JP6620290B2 - Vibro-acoustic analysis method and apparatus, and apparatus abnormal part estimation method and apparatus - Google Patents

Description

  The present invention relates to a vibroacoustic analysis method and apparatus, and an apparatus abnormal part estimation method and apparatus.

  In general, in a device including a rotating body such as an engine, in order to detect an abnormality, vibration and sound including distortion are measured and time series data is taken.

  In the analysis of the time series data, spectrum analysis using Fourier transform or the like is often performed. In simple spectral analysis, information in the time domain may be lost. Further, in a simple spectrum analysis, if the frequency distribution included in the measurement data is wide, it may be difficult to obtain an appropriate spectrum analysis.

  On the other hand, a method using wavelet transform has been proposed. In the wavelet transform, information on the time domain lost when the frequency characteristic is obtained by the Fourier transform can be left, and the analysis is performed by decomposing the information into the time domain and the frequency domain.

  For example, Patent Document 1 shows a general technical level related to the technique using the wavelet transform.

JP 2010-48684 A

  However, equipment including a rotating body such as an engine is composed of a large number of parts such as a bearing part, a gear part, and a power generation part, and information from each part is mixed in the measured vibroacoustic data. Appropriate analysis results were sometimes shown. In addition, since the output of the device is increased or decreased depending on the usage state, and the magnitude of the vibrational sound changes significantly, it is appropriate in terms of accuracy to estimate the abnormal part only by the magnitude of the measured vibrational sound data. It could not be said that it was a correct response.

Incidentally, in a relatively low speed bearing having a rotational speed of about several hundred [min ⁻¹ ], a large component appears in a high frequency band of several [kHz] to several tens [kHz] in the vibration acoustic data. Many. This is because vibroacoustic signals due to scratches or the like often vibrate around the bearing portion at its natural frequency due to a pulsed vibration source, and are mainly composed of high-frequency components. For this reason, even if the measurement signal is subjected to spectrum analysis, the peak frequency and amplitude in the low frequency region near the rotation speed may not be appropriately extracted. As a countermeasure, for example, it is common to extract a peak frequency and amplitude of a low cycle by performing spectrum analysis after performing envelope processing using a low-pass filter or the like. Although this method is suitable for analysis in the low frequency region, it cannot analyze signals in the high frequency region. In addition, the peak frequency and amplitude may not be correct values due to the performance and characteristics of the low-pass filter, making it difficult to use for evaluation.

  On the other hand, the following has been found by the study of the present inventors. That is, the measured time-series data is decomposed into a plurality of vectors in the time domain and the frequency domain by wavelet transform, and then the spectrum analysis is performed on the time-series vectors in each frequency band, so that the low frequency domain It is possible to appropriately extract the peak frequency and amplitude in a wide range from the high frequency region to the high frequency region.

  However, since one time-series data is decomposed into a large number of time-series data by performing analysis by wavelet transform, it is necessary for humans to judge this analysis result and it is practical. I could not say.

  The present invention has been made in view of the above-described conventional problems, and can perform vibroacoustic analysis including distortion and estimation of an abnormal part of a device accurately and easily by wavelet transformation, and early detection of an abnormal part. It is an object of the present invention to provide a vibroacoustic analysis method and apparatus, and an apparatus abnormal part estimation method and apparatus that can realize secondary damage prevention and planned maintenance based on remaining life prediction.

In order to achieve the above object, the vibroacoustic analysis method of the present invention includes a feature quantity extraction processing step for extracting feature quantities from time-series data of vibroacoustics measured in an operating device,
The feature amount extracted in the feature amount extraction processing step and the feature amount when the device is operating normally and the general measurement data of the device are accumulated, and the feature amount when the device is operating normally by machine learning Build a probability distribution model related to the correlation with the measurement data, give the general distribution data of the device and the feature quantity extracted in the feature quantity extraction processing step to the probability distribution model, and anomaly based on the probability distribution model A machine learning diagnosis step for calculating the degree.

