CN114358080A - Bearing vibration monitoring characteristic threshold self-learning method based on generation countermeasure network - Google Patents

Abstract

A bearing vibration monitoring characteristic threshold value self-learning method based on a generated countermeasure network comprises the following steps: acquiring a bearing vibration acceleration signal; carrying out mean value removal and segmentation on the acceleration signals to obtain a plurality of vibration signal sample sets under the same working condition; adopting a convolution pooling layer and a full connection layer to construct a discriminator of a self-learning neural network model, adopting a deconvolution layer to construct a self-learning sample generator, and training the self-learning neural network model by using a vibration signal sample set; generating a vibration signal sample by using a trained self-learning neural network model sample generator to calculate a characteristic value; and multiplying the characteristic value by a certain multiple to define a characteristic threshold value, and providing equipment monitoring use. The method avoids artificial setting of the characteristic threshold, can automatically provide dynamic adjustment of the characteristic threshold of the bearing vibration signal according to the equipment operation rule, and can provide the corresponding characteristic threshold under the background of multiple working conditions of bearing operation, so that the bearing vibration monitoring is more intelligent.

Description

Bearing vibration monitoring characteristic threshold self-learning method based on generation countermeasure network

Technical Field

The invention belongs to the technical field of monitoring of bearing running states, relates to the problem of monitoring of running states of rolling bearings of core components of rotating mechanical equipment, and particularly relates to a bearing vibration monitoring characteristic threshold value self-learning method based on a generated countermeasure network.

Background

In the 4.0 era of industry, rotary mechanical equipment is developing towards high performance, high reliability and the like, and rolling bearings are important components of the equipment, and the condition of the rolling bearings influences whether industrial equipment can play the maximum effect, so that the running condition of the bearings needs to be monitored in real time. At present, a vibration signal analysis method is mostly adopted for monitoring the running state of a bearing, namely, a sensor is utilized to collect vibration acceleration or vibration speed signals in the running process of the bearing, the vibration signals are analyzed and processed by means of a correlation algorithm, the correlation characteristic values of the vibration signals are extracted to represent the running state of the bearing, commonly used characteristic effective values comprise RMS root mean square value, kurtosis value, peak value, variance, signal characteristic frequency and the like, and then the running state of the bearing is diagnosed by utilizing the characteristic values.

In the existing bearing state monitoring algorithm, the state monitoring of the bearing is realized by setting a proper characteristic threshold, for example, a fault diagnosis method based on kurtosis and envelope analysis is provided in the morning, and the method is also complex in calculation and difficult to realize online monitoring. Meanwhile, as the service time of the bearing increases, some slight abrasion can occur to cause the vibration signal to change, and the set characteristic threshold value is not applicable any more. In addition, in the process of leaving a factory, characteristic vibration characteristic thresholds of some parts are also set, and when the vibration of the mechanical equipment exceeds the set threshold, an alarm signal is sent out.

A generated countermeasure network (GAN) is used as a novel data generation neural network in the deep learning field, and plays a great role in image recognition, data feature extraction and the like. The method is characterized in that the generated countermeasure network is based on a zero-sum game theory, a network model enables a sample generator to learn the distribution characteristics of real samples by establishing a sample generator and a discriminator and by means of two training modes of generation and countermeasure, so that the change characteristic rule and effective characteristic information of the sample data are mined, and the functions of data diagnosis, identification, prediction and early warning are achieved.

Based on the method, in order to realize the real-time monitoring of the bearing running state, the method carries out self-learning modeling on the historical data of the bearing running by means of the generation countermeasure network, obtains the bearing vibration change rule, and sets the corresponding characteristic threshold value to monitor the bearing running state in real time.

Disclosure of Invention

The invention aims to provide a bearing vibration monitoring characteristic threshold value self-learning method based on a generated countermeasure network.

