determination is made by a neural network as in the present invention, and the time is short. Makes it possible to make a desired determination.

Further, in the first embodiment, the case where the identification means 9 is performed by a neural network has been described. However, the present invention is not limited to this. If the correlation between the feature data and a plurality of feature data of the inspection target can be determined, a threshold is set for each feature data to determine whether the inspection target has an abnormality. Needless to say, it is possible to judge without doing.

In the first embodiment, the data obtained by the time-frequency analysis is used as the feature data. However, the present invention is not limited to this. For example, the standard deviation of the frequency component per unit time can be obtained by the time-series analysis. , Peak count, and
A plurality of feature data such as a variation width may be analyzed as feature data. Next, an example of the analysis result in this case will be described.

For example, if the signal obtained by the detection means 2 shown in FIGS. 4A and 5A indicates a sound pressure, the unit time is 0.2 sec, and A plurality of feature data such as the standard deviation of the sound pressure, the number of peaks, and the fluctuation range are obtained as shown in FIG. It is needless to say that the same effect as in the first embodiment can be obtained by determining these data as feature amount data in the same manner as in the first embodiment.

As another analysis method, the amplitude of each unit frequency may be analyzed as feature data by frequency analysis. In this case, the feature data is, for example, fast Fourier transform or 1/3 octave. What is necessary is just to obtain | require as a spectrum amplitude for every unit frequency by using analysis. An example of the analysis result in this case is shown.

For example, when the signal obtained by the detection means 2 shown in FIGS. 4A and 5A is obtained as a plurality of feature data of the spectrum amplitude for each unit frequency, good data is obtained. FIG. 7A shows the defective data as shown in FIG. 7B. It is needless to say that the same effect as in the first embodiment can be obtained by determining these data as feature amount data in the same manner as in the first embodiment.

Although not particularly shown in the first embodiment, as means for detecting the sound pressure and vibration of the generated sound, for example, a current detector or an ultrasonic detector may be used to detect the sound pressure and vibration of the generated sound. It goes without saying that any device can be used as long as the feature can be detected.

Further, in the first embodiment, the one provided with the filter 5 as the detecting means 2 is shown, but it is not limited to this, and it goes without saying that the filter 5 may not be provided.

[0043]

As described above, according to the first aspect of the present invention, the detecting means for extracting one or both of the sound pressure and the vibration of the sound generated from the object to be inspected, and the detecting means A calculation means for analyzing the extracted signal to obtain a plurality of feature data, and the same as a plurality of feature data of a sound generated from a model object which is the same as the object to be inspected which can determine the presence or absence of an abnormality. Type of feature data, and a comparison of the correlation between each feature data that can determine the presence or absence of these abnormalities and the feature data of the sound generated from the inspected object,
Since there is provided identification means for judging the presence or absence of an abnormality in the inspection object, an abnormal noise inspection device capable of determining the presence or absence of an abnormality in the inspection object by comprehensive judgment from a plurality of feature amount data is obtained. It becomes possible.

Further, the abnormal sound inspection apparatus according to a second aspect of the present invention is characterized in that, in the first aspect, there is provided teacher data creating means for creating feature amount data of a sound generated from the model object as teacher data, and the identification means is provided. The neural network is constructed by learning the neural network based on the teacher data input from the teacher data creating means and the presence / absence of an abnormality in the model object. Since the processing is performed on the network to determine the presence or absence of an abnormality in the inspection object, it is possible to obtain an abnormal sound inspection apparatus that can process a large number of feature data in a short time.

According to a third aspect of the present invention, in the first or second aspect, when the teacher data is composed of feature amount data of a model object of only a good product, the identification means includes Since it is determined that the target object is good or not good, it is possible to obtain an abnormal noise inspection device that can determine the presence or absence of abnormality in the inspected object based on feature amount data of only non-defective products.

According to a fourth aspect of the present invention, in the first or second aspect, the teacher data is a non-defective product,
In the case of being configured from the feature amount data of each model object with a plurality of types of defective products, the identification unit identifies good and bad types of the inspected object. It is possible to obtain an abnormal sound inspection device that can make a determination.