In the vibroacoustic analysis method, the feature amount extraction processing step includes:
Basic features that extract effective values, peak values, maximum peak differences, skewness, kurtosis, average crest factor, and absolute average amplitude as feature values from the measured acceleration of vibro-acoustic time-series data measured on the equipment in operation An extraction process,
The maximum peak order and its peak value, the second peak order and its peak value, the third peak order and its peak value, the fourth peak order and its peak value as feature quantities after spectral analysis of the time series data, ... A spectral analysis feature extraction process step for extracting the nth peak order and its peak value;
A wavelet transform of the time series data and decomposed into scale levels corresponding to a plurality of frequency bands, and a continuous wavelet transform processing step for obtaining power time series data for each scale level;
A wavelet transform feature quantity extraction process step for extracting an effective value, maximum value, maximum crest factor, and absolute average value as feature quantities of each frequency band from the power time-series data for each scale level obtained in the continuous wavelet transform process step; ,
The maximum peak order and its peak value, the second peak order and its peak value, the third peak order and its as feature quantities after spectral analysis of the power time series data for each scale level obtained in the continuous wavelet transform processing step A peak value, a fourth peak order and its peak value, a wavelet transform spectrum analysis feature extraction process step for extracting the nth peak order and its peak value.

  In the vibroacoustic analysis method, the general measurement data of the device may be at least one of a rotational speed, an output, a work amount, a pressure, a temperature, and a speed.

On the other hand, the vibroacoustic analysis apparatus of the present invention includes a feature amount extraction processing unit that extracts feature amounts from time-series data of vibroacoustics measured in a device in operation,
The feature amount extracted in the feature amount extraction processing unit and the feature amount in a state where the device is operating normally and the general measurement data of the device are accumulated, and the feature amount and the general state in the device normal operation state by machine learning Construct a probability distribution model related to the correlation with the measurement data, give the general measurement data of the device and the feature quantity extracted by the feature quantity extraction processing unit to the probability distribution model, and anomaly based on the probability distribution model A machine learning diagnostic unit for calculating the degree.

In the vibroacoustic analysis apparatus, the feature amount extraction processing unit includes:
Basic features that extract effective values, peak values, maximum peak differences, skewness, kurtosis, average crest factor, and absolute average amplitude as feature values from the measured acceleration of vibro-acoustic time-series data measured on the equipment in operation An extraction processing unit;
The maximum peak order and its peak value, the second peak order and its peak value, the third peak order and its peak value, the fourth peak order and its peak value as feature quantities after spectral analysis of the time series data, ... A spectral analysis feature amount extraction processing unit for extracting the n-th peak order and its peak value;
A continuous wavelet transform processing unit for wavelet transforming the time series data and decomposing it into scale levels corresponding to a plurality of frequency bands, and obtaining power time series data for each scale level;
A wavelet transform feature amount extraction processing unit that extracts an effective value, a maximum value, a maximum crest factor, and an absolute average value as feature amounts of each frequency band from power time-series data for each scale level obtained by the continuous wavelet transform processing unit; ,
The maximum peak order and its peak value, the second peak order and its peak value, the third peak order and its feature as a feature after spectral analysis of the power time-series data for each scale level obtained by the continuous wavelet transform processing unit A peak value, a fourth peak order and its peak value,..., An n-th peak order and its peak value, and a wavelet transform spectrum analysis feature quantity extraction processing unit.

  In the vibroacoustic analysis apparatus, the general measurement data of the device may be at least one of a rotational speed, an output, a work amount, a pressure, a temperature, and a speed.

Furthermore, in the machine learning diagnostic step of the vibroacoustic analysis method, the present invention pre-calculates a data set at the time of damage for each part of the device at the time of damage,
The degree of similarity for each part of the device by comparing the data set at the time of damage calculated in advance in the machine learning diagnosis step with the actual measurement data set of the degree of abnormality calculated from the general measurement data of the device Thus, the device abnormal part estimation method including the abnormal part estimation step of outputting as a part having a high possibility of damage in the descending order of the degree of similarity for each part of the equipment can be obtained.

