A self-learning method for characteristic threshold of bearing vibration monitoring based on generative adversarial network
By generating an adversarial network self-learning model to dynamically adjust the characteristic threshold of bearing vibration signal, the problem of poor adaptability of online monitoring and characteristic thresholds in the prior art is solved, real-time monitoring and intelligent adaptation of bearing status are achieved.
Patent Information
- Application Number
- CN202210003381.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-04
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-01-04
AI Technical Summary
The existing bearing status monitoring methods are difficult to achieve online monitoring, and the preset characteristic thresholds are changed due to slight wear during the bearing service, causing vibration signals to change and fail, making it difficult to adapt to complex actual operating conditions.
The self-learning method of bearing vibration monitoring feature threshold based on the generative adversarial network is adopted. Through self-learning of neural network models, the characteristic root mean square and kurtitude values of vibration signal characteristics are used to dynamically adjust the characteristic threshold to avoid artificial settings and adapt to multi-working environments.
Real-time monitoring of bearing operating status is realized, the characteristic threshold is automatically adjusted, and the intelligence and adaptability of monitoring is improved, and it is suitable for bearing equipment under complex working conditions.
Smart Images

Figure CN114358080B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of bearing operating status monitoring, and relates to the operating status monitoring problem of rolling bearings, which are the core components of rotating mechanical equipment. The present invention is specifically a bearing vibration monitoring feature threshold self-learning method based on a generative adversarial network. Background Art
[0002] Entering the era of Industry 4.0, rotating mechanical equipment is developing towards high performance and high reliability. As an important component of these equipment, the state of rolling bearings affects whether industrial equipment can play its maximum role, so it is necessary to monitor the operating status of bearings in real time. At present, the vibration signal analysis method is mostly used for bearing operating status monitoring, that is, the vibration acceleration or vibration velocity signal during the operation of the bearing is collected by the sensor, and the vibration signal is analyzed and processed with the help of relevant algorithms, and the relevant characteristic values of the vibration signal are extracted to characterize the operating status of the bearing. The commonly used characteristic effective values are RMS root mean square value, kurtosis value, peak-to-peak value, variance, signal characteristic frequency, etc., and then these characteristic values are used to diagnose the operating status of the bearing.
[0003] In the existing bearing condition monitoring algorithm, the bearing condition monitoring is achieved by setting appropriate characteristic thresholds. For example, Yang Chen proposed a fault diagnosis method based on kurtosis and envelope analysis. This method is also complex in calculation and difficult to achieve online monitoring. At the same time, as the service time of the bearing increases, some slight wear will occur, causing the vibration signal to change, and the set characteristic threshold is no longer applicable. In addition, in the process of some equipment leaving the factory, characteristic vibration characteristic thresholds of some components are also set. When the vibration of the mechanical equipment exceeds the set threshold, an alarm signal is issued. However, the setting of these characteristic values is based on standard test conditions, and the actual operating conditions of the bearing equipment are complex, and the pre-set characteristic values are difficult to apply.
[0004] Generative adversarial network (GAN) is a new type of data generation neural network in the field of deep learning, which has played a huge role in image recognition, data feature extraction, etc. Generative adversarial network is based on zero-sum game theory. The network model establishes a sample generator and a discriminator. With the help of two training modes, generation and adversarial, the sample generator learns the distribution characteristics of real samples, thereby mining the changing characteristic laws and effective feature information of sample data, which plays a role in data diagnosis, recognition, prediction and early warning.
[0005] Based on this, in order to realize real-time monitoring of the bearing operating status, this method uses a generative adversarial network to perform self-learning modeling on the historical data of the bearing operation, obtain the bearing vibration change law, and set the corresponding feature threshold to monitor the bearing operating status in real time. Summary of the invention
[0006] The purpose of the present invention is to provide a bearing vibration monitoring feature threshold self-learning method based on a generative adversarial network.
[0007] The technical solution of the present invention:
[0008] A bearing vibration monitoring feature threshold self-learning method based on a generative adversarial network includes the following steps:
[0009] Step S1: Self-learning feature value selection
[0010] The root mean square and kurtosis values of the vibration signal characteristics are selected as the characteristic threshold design benchmarks for bearing condition monitoring; Step S2: Self-learning sample preprocessing
[0011] The bearing operation status data within a period T is collected by using a vibration acceleration sensor, and the bearing operation status data samples are preprocessed by removing the mean value, and then segmented to obtain multiple short samples x with the same length n. r ; Each short sample marks the operating condition information, including the bearing operating speed, load and sampling time, and stores it in the database; from the database, select a bearing operating status data sample with a period T according to the same operating condition as the training sample set X of the self-learning neural network model train ;
[0012] According to the operating conditions of the bearing, a periodic sample is collected for each operating condition and stored in the sample database to construct a training sample set of the self-learning neural network model under each operating condition;
[0013] Step S3: Constructing a feature threshold self-learning neural network model
[0014] Based on deep learning generative adversarial neural network, the self-learning neural network model is mainly composed of generator G and discriminator D; generator G contains three layers of one-dimensional convolutional layers, Decov1-Decov2-Decov3, and the input of generator G is a z-dimensional Gaussian white noise data that is different from the training sample set in S2. Through the three layers of convolutional layers, data upsampling is achieved, and the output is a data sample x with a length of n g ; The discriminator D is a binary convolutional network, which includes two convolutional pooling layers and three fully connected layers, Cov1+Pool1-Cov2+Pool2-FC1-FC2-FC3. The input of the discriminator D is a data sample x of length n, and the output is converted into a matrix M(m1,m2) of length 2 through the Sigmoid function, where Where m1 represents the input x is x r , and m2 means that the input x is x g The probability of
[0015] Step S4: using the training sample set Xtrain Training a self-learning neural network model
[0016] Using the training sample set constructed in step 2, the self-learning neural network model is trained. The training objective function is shown in formula (1). The training is mainly divided into the training of the discriminator D and the generator G.
