JP3340541B2 - Inspection device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an inspection apparatus for inspecting the presence or absence of an abnormality in an object based on vibration of the object or a sound emitted from the object.
The present invention relates to an inspection apparatus suitable for abnormal sound / vibration inspection of a relatively small rotating body such as about 200 W).

[0002]

2. Description of the Related Art Since electric motors for automobiles and motors for home electric appliances are used close to end users, strict control of abnormal noise and vibration is required. Heretofore, this type of inspection has been mostly performed by a sensory test of a skilled sensual person. In recent years, F
There are inspection devices that monitor a specific frequency part by combining an FT analyzer and a comparator function. However, if the cause of the abnormality is wide-ranging, it is necessary to determine a specific frequency for each cause and adjust the level. It was difficult to match the sensory test of humans, for example, if the level slightly exceeded the comparator level, it was determined to be NG.

[0003] A relatively large number of motor abnormality causes are sudden sound abnormalities caused by the incorporation of chips, dust and the like. Sudden sound abnormalities do not occur constantly, such as often appearing only at the time of startup, and therefore, it has been difficult to detect them using the conventional method. Further, in the case of a geared motor, the levels of sound and vibration are largely different between forward rotation and reverse rotation in many cases, and it is often not possible to judge well by the level adjustment of a single comparator.

[0004]

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and has as its object to provide an inspection apparatus capable of more accurately inspecting for the presence or absence of an abnormality as compared with the related art.

[0005]

According to a first aspect of the present invention, there is provided an inspection apparatus for inspecting the presence or absence of an abnormality in a subject based on vibration of the subject or sound emitted from the subject. (1) A sensor that obtains a detection signal by detecting a vibration of a subject or a sound emitted from the subject (2) An envelope signal representing an envelope of the detection signal is generated by detecting an envelope of the detection signal. Envelope detection unit (3) Fourier transform unit that generates a Fourier transform signal by Fourier-transforming the detection signal and / or the envelope signal. (4) (a) The Fourier transform signal is defined for each frequency. First calculating means for obtaining a certainty of the presence or absence of abnormality of the subject by comparing with a predetermined threshold value, a crest factor of a detection signal and / or an envelope signal, and / or
A second calculating means for obtaining a kurtosis, obtaining a certainty degree of the presence or absence of an abnormality of the subject based on the crest factor and / or kurtosis, and obtaining an overall value of the Fourier transform signal, based on the overall value; And (b) a Fourier transform signal or a Fourier transform signal generated based on the Fourier transform signal, the computing unit comprising at least one selected from a group consisting of third computing means for determining the degree of certainty as to whether or not the subject has an abnormality. A neural network for inputting a feature vector signal representing the feature of the transformed signal and obtaining the certainty of presence / absence of an abnormality in the subject; and (c) a fuzzy calculation based on the certainty obtained by both the arithmetic unit and the neural network. A fuzzy judgment unit for judging the presence / absence of an abnormality; and a judgment processing unit having (a) to (c). It is intended to.

[0006]

According to the present invention, (i) a combination of the judgment by the neural network and the judgment by the comparator, (ii) the judgment by the neural network and the judgment by the crest factor (crest factor) and / or the cretosis (peakness) (Iii) Combination of determination by neural network and determination by overall value At least one of the three combinations, and a fuzzy determination unit that combines them and performs a comprehensive determination by fuzzy operation It is characterized by.

Hereinafter, each combination of the above (i) to (iii) will be described. (Combination of neural network and comparator)
The neural network has a feature that, when the relative level of normal / abnormal is unknown, learning of the normal / abnormal work can automatically learn the level. Further, for example, it is difficult to determine the balance between the rotational order of the motor and its harmonics, and to identify a partial pattern in a specific frequency band by a comparator, but it can be easily determined by a neural network.

However, since the neural network learns and recognizes the entire waveform pattern, it is not suitable for determining only the difference in the level of a specific frequency. In the case where a difference in the level of a specific frequency is a problem, such as gear meshing vibration, there may be a case where the neural network alone cannot determine the level. A system configuration for determining abnormality can be provided.