The technical scheme of the invention is as follows:

a bearing vibration monitoring characteristic threshold value self-learning method based on a generated countermeasure network comprises the following steps:

step S1: self-learning eigenvalue selection

Selecting a vibration signal characteristic root mean square value and a kurtosis value as a characteristic threshold design reference for monitoring the bearing state; step S2: self-learning sample preprocessing

Collecting bearing running state data in a period T by using a vibration acceleration sensor, carrying out mean value removing pretreatment on bearing running state data samples, and then segmenting to obtain a plurality of short samples x with the same length of n_r(ii) a Each short sample marks working condition information including bearing running rotating speed, load and sampling time and stores the working condition information in a database; selecting a bearing running state data sample of a period T from a database according to the same working condition as a training sample set X of the self-learning neural network model_train；

According to the bearing operation working conditions, correspondingly collecting a period sample in each working condition and storing the period sample in a sample database so as to construct a training sample set of the self-learning neural network model in each working condition;

step S3: method for constructing characteristic threshold value self-learning neural network model

Generating an antagonistic neural network based on deep learning, wherein a self-learning neural network model mainly comprises a generator G and a discriminator D; the generator G comprises three layers of one-dimensional convolution layers, Decov1-Decov2-Decov3The input of the generator G is a Z-dimensional Gaussian white noise data different from the training sample set in S2, the data upsampling is realized through three layers of convolution layers, and the output is a data sample x with the length of n_g(ii) a The discriminator D is a two-classification convolution network comprising two convolution pooling layers and three full-connection layers, i.e. Cov1+ Pool1-Cov2+ Pool2-FC1-FC2-FC3, the input of the discriminator D is a data sample x with the length of n, and the output of the discriminator D is converted into a matrix M (M is a matrix M with the length of 2) through a Sigmoid function₁,m₂) Wherein

Wherein m is₁Expressed as input x being x_rProbability of (m) of₂Then it means that the input x is x_gThe probability of (d);

step S4, using training sample set X_trainTraining self-learning neural network model

Training a self-learning neural network model by using the training sample set constructed in the step 2, wherein a training target function is shown as a formula (1), and the training is mainly divided into the training of a discriminator D and a generator G;

firstly, training a discriminator D, and collecting a training sample set X_trainInputting the result into a discriminator D, calculating the loss of the discriminator D by using the cross entropy loss function shown in the formula (2)₁(ii) a The generator G then generates an equal number of data samples x_gAnd input into a discriminator D to calculate the loss by using the cross entropy loss function of the formula (2)₂Merging loss₁And loss₂Feeding back to a discriminator D and carrying out gradient updating through an Adam function;

next, generator G training is performed, generator G regenerates data sample x'_gAnd data sample x_gThe flag 1 is input to a discriminator D for evaluation, i.e. the data sample x is calculated_gBelong to X_trainCalculating the loss of the current time by using a cross entropy loss function of the formula (2), and then transferring the loss to a generator G and updating the gradient of the G through an Adam function;

in the formula (1), D (x) represents x_rProbability of belonging to 1, and D (G (z)) represents x_gA probability of belonging to 0;

in the formula (2), y_iThe label representing the sample i, for the present method, the sample x_rIs marked as 1, and sample x_rIs marked as 0, and p_iThen the probability that sample i is identified as 1 is indicated;

then, carrying out second iterative training, namely repeating the training process of the discriminator D and the generator G, and updating the parameters of the discriminator D, so that the discriminator D can accurately identify the real sample x_rAnd generating a sample x_g(ii) a Iteratively training the generator G to generate a sample x_gIdentified as true sample x by discriminator D_rThe probability of (a) is continuously improved, namely the generated sample x is obtained_gSample distribution of (2) and true sample x_rThe closer the sample distribution of (a); by analogy, multiple times of self-learning neural network model training are completed, the generator G and the discriminator D of the epoch generation are obtained, and the discriminator D cannot distinguish that the input sample x is the self-learning generated sample x_gAnd true sample x_r；

Step S5: generating self-learning samples using a trained sample generator G

Generating a group of generated samples x from the randomly distributed white Gaussian noise data by using the self-learning sample generator G trained in the step 4_gThis generates a sample x_gThe sample is obtained through self-learning of the bearing vibration signal sample, and the vibration signal characteristic threshold value is calculated based on the generated sample;

step S6: feature threshold calculation

Generating sample x obtained in step 5_gAs characteristic threshold value self-learning reference data under corresponding working conditions, calculating and generating sample x by using formula (4) and formula (5)_gRMS and kurtosis value K, then respectivelyMultiplying by 1.2 times to obtain characteristic threshold value VT of final self-learning neural network model self-learning sample₁And VT₂；

Wherein n is the length of a single sample, and x (i) is the signal value;

step S7: feature threshold update

And selecting an updating period T of the characteristic threshold, repeating the steps S1 to S6, and learning and updating the characteristic threshold of the bearing vibration signal by using the vibration signal samples within the period T.