Further, the present invention is configured such that the machine learning diagnostic unit of the vibroacoustic analysis apparatus calculates in advance a data set at the time of damage for each part of the device at the time of damage,
The degree of similarity for each part of the device by comparing the data set at the time of damage calculated in advance by the machine learning diagnostic unit with the actual measurement data set of the degree of abnormality calculated from the general measurement data of the device Thus, the device abnormal part estimation device including the abnormal part estimation unit that outputs the parts with high possibility of damage in descending order of the degree of similarity for each part of the equipment can be obtained.

  According to the vibroacoustic analysis method and apparatus and the apparatus abnormal part estimation method and apparatus of the present invention, the vibroacoustic analysis and the estimation of the abnormal part of the apparatus can be performed accurately and easily by wavelet transform. It is possible to achieve an excellent effect of preventing secondary damage due to discovery and realizing planned maintenance by predicting the remaining life.

It is a flowchart which shows the Example of the vibration acoustic analysis method and apparatus of this invention, and an apparatus abnormal region estimation method and apparatus. It is a flowchart which shows the detail of the feature-value extraction process in the Example of the vibroacoustic-analysis method and apparatus of this invention, and an apparatus abnormal region estimation method and apparatus. It is a block diagram in the Example of the vibration acoustic analysis method and apparatus of this invention, and an apparatus abnormal region estimation method and apparatus. It is a diagram which shows the example of a waveform analyzed by the example of the waveform of the measurement signal in the Example of the vibroacoustic analysis method and apparatus of this invention, an apparatus abnormal region estimation method, and an apparatus, and a wavelet transform. It is a diagram which shows the pattern at the time of the abnormality of the site | part in the Example of the vibration acoustic analysis method and apparatus of this invention, and an apparatus abnormal site | part estimation method and apparatus.

  Embodiments of the present invention will be described below with reference to the accompanying drawings.

  FIG. 1 to FIG. 5 show an embodiment of the vibroacoustic analysis method and apparatus and the apparatus abnormal part estimation method and apparatus of the present invention.

  In the case of the vibroacoustic analysis method of the present embodiment, as shown in FIG. 1, it has a feature amount extraction processing step and a machine learning diagnosis step.

  In the feature quantity extraction processing step, the feature quantity is extracted from the time-series data of the vibration and acoustics measured in the operating device.

  In the machine learning diagnosis step, the feature amount extracted in the feature amount extraction processing step and the device is operating normally and general measurement data of the device (for example, rotation speed, output, work amount, pressure) , Temperature, speed) and build a probability distribution model related to the correlation between the feature quantity in the normal operation state of the device and the general measurement data by machine learning, and the general measurement data of the device and the probability distribution model The feature amount extracted in the feature amount extraction processing step is given, and the degree of abnormality is calculated based on the probability distribution model.

  In the example of FIG. 1, measurement is started as step S10, and in the case of multipoint measurement, measurement is performed simultaneously. In step S11, time-series data for a certain time is input, and in step S12, the measurement is stopped.

  Thereafter, as the steps S13 and S14, the feature amount extraction processing step and the machine learning diagnosis step are performed.

  The degree of abnormality calculated in the machine learning diagnosis step is stored and accumulated as step S15.

  After storing and accumulating the degree of abnormality in step S15, it is determined in step S16 whether or not the set measurement interval time has passed. If the set measurement interval time has not passed, the process waits for a predetermined time as step S17. Returning to S16, if the set measurement interval time has elapsed, the process returns to step S10 to repeat the measurement.

  The degree of abnormality calculated in the machine learning diagnosis step of step S14 is compared with a normal range defined as step S18. If the degree of abnormality is outside the normal range, an alarm is issued as step S19, As step S20, an abnormal site estimation step is performed. The abnormal site estimation step is a step performed in the device abnormal site estimation method.

  More specifically, the feature quantity extraction processing step uses time-series data for a predetermined time input as measurement data in step S11, and as shown in FIG. 2, a basic feature quantity extraction processing step as step S100, Spectrum analysis feature quantity extraction process step as step S200, Continuous wavelet transform process step as step S300, Wavelet transform feature quantity extraction process step as step S400, Wavelet transform spectrum analysis feature quantity extraction process step as step S500 And have.