(Combination of Neural Network and Cress Factor, Kurtosis) The i-th measured value on the time axis is X _i , the maximum value among the measured values is X _max , the effective value of the measured values is X _rms , and the total number of measured values is C assuming N
R is the following formula

[0010]

(Equation 1)

Is defined by When the i-th measurement value on the time axis is X _i , the average value of the measurement values is X, the standard deviation of the measurement values is σ, and the total number of measurements is N, Kurtosis μ ⁴
Is the following formula

[0012]

(Equation 2)

[0013] neural network,
For both comparators, the determination is for data in the frequency domain. Waveform data in the frequency domain is added and averaged to stabilize the signal.However, the sudden sound due to dust inclusion may be output only once or twice during the measurement period, and the averaged data may be buried. There are many. By measuring and judging kurtosis and crest factor from data in the time domain where averaging is not performed,
Can handle vibration.

(Combination of Neural Network and Overall Value) Assuming that the measured value on the time axis is X _i and the total number of measurements is N, the overall value OA is

[0015]

(Equation 3)

Is required. As described above, the neural network learns and recognizes the entire waveform pattern. At this time, most are input as a waveform pattern standardized by pre-processing (normalized by setting the power of the frequency having the maximum power to 1 or normalized by an overall value). So the difference in power is:
It can be said that the discrimination function of the neural network is insensitive. Anomalies of the type that can be determined to be anomalous when the power differs greatly from the learned ones are often indistinguishable by normalization.In such a case, a threshold for the overall value should be set. In addition, by comparing and determining the overall value and the threshold value, it is possible to determine that an abnormality caused by a difference in power is also abnormal.

(A Combination of Any of the Above and Fuzzy Calculation) In the present invention described above, the determination result of each of the above items (neural network, comparator, crest factor, kurtosis, and overall value) is obtained as a certainty factor. In the present invention, a specific calculation method of the certainty factor is arbitrary, but is determined so as to reflect a degree considered to be normal and a degree considered to be abnormal.

The present invention provides at least one of the above-mentioned combinations.
This is a combination of a fuzzy calculation and a certainty. Comprehensive determination of the certainty calculated for each item by the fuzzy calculation enables more reliable determination. That is, with the certainty factor of each item, it is possible to eliminate the inconvenience of abnormal determination due to a slight over-setting of the specific determination item. In addition, if an arbitrary combination of determination items corresponding to individual inspection lines is allowed, flexible determination closer to a sensory test can be performed.

[0019]

Embodiments of the present invention will be described below. FIG. 1 is a system block diagram of an embodiment of the inspection apparatus of the present invention, and FIG. 2 is a block diagram showing a configuration of a determination processing unit shown by a block in FIG. [Configuration of this system] In this system, a vibration acceleration detector is used for signal input, and a signal (detection signal) obtained by passing the signal from the vibration acceleration detector through an amplifier and a high-pass filter (Equation 1)
The signal (envelope signal) having passed through the envelope detector and the envelope detector is input to one or a plurality of FFT analyzers (converters). The two-channel waveforms each converted into a power spectrum, the detection signal output from the amplifier before conversion into the power spectrum, and the envelope signal output from the envelope detector are transferred to the determination processing unit. In this embodiment, a personal computer is used as the determination processing unit.

The personal computer has a learning / recognition function for identifying the power spectrum of the two channels as a waveform, a comparator function for comparing each frequency of the power spectrum, a function of determining the overall value of the two channels, and a time domain. It has a crest factor and kurtosis determination function from data. The level comparator and overall value for each frequency are
Since the judgment weights are different, the judgment can be made independently.

The learning / recognition function has a pattern learning / recognition function using a neural network. The discrimination function of the neural network and the discrimination function by the comparator are configured to perform a plurality of processes under different condition settings. Each recognition and determination unit outputs independent normal / abnormal determination and certainty factor (10 levels), and the overall determination unit comprehensively determines each result by fuzzy OR.