The invention has the beneficial effects that: the invention provides a bearing vibration signal characteristic threshold self-learning method based on a generated countermeasure network, which comprises the following steps: acquiring a bearing vibration acceleration signal; carrying out mean value removal and segmentation on the acceleration signals to obtain a plurality of vibration signal sample sets under the same working condition; adopting a convolution pooling layer and a full connection layer to construct a discriminator of a self-learning neural network model, adopting a deconvolution layer to construct a self-learning sample generator, and training the self-learning neural network model by using a vibration signal sample set; generating a vibration signal sample by using a trained self-learning neural network model sample generator to calculate a characteristic value; and multiplying the characteristic value by a certain multiple to define a characteristic threshold value, and providing equipment monitoring use.

Compared with the prior art, the method avoids artificial setting of the characteristic threshold, dynamic adjustment of the characteristic threshold of the bearing vibration signal can be automatically given according to the equipment operation rule, and meanwhile, the method can give the corresponding characteristic threshold under the background of multiple working conditions of bearing operation, so that the bearing vibration monitoring is more intelligent.

Drawings

FIG. 1 is a flow chart of a rolling bearing vibration monitoring characteristic threshold value self-learning method based on generation of a countermeasure network provided by the invention

FIG. 2 is a self-learning neural network model architecture diagram based on generating a countermeasure network

FIG. 3 shows 4 sets of bearing vibration signal samples obtained by training the model of the present method using 4 sets of data sets, where (a) is the data set No. 1, (b) is the data set No. 2, (b) is the data set No. 3, and (d) is the data set No. 4.

Detailed Description

In order to make the technical problems, technical solutions and advantages of the present invention more apparent, embodiments of the present invention will be described in detail below with reference to the accompanying drawings and experimental data analysis.

In the implementation, a self-learning method for a rolling bearing vibration monitoring characteristic threshold based on a generated countermeasure network is provided, and fig. 1 shows a flow of the self-learning updating method for the bearing vibration monitoring characteristic threshold, and the method comprises the following steps.

Step S1: vibration signal sample acquisition and pre-processing

The method comprises the steps of collecting vibration acceleration signals of a bearing running state by adopting a vibration acceleration sensor, removing the mean value of the obtained vibration time domain signals, and then dividing a sample to obtain a plurality of small samples x_rWherein each small sample x_rAnd marking the rotating speed of the running of the bearing, the load information load and the corresponding sample acquisition time stamp, and storing the data sample into a database.

In order to explain the implementation process details of the method in more detail, the invention introduces the execution process of the method in detail on the basis of a bearing test platform in a vibration engineering laboratory of university of great university of great university of great university of great university of china, and great university of great university of great university of great university of great. In order to simulate the physical changes of bearing parts along with the running of time, the experimental scheme selects the bearings in two states to carry out vibration signal data acquisition, wherein the first state is that a brand new bearing is used for simulating the initial running state of the bearing, and the other state is that the bearing with the inner ring worn by 0.1mm simulates the bearing running for a long time. Are respectively arranged on the bearing test boardThe two bearings are installed for collecting vibration signal data, and the collection scheme is shown in table 1. Based on the experimental acquisition scheme of table 1, four vibration signal samples with the length of 310K are acquired by using the vibration acceleration sensor, wherein sample numbers 1 and 3 are comparison groups under one working condition, and samples 2 and 4 are comparison groups under another working condition. Firstly, the four long samples are subjected to mean value removing treatment, then a sample segmentation scheme with a window length of 1024 points and a sliding distance of 512 points is selected, the four long samples are segmented, and four types of sample sets X are obtained_train∈{X₁,X₂,X₃,X₄The first 600 samples, i.e. the number of samples (X), are selected from each type of sample set_train) 600, and a single sample length, length (x)_r)＝1024。