  Incidentally, when the number of measurement parts of the device is n, there are n pieces of time series data input as measurement data in step S11 corresponding to the number of measurement parts of the device.

ピーク平均値ＭＰ２、

（但し、ピークは、極大値、極小値の両方）

最大ピーク差ＭＰ３、

（但し、x_+p：サンプル期間中の最大値、x_-p：サンプル期間中の最小値）

歪度ＭＰ４、

但し、σ：標準偏差

尖度ＭＰ５、

平均波高率ＭＰ６、

（但し、［波高率(Crest factor)］＝［ピーク値］／［実効値］であるが、ピーク値はピーク平均値ＭＰ２を取っている。）

絶対平均振幅ＭＰ７

を抽出するようになっている（ステップＳ１１０参照）。尚、前記ステップＳ１００の基本特徴量抽出処理工程においては、ステップＳ１１０で抽出された特徴量を、ステップＳ１２０として定義済みの正常範囲と比較し、前記特徴量が正常範囲外である場合には、ステップＳ１３０として警報を発令するようになっている。 The basic feature amount extraction processing step includes an effective value MP1, as a feature amount, from a measured acceleration of time-series data of vibroacoustics measured in a device in operation.

Peak average value MP2,

(However, the peak is both maximum and minimum)

Maximum peak difference MP3,

(Where x _{+ p is} the maximum value during the sample period, x _{-p is} the minimum value during the sample period)

Skewness MP4,

Where σ: standard deviation

Kurtosis MP5,

Average crest factor MP6,

(However, [Crest factor] = [Peak value] / [Effective value], but the peak value is the peak average value MP2.)

Absolute average amplitude MP7

Are extracted (see step S110). In the basic feature amount extraction process in step S100, the feature amount extracted in step S110 is compared with the normal range defined as step S120. If the feature amount is outside the normal range, In step S130, an alarm is issued.

  The spectral analysis feature quantity extraction processing step includes a maximum peak order and its peak value, a second peak order and its peak value, a third peak order and its peak value, as feature quantities after spectral analysis of the time series data. The four peak orders and their peak values,... Nth peak order and their peak values are extracted (see step S210). In the spectrum analysis feature quantity extraction process in step S200, the feature quantity extracted in step S210 is compared with the normal range defined as step S220. If the feature quantity is outside the normal range, In step S230, an alarm is issued.

  In the continuous wavelet transform processing step, the time-series data is subjected to wavelet transform and decomposed into scale levels corresponding to a plurality of (for example, m) frequency bands to obtain power time-series data for each scale level. (See step S310).

最大値ＷＰ２、

最大波高率ＷＰ３、

絶対平均値ＷＰ４

を抽出するようになっている（ステップＳ４１０参照）。尚、前記ステップＳ４００のウェーブレット変換特徴量抽出処理工程においては、ステップＳ４１０で抽出された特徴量を、ステップＳ４２０として定義済みの正常範囲と比較し、前記特徴量が正常範囲外である場合には、ステップＳ４３０として警報を発令するようになっている。 The wavelet transform feature amount extraction processing step includes an effective value WP1 as a feature amount of each frequency band from the power time series data for each scale level obtained in the continuous wavelet transform processing step.

Maximum value WP2,

Maximum crest factor WP3,

Absolute average value WP4

Are extracted (see step S410). In the wavelet transform feature quantity extraction process in step S400, the feature quantity extracted in step S410 is compared with the normal range defined as step S420, and the feature quantity is outside the normal range. In step S430, an alarm is issued.

  The wavelet transform spectrum analysis feature amount extraction processing step includes a maximum peak order and its peak value, a second peak as a feature amount after spectrum analysis of power time series data for each scale level obtained in the continuous wavelet transform processing step. The order and its peak value, the third peak order and its peak value, the fourth peak order and its peak value,... Nth peak order and its peak value are extracted (see step S510). In the wavelet transform spectrum analysis feature quantity extraction process in step S500, the feature quantity extracted in step S510 is compared with the normal range defined as step S520, and the feature quantity is outside the normal range. In step S530, an alarm is issued.