Each recognition / judgment unit has a judgment condition (neural network learning content, comparator judgment level, etc.) for the normal rotation and the reverse rotation, respectively, and inputs a normal rotation and a reverse rotation signal as external signals at the time of judgment. Thus, the determination is made by selecting the individual determination conditions for the forward rotation and the reverse rotation. A vibration acceleration detector is attached near the bearing of the target rotating device (the motor in FIG. 1), and the vibration acceleration signal and the envelope detection signal are input to the FFT analyzer. The frequency analysis waveform subjected to the addition processing or the peak processing of both signals is taken into the determination processing unit as one waveform pattern. In the determination processing section, first, in the data preprocessing section (see FIG. 2), the waveform pattern is subjected to preprocessing such as smoothing, normalization, and compression as necessary, and then waveform recognition is performed using a neural network. Then, a normal / abnormal judgment and a certainty factor are calculated. The calculation of the certainty factor of the determination by the neural network will be described later.

Further, the overall value of the detection signal output from each amplifier of the above two signals and the overall value of the envelope signal output from the envelope detector are obtained, and the crest factor and kurtosis value are calculated from the envelope signal. Then, the normal / abnormal judgment is performed individually by the level judgment with the respective abnormality judgment values. In the determination unit by the comparator, for example, when mainly discriminating the gear meshing sound, since the level comparison function of the specific frequency may be important, the level of the frequency analysis waveform of both signals with respect to an arbitrary frequency is determined. It also has a comparison function.

The method of determining the level of each value and the method of calculating the degree of certainty are as follows. Although omitted in the description, each level is set independently for normal rotation and reverse rotation. If the measured value is X, the abnormality determination level is NGlevel, the certainty calculation range setting value is NGmax, and the certainty is CF, if X
Is equal to or larger than NGlevel, it is determined to be NG.

If X is smaller than NGlevel, it is OK. The confidence calculation is as follows. CF = (X−NGlevel) / (NGlevel + NGmax) * 10.0 (4) That is, the CF value is + when the abnormality is normal, and − when the normality is normal. . The certainty factor calculation range setting value is set in advance as an empirical value.

Next, the result of recognition by the neural network and the result of level comparison of each value are comprehensively determined by fuzzy OR. Next, the calculation in each recognition / determination unit will be described. (Calculation by Neural Network) FIG. 3 is a schematic configuration diagram of a hierarchical neural network.

It is assumed that two neurons in the output layer of the neural network correspond to normal (OK) and abnormal (NG), respectively. When the measurement data is presented to the input side, if only one output neuron exceeds a preset threshold or if multiple neurons correspond to the same normal / abnormal judgment, the normal / abnormal judgment is made. It is output with a certainty factor of 10.

If there is no neuron exceeding the threshold value, the output is unknown, that is, the confidence level is 0. When there are two or more neurons exceeding the threshold value and the two neurons with higher firing intensities correspond to different judgments (one is normal judgment and the other is abnormal judgment), the confidence is obtained by the following calculation. .

The output of the neuron corresponding to the abnormality is NN1,
Assuming that the output of the normal neuron is NN2, the maximum value of the output difference of the neuron is Dmax, and the confidence is CF, CF = (NN1-NN2) /Dmax*10.0 (5) That is, in the case of an abnormality, If the certainty factor CF value is + and the confidence value is normal, the CF value assumes a negative value. Therefore,
The possible range of the CF value is from -10.0 to +10.0. In addition, an empirical value is set as the Dmax value in advance.

(Comparator operation) The measured value of the power at each frequency is Xf, the set power level at each frequency is Xfn, the measured value of the power at the frequency at which (Xf-NGlevel) is the maximum value, the confidence operation range Set value to NGma
Assuming that x and the certainty are CF, the certainty calculation is obtained by the following equation.

CF = (Xfn−NGlevel) /NGmax*10.0 (6) That is, in the case of NG, the certainty factor CF value takes a value of +, and in the case of OK, the CF value takes a value of −. I do. NGma
As the x value, an empirical value is set in advance. (Crest factor, kurtosis, overall value calculation) The crest factor CR, kurtosis μ ⁴ , and overall value OA are (1), (2),
It is determined by equation (3). Those crest factors C
The degree of certainty based on R, kurtosis μ ⁴ , and overall value OA is calculated from the value obtained by the equations (1) to (3) as a measured value X
Is obtained by equation (4).