Table 1 data acquisition protocol

Serial number

Bearing condition

Simulation of states

Sampling rate

Rotational speed

Load(s)

Sampling time

1

Brand-new bearing

Initial operation

10Ks/S

1200rpm

0

31S

2

Brand-new bearing

Initial operation

10Ks/S

900rpm

0

31S

3

Abrasion of 0.1mm

Late stage operation

10Ks/S

1200rpm

0

31S

4

Abrasion of 0.1mm

Late stage operation

10Ks/S

900rpm

0

31S

Step S2: construction of self-learning neural network model

The method comprises the steps that an antagonistic neural network is generated based on deep learning, a characteristic threshold self-learning neural network model is mainly composed of a generator network G and a discriminator D, wherein the generator G comprises three groups of one-dimensional convolutional layers, Decov1-Decov2-Decov3, the input of the generator G is one piece of Gaussian white noise data with the dimension of z being 100, data up-sampling is achieved through three layers of deconvolution layers, and the output is one piece of white noise data with the length of n being 1001024 sequence data x_g. The discriminator D is a two-classification convolution network comprising two convolution pooling layers and three full-connection layers, i.e. Cov1+ Pool1-Cov2+ Pool2-FC1-FC2-FC3, the input of the discriminator D is sequence data x with the length of n being 1024, and the output of the discriminator D is converted into a matrix M (M) with the length of 2 through a Sigmoid function₁,m₂) Wherein

Wherein m is₁Expressed as input x being x_rProbability of (m) of₂Then it is denoted as x_gThe probability of (c).

Based on the deep learning framework of the pyrrch, the built self-learning neural network model network realizes model construction and subsequent model training by utilizing the python language.

Step S3: self-learning neural network model training process

In step 2, a self-learning neural network model is written by using python language, and four groups of sample sets X preprocessed in step 1 are used in the step_trainAnd respectively training the self-learning neural network model.

The training of the self-learning neural network model is divided into two processes, firstly, the training of a discriminator D of the self-learning neural network model is carried out, and a group of sample sets X_trainInput to D for recognition, output converted to one [0,1 ] by Sigmoid]Interval value D (x) of_r) Calculation of [ D (x) using Cross-entry_r),1]Loss of₁Generating samples x from randomly distributed Gaussian noise of length 100 using a self-learning sample generator G_fInput the discriminator x_fCross-entry was also used to calculate [ D (x)_f),0]Loss of₂Will lose₁And loss₂And adding the sum and feeding the sum back to a discriminator, and performing individual updating of the D gradient of the discriminator by using an Adam optimization algorithm.

Then, iterative training of a generator G is carried out, firstly Gaussian white noise data with the length of 100 is constructed, and then the Gaussian white noise data is input into the generator G for up-sampling to obtain a sequence x with the length of 1024_fX is to be_fInputting the result into a discriminator D for identification, and mapping the output result to [0,1 ] through Sigmoid]Interval(s)A value D (x)_f) And calculated by Cross-entry [ D (x)_f),1]And (4) model loss, wherein the loss is transmitted to a generator G by using an Adam function to perform gradient descent, and parameters of G are updated.

Further, in the process, the training of the two models G and D is one-time iterative training of the sample self-learning neural network model, and according to the specific sample training condition, the model iterative training ending condition is set or the total model training iteration times is set.

Further, four groups of bearing vibration data samples preprocessed in the step 2 are provided, wherein 1, 3, 2 and 4 are comparison groups, the four groups of data sets repeat the training method of the self-learning neural network model in the step, the total number of times of single iterative training of each model is 4000, the learning rate is 0.001, and finally four self-learning neural network models, namely 4 self-learning sample generators G are obtained.

And 4, step 4: self-learning sample generation

Constructing Gaussian white noise data with the length of 100 based on the 4 self-learning neural network model sample generators G trained in the step 3, respectively inputting the Gaussian white noise data into the 4 generators G to obtain 4 groups of sequence samples with the length of 1024, namely the sample x obtained by the model self-learning_l1,x_l2,x_l3,x_l4The resulting sample is shown in fig. 3.