According to a fifth aspect of the present invention, in any one of the first to fourth aspects, the calculating means comprises:
Since the signal extracted from the detection means is analyzed in time series to obtain a plurality of feature data per unit time, it is possible to obtain an abnormal noise inspection device capable of performing desired identification.

According to a sixth aspect of the present invention, in any one of the first to fourth aspects, the calculating means comprises:
Since the signal extracted from the detection means is subjected to frequency analysis to obtain characteristic amount data having an amplitude for each unit frequency, it is possible to obtain an abnormal noise inspection device capable of performing desired identification.

According to a seventh aspect of the present invention, in any one of the first to fourth aspects, the calculating means comprises:
Obtain an abnormal noise inspection device that can perform highly accurate abnormality identification because time-frequency analysis is performed on the signal extracted from the detection means to obtain feature amount data composed of the amplitude of a unit frequency component per unit time. Becomes possible.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration of an abnormal noise inspection device according to Embodiment 1 of the present invention.

FIG. 2 is a view showing a flowchart of an operation of the abnormal sound inspection apparatus shown in FIG.

FIG. 3 is a view showing a flowchart of an operation of the abnormal sound inspection apparatus shown in FIG. 1;

FIG. 4 is a diagram showing data of non-defective products according to the first embodiment of the present invention.

FIG. 5 is a diagram showing data of defective products according to the first embodiment of the present invention.

FIG. 6 is a diagram showing data according to the first embodiment of the present invention as data obtained by time-series analysis.

FIG. 7 is a diagram showing data according to the first embodiment of the present invention as data obtained by frequency analysis.

[Explanation of symbols]

1 rotating machine, 2 detecting means, 7 calculating means, 8 teacher data creating means, 9 identifying means.

Claims (7)

[Claims] 1. A detecting means for extracting one or both of a sound pressure and a vibration signal of a sound generated from an object to be inspected, and a plurality of feature data obtained by analyzing the signal extracted by the detecting means. Calculating means and inputting the same kind of feature amount data as the plurality of feature amount data of the sound generated from the model object which is the same as the object to be inspected which can determine the presence or absence of an abnormality; A comparison means for comparing the correlation between the characteristic data of each of which the presence or absence of the object can be determined and the characteristic data of the sound generated from the object to be inspected, and identifying means for judging the presence or absence of an abnormality of the object to be inspected. Characteristic abnormal noise inspection device. 2. A teacher data creating means for creating feature amount data of a sound generated from a model object as teacher data,
The identification means is performed by a neural network, and the neural network is constructed by learning based on the teacher data input from the teacher data creation means and the presence / absence of an abnormality in the model object. The abnormal noise inspection apparatus according to claim 1, wherein the characteristic amount data is processed by the neural network to identify the presence or absence of an abnormality in the inspection object. 3. The method according to claim 1, wherein when the teacher data is composed of feature amount data of a model object including only non-defective products, the identification unit identifies whether the inspection object is good or not. Alternatively, the abnormal noise inspection device according to claim 2. 4. When the teacher data is composed of characteristic amount data of each model object of a good product and a plurality of types of defective products, the classifying means determines good and bad types of the object to be inspected. The abnormal noise inspection device according to claim 1, wherein the device is identified. 5. The calculation unit according to claim 1, wherein the calculation unit obtains a plurality of feature amount data per unit time by performing a time series analysis on the signal extracted from the detection unit. Abnormal noise inspection equipment. 6. The apparatus according to claim 1, wherein the calculating means analyzes the frequency of the signal extracted from the detecting means and obtains characteristic amount data having an amplitude for each unit frequency. Abnormal noise inspection device as described. 7. The method according to claim 1, wherein the calculating means performs a time-frequency analysis on the signal extracted from the detecting means, and obtains characteristic amount data including an amplitude of a unit frequency component for each unit time. 5. The abnormal noise inspection device according to any one of 4.