  The feature amount extracted in the basic feature amount extraction processing step (see step S110), the feature amount extracted in the spectrum analysis feature amount extraction processing step (see step S210), and the wavelet transform feature amount extraction processing step The extracted feature quantity (see step S410), the feature quantity extracted in the wavelet transform spectrum analysis feature quantity extraction processing step (see step S510), and general measurement data of the device measured as step S600 (for example, rotation) Number, output, work, pressure, temperature, speed) are stored and accumulated as step S700.

  The abnormal part estimation step compares the data set of the degree of abnormality calculated in advance in the machine learning diagnosis step with the actual measurement data set of the degree of abnormality calculated from the general measurement data of the device. The degree of similarity for each part of the device is obtained and output as the part having a high possibility of damage in descending order of the degree of similarity for each part of the device.

  On the other hand, FIG. 3 is a block diagram in the present embodiment, which includes a feature quantity extraction processing unit 10 and a machine learning diagnosis unit 20 as a vibroacoustic analysis device, and an apparatus abnormal site estimation device as the vibroacoustic analysis device. A feature amount extraction processing unit 10, a machine learning diagnosis unit 20, and an abnormal part estimation unit 30 are provided.

  The feature quantity extraction processing unit 10 is configured to extract feature quantities from time-series data of vibroacoustics measured in a device in operation.

  The machine learning diagnosis unit 20 accumulates the feature amount and the general measurement data of the device that are extracted by the feature amount extraction processing unit 10 and the device is operating normally. A probability distribution model relating to the correlation between the feature quantity in the normal operating state and the general measurement data is constructed, and the general measurement data of the device and the feature quantity extracted by the feature quantity extraction processing unit 10 are added to the probability distribution model. Given, the degree of abnormality is calculated based on the probability distribution model.

  The feature amount extraction processing unit 10 includes a basic feature amount extraction processing unit 11, a spectrum analysis feature amount extraction processing unit 12, a continuous wavelet transform processing unit 13, a wavelet transform feature amount extraction processing unit 14, and a wavelet transform spectrum analysis. And a feature amount extraction processing unit 15.

  The basic feature amount extraction processing unit 11 has an effective value, a peak value, a maximum peak difference, a skewness, a kurtosis, and an average crest factor as a feature amount from the measured acceleration of time-series data of vibroacoustics measured in a driving device. The absolute average amplitude is extracted.

  The spectrum analysis feature amount extraction processing unit 12 has a maximum peak order and its peak value, a second peak order and its peak value, a third peak order and its peak value as a feature amount after spectral analysis of the time series data, The fourth peak order and its peak value,... Nth peak order and its peak value are extracted.

  The continuous wavelet transform processing unit 13 performs wavelet transform on the time series data, decomposes it into scale levels corresponding to a plurality of frequency bands, and obtains power time series data for each scale level.

  The wavelet transform feature quantity extraction processing unit 14 calculates the effective value, the maximum value, the maximum crest factor, and the absolute average as the feature quantity of each frequency band from the power time series data for each scale level obtained by the continuous wavelet transform processing unit 13. The value is extracted.

  The wavelet transform spectrum analysis feature quantity extraction processing unit 15 is configured to analyze the power time series data for each scale level obtained by the continuous wavelet transform processing unit 13 and analyze the maximum peak order and its peak value as a feature quantity. The second peak order and its peak value, the third peak order and its peak value, the fourth peak order and its peak value,... The n th peak order and its peak value are extracted.

  The abnormal site estimation unit 30 compares the abnormality degree data set calculated in advance by the machine learning diagnosis unit 20 with the actual measurement data set of abnormality degree calculated from the general measurement data of the device. The degree of similarity for each part of the device is obtained and output as a part having a high possibility of damage in the order of the degree of similarity for each part of the device.

  Next, the operation of the above embodiment will be described.