And 5: feature threshold calculation and update

Respectively calculating 4 groups of generated samples x in step 4_l1,x_l2,x_l3,x_l4RMS value and kurtosis value R of_l,K_lObtaining 8 characteristic values R_l1,K_l1,R_l2,K_l2R_l3,K_l3R_l4,K_l4Each of the combinations R_l,K_lAs the vibration signal characteristic value under the corresponding working condition, multiplying the calculated characteristic value by 1.5 to be used as the final self-learning characteristic threshold value VT of the sample₁,VT₂Then a total of eight characteristic thresholds VT are obtained₁₁,VT₁₂,VT₂₁,VT₂₂,VT₃₁,VT₃₂,VT₄₁,VT₄₂Two features per setThreshold VT₁,VT₂The characteristic threshold value of the vibration monitoring signal of the bearing equipment under the corresponding working condition obtained by the self-learning of the model is transmitted to a corresponding monitoring system for use.

Step 6: feature threshold self-learning self-updating

Setting an updating period of the self-learning neural network model, if taking 10 days as a standard, carrying out model self-learning updating on bearing vibration signal samples accumulated in nearly 10 days, repeating the steps 1 to 5, obtaining the latest vibration signal characteristic threshold value under the corresponding working condition, sending the latest vibration signal characteristic threshold value to a monitoring end for monitoring the running state of the bearing, and updating the model to obtain the vibration characteristic threshold value when the monitoring data of the next 10 days is full, so as to realize dynamic adjustment of the vibration characteristic threshold value.

Further, 4 vibration signal sample sets { X } have been introduced in step 1₁,X₂,X₃,X₄In which X is₁,X₃For vibration signal data sets of the initial stage and the later stage of the bearing operation under the same working condition, and for comparison, the threshold value monitoring of the initial vibration characteristic of the working condition is carried out by VT₁₁,VT₁₂For reference, after the later stage, through sample self-learning, VT is used₃₁,VT₃₂As a reference. In the second condition, i.e. data sample set { X₂,X₄Vibration signature threshold monitoring will be from the initial VT₂₁,VT₂₂To VT₄₁,VT₄₂And (6) updating.

Further, 4 sets of sample data { X }₁,X₂,X₃,X₄Practical tests are carried out to obtain the characteristic threshold, and as shown in the result table 2, the signal characteristic threshold can be seen to be changed, so that the effectiveness of the method is further verified.

TABLE 2 self-learning neural network model feature threshold

Serial number	Bearing condition	Rotational speed	Load(s)	Rl	Kl	VT1	VT2
								1	Brand-new bearing	1200rpm	0	0.078	3.178	0.117	4.767
2	Brand-new bearing	900rpm	0	0.058	3.372	0.087	5.058
								3	Abrasion of 0.1mm	1200rpm	0	0.243	7.322	0.3645	10.983
4	Abrasion of 0.1mm	900rpm	0	0.219	8.617	0.3285	12.9255

Although the embodiments of the present invention have been shown and described, it is understood that the above embodiments are only for illustrating the technical solutions of the present invention and should not be construed as limiting the present invention, and those skilled in the art can make modifications and substitutions to the above embodiments within the scope of the present invention without departing from the principle and spirit of the present invention.

Claims (1)

1. A bearing vibration monitoring characteristic threshold value self-learning method based on a generated countermeasure network is characterized by comprising the following steps:

step S1: self-learning eigenvalue selection

Selecting a vibration signal characteristic root mean square value and a kurtosis value as a characteristic threshold design reference for monitoring the bearing state;

step S2: self-learning sample preprocessing

Collecting bearing running state data in a period T by using a vibration acceleration sensor, carrying out mean value removing pretreatment on bearing running state data samples, and then segmenting to obtain a plurality of short samples x with the same length of n_r(ii) a Each short sample marks working condition information including bearing running rotating speed, load and sampling time and stores the working condition information in a database; from a databaseSelecting a bearing running state data sample of a period T as a training sample set X of the self-learning neural network model according to the same working condition_train；