(Fuzzy OR operation method) FIGS. 4 and 5
Are examples of antecedent / consequent part membership functions, respectively.
FIG. 19 is a diagram illustrating another example of the consequent part membership function. Each membership function is set, for example, as shown in FIG. The membership function of the consequent part can be set independently for each item.

For example, the judgment certainty (abnormal certainty) of the neural network is NN, and the judgment certainty of the comparator is CMP.
And the following rule operation is performed. When the number of judgment items increases, the calculation is performed by increasing the rules similar to (1) to (7). A fuzzy-or operation is performed on the output calculated according to each rule to obtain a non-fuzzy output. The fuzzy operation used is the MAX-MIN method, MAX-DO
Any T method is acceptable.

(1) If the NG certainty factor of the NN is NB, the total certainty factor is also NB. (2) If the NG certainty factor of the NN is NM, the total certainty factor is also NM. (3) If the NG confidence of the NN is NS, the overall confidence is also NS. (4) If the NG certainty factor of the NN is ZO, the total certainty factor is also ZO. (5) If the NG certainty factor of the NN is PS, the total certainty factor is also PS.

(6) If the NG certainty factor of the NN is PM, the total certainty factor is also PM. (7) If the NG confidence of the NN is PB, the overall confidence is also PB. (8) If the NG certainty factor of the CMP is NB, the total certainty factor is also NB. (9) If the NG confidence of CMP is NM, the overall confidence is also NM. (10) If the NG certainty factor of the CMP is NS, the total certainty factor is also NS.

(11) If the NG confidence of CMP is ZO
Then, the total certainty is also ZO. (12) If the NG certainty factor of the CMP is PS, the total certainty factor is also PS. (13) If the NG certainty factor of the CMP is PM, the total certainty factor is also PM. (14) If the NG certainty factor of the CMP is PB, the total certainty factor is also PB. Here, for simplicity, the above two rules (1), (1)
Considering the case consisting only of 0), a fuzzy OR is output as follows. As the membership function of the antecedent part (condition part) / consequent part (conclusion part), one having a shape as shown in FIG. 3 is employed. As the membership function of the condition part, an independent function can be used for each item (here, a neural network and a comparator). The membership function of the consequent part may have a shape as shown in FIG. 4 or a singleton having no distribution as shown in FIG. Here, the description will be made using the MAX-MIN method.

Each of the inputted abnormal certainty factors is represented by NN
(Abnormal confidence calculated by the neural network) and CMP (abnormal confidence calculated by the comparator). From the rule (1), a height α that is compatible with the membership function (NB) of the antecedent part is obtained as shown in FIG.
The membership function (NB) of the consequent part is truncated based on the height (degree of conformity) α.

Similarly, from the rule (10), a height β that is compatible with the membership function (NS) of the antecedent is obtained,
The membership function (NS) of the consequent part is truncated based on the height (fitness) β. The maximum value combination (OR set) of the truncated membership functions of each consequent part is obtained as shown in FIG. 8, and this is defined as a function f (x). The center of gravity of f (x) is calculated as in the following equation (7), and the total confidence CF
Ask for.

[0039]

(Equation 4)

The system according to the above embodiment has the following effects. (1) By recognizing the power spectrum of the two channels of the vibration detection signal and the envelope signal as one pattern, it is possible to determine the intermittent sound that sounds like noise and the continuous sound that sounds like noise. Since the intermittent allophone and the continuous allophone have different frequencies, use two FFT analyzers with frequency ranges suitable for each analysis.
Recognition as a two-channel power spectrum pattern enables recognition of both abnormal sounds. For example,
The continuous sound often has a pattern characteristic of the power of several kHz of the vibration acquisition signal, and the intermittent sound often has a characteristic in the waveform of the low-frequency part and the waveform of the high-frequency part of the envelope signal.