  In FIG. 1, first, measurement of each part of the device is started (see step S10), time series data is input for a certain time (see step S11), and measurement is stopped (see step S12). A feature amount extraction processing step is performed (see step S13), followed by a machine learning diagnosis step (see step S14).

  The degree of abnormality calculated in the machine learning diagnosis step is stored and accumulated (see step S15). Thereafter, it is determined whether or not the set measurement interval time has passed (see step S16), and the set measurement interval time is determined. If it has not elapsed, it waits for a fixed time (see step S17) and returns to step S16. If the set measurement interval time has elapsed, it returns to step S10 and measurement is repeated.

  The degree of abnormality calculated in the machine learning diagnosis step is compared with a defined normal range (see step S18). If the degree of abnormality is outside the normal range, an alarm is issued (see step S19). ), An abnormal site estimation step is performed (see step S20).

  In the feature quantity extraction processing step, time-series data for a predetermined time inputted as measurement data is used (see step S11), and as shown in FIG. 2, the basic feature quantity extraction processing step (see step S100), spectrum Analysis feature extraction process step (see step S200), continuous wavelet transform process step (see step S300), wavelet transform feature extraction process step (see step S400), wavelet transform spectrum analysis feature extraction process step (step (See S500).

  In the basic feature amount extraction processing step, an effective value MP1 (see [Equation 1]) and a peak average value MP2 ([Equation 2]) are measured as feature amounts from the measured acceleration of the vibration-sound time-series data measured in the device in operation. Reference), maximum peak difference MP3 (see [Equation 3]), skewness MP4 (see [Equation 4]), kurtosis MP5 (see [Equation 6]), average crest factor MP6 (see [Equation 7]), absolute The average amplitude MP7 (see [Equation 8]) is extracted (see step S110). The feature quantity extracted in the basic feature quantity extraction processing step is compared with a defined normal range (see step S120), and an alarm is issued when the feature quantity is outside the normal range ( (See step S130).

  In the spectral analysis feature quantity extraction processing step, the maximum peak order and its peak value, the second peak order and its peak value, the third peak order and its peak value, as the feature quantity after spectral analysis of the time series data, The four peak orders and their peak values,... Nth peak order and their peak values are extracted (see step S210). The feature amount extracted in the spectrum analysis feature amount extraction processing step is compared with a defined normal range (see step S220), and an alarm is issued if the feature amount is outside the normal range. (See step S230).

  In the continuous wavelet transform processing step, the time series data is wavelet transformed and decomposed into scale levels corresponding to a plurality of (for example, m) frequency bands, and power time series data for each scale level is obtained (step S310). reference).

Here, the waveform of the measurement signal and the waveform analyzed by wavelet transform in the present embodiment are as shown in FIG. 4, for example. In the example of FIG. 4, the scale parameter is set to 2 ^s (s = 0 to 7, that is, m = 8), and is decomposed every octave, and analysis is performed up to 8 scales every octave. However, the vertical axis of each waveform is not the same, and is appropriately adjusted according to the plot width. It should be noted that the decomposition method is arbitrary, and a scale parameter of 1 or less can be applied when it is desired to see a higher frequency component.

  In the wavelet transform feature quantity extraction process step, the effective value WP1 (see [Equation 9]) and the maximum value WP2 (see [Equation 9]) as the feature quantity of each frequency band from the power time series data for each scale level obtained in the continuous wavelet transform process step. The maximum crest factor WP3 (see [Expression 11]) and the absolute average value WP4 (see [Expression 12]) are extracted (see Step S410). The feature quantity extracted in the wavelet transform feature quantity extraction processing step is compared with a defined normal range (see step S420), and an alarm is issued when the feature quantity is outside the normal range. (See step S430).

  In the wavelet transform spectrum analysis feature quantity extraction process step, the maximum peak order and its peak value, the second peak as the feature quantity after spectrum analysis of the power time series data for each scale level obtained in the continuous wavelet transform process step The order and its peak value, the third peak order and its peak value, the fourth peak order and its peak value,... Nth peak order and its peak value are extracted (see step S510). Note that the feature amount extracted in the wavelet transform spectrum analysis feature amount extraction processing step in step S500 is compared with a defined normal range (see step S520), and when the feature amount is outside the normal range, An alarm is issued (see step S530).