According to the bearing operation working conditions, correspondingly collecting a period sample in each working condition and storing the period sample in a sample database so as to construct a training sample set of the self-learning neural network model in each working condition;

step S3: method for constructing characteristic threshold value self-learning neural network model

Generating an antagonistic neural network based on deep learning, wherein a self-learning neural network model mainly comprises a generator G and a discriminator D; the generator G comprises three layers of one-dimensional convolutional layers, namely Decov1-Decov2-Decov3, the input of the generator G is one piece of Z-dimensional Gaussian white noise data different from the training sample set in S2, data upsampling is realized through the three layers of convolutional layers, and the output is a data sample x with the length of n_g(ii) a The discriminator D is a two-classification convolution network comprising two convolution pooling layers and three full-connection layers, i.e. Cov1+ Pool1-Cov2+ Pool2-FC1-FC2-FC3, the input of the discriminator D is a data sample x with the length of n, and the output of the discriminator D is converted into a matrix M (M is a matrix M with the length of 2) through a Sigmoid function₁,m₂) Wherein

Wherein m is₁Expressed as input x being x_rProbability of (m) of₂Then it means that the input x is x_gThe probability of (d);

step S4, using training sample set X_trainTraining self-learning neural network model

Training a self-learning neural network model by using the training sample set constructed in the step 2, wherein a training target function is shown as a formula (1), and the training is mainly divided into the training of a discriminator D and a generator G;

firstly, training a discriminator D, and collecting a training sample set X_trainInputting the result into a discriminator D, calculating the loss of the discriminator D by using the cross entropy loss function shown in the formula (2)₁(ii) a The generator G then generates an equal number of data samples x_gAnd input into a discriminator D to calculate the loss l by using the cross entropy loss function of the formula (2)oss₂Merging loss₁And loss₂Feeding back to a discriminator D and carrying out gradient updating through an Adam function;

next, generator G training is performed, generator G regenerates data sample x'_gAnd data sample x_gThe flag 1 is input to a discriminator D for evaluation, i.e. the data sample x is calculated_gBelong to X_trainCalculating the loss of the current time by using a cross entropy loss function of the formula (2), and then transferring the loss to a generator G and updating the gradient of the G through an Adam function;

in the formula (1), D (x) represents x_rProbability of belonging to 1, and D (G (z)) represents x_gA probability of belonging to 0;

in the formula (2), y_iThe label representing the sample i, for the present method, the sample x_rIs marked as 1, and sample x_rIs marked as 0, and p_iThen the probability that sample i is identified as 1 is indicated;

then, carrying out second iterative training, namely repeating the training process of the discriminator D and the generator G, and updating the parameters of the discriminator D, so that the discriminator D can accurately identify the real sample x_rAnd generating a sample x_g(ii) a Iteratively training the generator G to generate a sample x_gIdentified as true sample x by discriminator D_rThe probability of (a) is continuously improved, namely the generated sample x is obtained_gSample distribution of (2) and true sample x_rThe closer the sample distribution of (a); by analogy, multiple times of self-learning neural network model training are completed, the generator G and the discriminator D of the epoch generation are obtained, and the discriminator D cannot distinguish that the input sample x is the self-learning generated sample x_gAnd true sample x_r；

Step S5: generating self-learning samples using a trained sample generator G

Generating a group of generated samples x from the randomly distributed white Gaussian noise data by using the self-learning sample generator G trained in the step 4_gThis generates a sample x_gThe sample is obtained through self-learning of the bearing vibration signal sample, and the vibration signal characteristic threshold value is calculated based on the generated sample;

step S6: feature threshold calculation

Generating sample x obtained in step 5_gAs characteristic threshold value self-learning reference data under corresponding working conditions, calculating and generating sample x by using formula (4) and formula (5)_gThe RMS value and the kurtosis value K are multiplied by 1.2 times respectively to be used as the characteristic threshold value VT of the final self-learning neural network model self-learning sample₁And VT₂；

Wherein n is the length of a single sample, and x (i) is the signal value;

step S7: feature threshold update

And selecting an updating period T of the characteristic threshold, repeating the steps S1 to S6, and learning and updating the characteristic threshold of the bearing vibration signal by using the vibration signal samples within the period T.

Images