(2) Since the above function takes a waveform as a pattern, it is relatively insensitive to a minute difference in a specific frequency. Further, when the number of abnormal samples is small, learning cannot be converged, and pattern recognition alone may not be sufficient. On the other hand, it may be important to monitor the level of a particular line, such as a gear meshing abnormality.
In the above embodiment, a power spectrum comparator function is additionally provided, so that abnormality detection can be performed by comparing the level of a specific frequency.

(3) Since an overall value determination function is also provided, even when the learned work and the sound / vibration level are extremely different, it is possible to more reliably determine an abnormality by determining the overall value. (4) In the case of a sudden sound abnormality due to dust mixing or the like, since it is difficult to detect by the functions of the items (1) and (2) on the averaged spectrum waveform, the sudden sound abnormality is determined by the crest factor or the like. By including the value of Kurtsis (fourth moment) in the determination, the abnormality can be determined.

(5) By performing a comprehensive judgment by a logical OR operation using fuzzy inference in which the judgment results of the above functions are weighted, a subtle abnormality judgment of a product can be made closer to a sensory test. For example, if the value of the comparator or each item slightly exceeds NG
NG determination can be made based on the balance with other NG factors.

(6) By having the learning contents and judgment conditions for each of the normal rotation and the reverse rotation, detailed setting and judgment can be made for each rotation direction.

[0045]

As described above, according to the inspection apparatus of the present invention, it is possible to more accurately inspect an object to be inspected for an abnormality.

[Brief description of the drawings]

FIG. 1 is a system block diagram of an embodiment of an inspection apparatus according to the present invention.

FIG. 2 is a block diagram illustrating a configuration of a determination processing unit indicated by blocks in FIG. 1;

FIG. 3 is a schematic configuration diagram of a hierarchical neural network.

FIG. 4 is a diagram showing an example of an antecedent / consequent part menzip function.

FIG. 5 is a diagram showing another example of the consequent part membership function.

FIG. 6 is a diagram showing a calculation method when shifting from the antecedent part to the consequent part.

FIG. 7 is a diagram illustrating a calculation method when shifting from the antecedent part to the consequent part.

FIG. 8 is a diagram showing a method of obtaining a total certainty factor.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-5-72026 (JP, A) JP-A-6-135395 (JP, A) JP-A-5-52644 (JP, A) JP-A-58-58 219424 (JP, A) JP-A-1-152335 (JP, A) JP-A-62-237339 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G01H 17/00 G01M 19 / 00

Claims (1)

(57) [Claims] An inspection apparatus for inspecting the presence or absence of an abnormality in a subject based on vibration of the subject or sound emitted from the subject, wherein the testing apparatus detects the vibration of the subject or sound emitted from the subject. An envelope detector that generates an envelope signal representing an envelope of the detection signal by performing an envelope detection of the detection signal, and detects the detection signal and / or the envelope signal. Fourier transform unit that generates a Fourier transform signal by performing Fourier transform, and the certainty degree of the presence or absence of abnormality of the subject by comparing the Fourier transform signal with each predetermined threshold value determined for each frequency. First calculating means for obtaining the crest factor and / or kurtosis of the detection signal and / or the envelope signal, and obtaining the crest factor and / or Second computing means for determining the degree of certainty regarding the presence or absence of an abnormality in the subject based on Kurtosis, and determining an overall value of the Fourier transform signal, and determining whether or not there is an abnormality in the subject based on the overall value. A computing unit consisting of at least one selected from a group consisting of third computing means for obtaining a certainty factor, and a characteristic representing the Fourier transform signal or a feature of the Fourier transform signal generated based on the Fourier transform signal A neural network that receives a vector signal to determine the degree of certainty regarding the presence or absence of an abnormality in the subject, and determines the presence or absence of an abnormality in the subject by fuzzy arithmetic based on the certainty determined by both the arithmetic unit and the neural network An inspection apparatus comprising: a determination processing unit having a fuzzy determination unit that performs the determination.