  The feature amount extracted in the basic feature amount extraction processing step (see step S110), the feature amount extracted in the spectrum analysis feature amount extraction processing step (see step S210), and the wavelet transform feature amount extraction processing step The extracted feature quantity (see step S410), the feature quantity extracted in the wavelet transform spectrum analysis feature quantity extraction processing step (see step S510), and the number of revolutions, output, work, pressure, temperature, speed, for example The general measurement data (see step S600) of the devices such as these are stored and accumulated (see step S700). Here, the number of measurement parts of the device is n, but the general measurement data is one set for the device (one unit) to be diagnosed.

  In the abnormal part estimation step, the damage degree data set calculated in advance in the machine learning diagnosis step is compared with the actual measurement data set of the abnormality degree calculated from the general measurement data of the device. The degree of similarity for each part of the device is obtained, and the part is output as a part having a high possibility of damage in descending order of the degree of similarity for each part of the device.

  FIG. 5 is a diagram showing a pattern at the time of abnormality of a part in the present embodiment. In the diagram shown in FIG. 5, the horizontal axis represents the frequency level, the vertical axis represents the degree of abnormality, and the measurement positions “A”, “B”... The pattern at the time of abnormality and the pattern at the time of abnormality of the site | part B are displayed, and the site | part with high possibility of damage is attained by the abnormality level in which frequency level is high.

  As a result, in this embodiment, the measured time series data is decomposed into a plurality of vectors in the time domain and the frequency domain by wavelet transform, and then spectrum analysis is performed on the time series vectors in each frequency band. Thus, it is possible to appropriately extract the peak frequency and amplitude in a wide range from the low frequency region to the high frequency region.

  In addition, it is not necessary for a person to determine the analysis results of a large number of time series data decomposed from one time series data by wavelet transformation, so that much labor is not required and practical.

  In this way, vibroacoustic analysis including distortion and estimation of abnormal parts of equipment can be performed accurately and easily by wavelet transform, prevention of secondary damage by early detection of abnormal parts, and planned maintenance by predicting remaining life. Can be realized.

  Note that the vibroacoustic analysis method and apparatus and the apparatus abnormal part estimation method and apparatus of the present invention are not limited to the above-described embodiments, and various modifications can be made without departing from the scope of the present invention. Of course.

10 Feature Extraction Processing Unit 11 Basic Feature Extraction Processing Unit 12 Spectrum Analysis Feature Extraction Processing Unit 13 Continuous Wavelet Transform Processing Unit 14 Wavelet Transform Feature Extraction Processing Unit 15 Wavelet Transform Spectrum Analysis Feature Extraction Processing Unit 20 Machine Learning Diagnosis Unit 30 Abnormal part estimation part

Claims (8)

A feature quantity extraction process for extracting feature quantities from time-series data of vibroacoustics measured in a driving device;
The feature amount extracted in the feature amount extraction processing step and the feature amount when the device is operating normally and the general measurement data of the device are accumulated, and the feature amount when the device is operating normally by machine learning Build a probability distribution model related to the correlation with the measurement data, give the general distribution data of the device and the feature quantity extracted in the feature quantity extraction processing step to the probability distribution model, and anomaly based on the probability distribution model A vibroacoustic analysis method comprising: a machine learning diagnosis step of calculating a degree. The feature amount extraction processing step includes:
Basic features that extract effective values, peak values, maximum peak differences, skewness, kurtosis, average crest factor, and absolute average amplitude as feature values from the measured acceleration of vibro-acoustic time-series data measured on the equipment in operation An extraction process,
The maximum peak order and its peak value, the second peak order and its peak value, the third peak order and its peak value, the fourth peak order and its peak value as feature quantities after spectral analysis of the time series data, ... A spectral analysis feature extraction process step for extracting the nth peak order and its peak value;
A wavelet transform of the time series data and decomposed into scale levels corresponding to a plurality of frequency bands, and a continuous wavelet transform processing step for obtaining power time series data for each scale level;
A wavelet transform feature quantity extraction process step for extracting an effective value, maximum value, maximum crest factor, and absolute average value as feature quantities of each frequency band from the power time-series data for each scale level obtained in the continuous wavelet transform process step; ,
The maximum peak order and its peak value, the second peak order and its peak value, the third peak order and its as feature quantities after spectral analysis of the power time series data for each scale level obtained in the continuous wavelet transform processing step The vibroacoustic analysis method according to claim 1, further comprising: a peak value, a fourth peak order and its peak value, a wavelet transform spectrum analysis feature quantity extraction process step for extracting the nth peak order and its peak value.   The vibroacoustic analysis method according to claim 1 or 2, wherein the general measurement data of the device is at least one of a rotation speed, an output, a work amount, a pressure, a temperature, and a speed. A feature amount extraction processing unit that extracts feature amounts from time-series data of vibroacoustics measured in a driving device;
The feature amount extracted in the feature amount extraction processing unit and the feature amount in a state where the device is operating normally and the general measurement data of the device are accumulated, and the feature amount and the general state in the device normal operation state by machine learning Construct a probability distribution model related to the correlation with the measurement data, give the general measurement data of the device and the feature quantity extracted by the feature quantity extraction processing unit to the probability distribution model, and anomaly based on the probability distribution model A vibroacoustic analysis apparatus comprising: a machine learning diagnosis unit that calculates a degree. The feature amount extraction processing unit
Basic features that extract effective values, peak values, maximum peak differences, skewness, kurtosis, average crest factor, and absolute average amplitude as feature values from the measured acceleration of vibro-acoustic time-series data measured on the equipment in operation An extraction processing unit;
The maximum peak order and its peak value, the second peak order and its peak value, the third peak order and its peak value, the fourth peak order and its peak value as feature quantities after spectral analysis of the time series data, ... A spectral analysis feature amount extraction processing unit for extracting the n-th peak order and its peak value;
A continuous wavelet transform processing unit for wavelet transforming the time series data and decomposing it into scale levels corresponding to a plurality of frequency bands, and obtaining power time series data for each scale level;
A wavelet transform feature amount extraction processing unit that extracts an effective value, a maximum value, a maximum crest factor, and an absolute average value as feature amounts of each frequency band from power time-series data for each scale level obtained by the continuous wavelet transform processing unit; ,
The maximum peak order and its peak value, the second peak order and its peak value, the third peak order and its feature as a feature after spectral analysis of the power time-series data for each scale level obtained by the continuous wavelet transform processing unit The vibroacoustic analysis apparatus according to claim 4, further comprising: a peak value, a fourth peak order and its peak value, a wavelet transform spectrum analysis feature quantity extraction processing unit that extracts the nth peak order and its peak value.   6. The vibroacoustic analysis apparatus according to claim 4, wherein the general measurement data of the device is at least one of a rotational speed, an output, a work amount, a pressure, a temperature, and a speed. In the machine learning diagnostic step of the vibroacoustic analysis method according to any one of claims 1 to 3, a data set at the time of damage of the degree of abnormality occurring at the time of damage for each part of the device is calculated in advance.
The degree of similarity for each part of the device by comparing the data set at the time of damage calculated in advance in the machine learning diagnosis step with the actual measurement data set of the degree of abnormality calculated from the general measurement data of the device A device abnormal part estimation method comprising: an abnormal part estimation step for obtaining a part having a high possibility of damage in descending order of similarity for each part of the device. The machine learning diagnostic unit of the vibroacoustic analysis apparatus according to any one of claims 4 to 6 is configured to calculate in advance a damage data set of an abnormality degree that occurs at the time of damage for each part of the device,
The degree of similarity for each part of the device by comparing the data set at the time of damage calculated in advance by the machine learning diagnostic unit with the actual measurement data set of the degree of abnormality calculated from the general measurement data of the device An apparatus abnormal part estimation apparatus provided with an abnormal part estimation part which calculates | requires and outputs as a part with high possibility of damage in order with the high similarity for every site | part of this apparatus.

Images