[0017] First, the discriminator D is trained and the training sample set X train Input the discriminator D, and use the cross entropy loss function shown in formula (2) to calculate the loss1 of the discriminator D; then the generator G generates the same amount of data samples x g And input the discriminator D, use the cross entropy loss function of formula (2) to calculate the loss loss2, merge loss1 and loss2 and feed them back to the discriminator D, and perform gradient update through the Adam function;
[0018] Then the generator G is trained and the generator G regenerates the data sample x′ g , and the data sample x g Input the label 1 into the discriminator D for identification, that is, calculate the data sample x g Belong to X train The probability of , using the cross entropy loss function of formula (2) to calculate the loss at this time, and then the loss will be passed to the generator G and the gradient of G will be updated through the Adam function;
[0019]
[0020] In formula (1), D(x) represents x r The probability of belonging to 1, and D(G(z)) represents x g The probability of belonging to 0;
[0021]
[0022] In formula (2), y i represents the label of sample i. For this method, sample x r is marked as 1, and the sample x r The sign of is 0, and p i It represents the probability that sample i is identified as 1;
[0023] Then, the second iterative training is carried out, that is, the training process of the discriminator D and the generator G is repeated, and the parameters of the discriminator D are updated so that the discriminator D can accurately identify the real sample x r And generate samples x g ; Iteratively train the generator G to generate samples x g Identified as a true sample x by the discriminator D r The probability of generating samples x is continuously increased. gThe sample distribution and true sample x r The sample distribution of is getting closer; and so on, after completing multiple self-learning neural network model training, the generator G and discriminator D of the epoch generation are obtained. At this time, the discriminator D cannot distinguish whether the input sample x is the self-learning generated sample x g and the real sample x r ;
[0024] Step S5: Generate self-learning samples using the trained sample generator G
[0025] Using the self-learning sample generator G trained in step 4, generate a set of generated samples x from the randomly distributed Gaussian white noise data g , which generates a sample x g That is, the sample is obtained through self-learning of the bearing vibration signal sample, and the vibration signal characteristic threshold will be calculated based on the generated sample;
[0026] Step S6: Feature threshold calculation
[0027] Generate sample x obtained in step 5 g As the characteristic threshold self-learning benchmark data under the corresponding working conditions, the sample x is calculated and generated using equations (4) and (5) g The root mean square value RMS and kurtosis value K are then multiplied by 1.2 times as the feature thresholds VT1 and VT2 of the self-learning samples of the final self-learning neural network model;
[0028]
[0029]
[0030] Where n is the length of a single sample, and x(i) is the signal value;
[0031] Step S7: Feature threshold update
[0032] Select an update period T of the characteristic threshold, repeat steps S1 to S6, and use the vibration signal samples within the period T to learn and update the characteristic threshold of the bearing vibration signal.
[0033] The beneficial effects of the present invention are as follows: the present invention provides a method for self-learning of characteristic thresholds of bearing vibration signals based on a generative adversarial network, the method comprising the following steps: obtaining a bearing vibration acceleration signal; de-averaging and segmenting the acceleration signal to obtain a plurality of vibration signal sample sets under the same working condition; using a convolutional pooling layer and a fully connected layer to construct a discriminator of a self-learning neural network model, using a deconvolution layer to construct a self-learning sample generator, and using a vibration signal sample set to train the self-learning neural network model; using the trained self-learning neural network model sample generator to generate vibration signal samples for eigenvalue calculation; multiplying the eigenvalue by a certain multiple to define it as a characteristic threshold, and providing it for equipment monitoring.
[0034] Compared with the existing technology, this method avoids manually setting characteristic thresholds and can automatically give dynamic adjustments to the characteristic thresholds of bearing vibration signals according to the operating rules of the equipment. At the same time, in the context of multiple operating conditions of the bearing, this method can give corresponding characteristic thresholds, making bearing vibration monitoring more intelligent. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 This is a flow chart of the self-learning method for rolling bearing vibration monitoring feature threshold based on generative adversarial network provided by the present invention
[0036] Figure 2 This is a self-learning neural network model architecture diagram based on the generative adversarial network
[0037] Figure 3 The four groups of bearing vibration signal samples under different working conditions are obtained by training the model of this method using four groups of data sets, among which (a) is data set No. 1, (b) is data set No. 2, (b) is data set No. 3, and (d) is data set No. 4. DETAILED DESCRIPTION
[0038] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, the specific implementation methods of the present invention will be described in detail below in combination with the accompanying drawings and analysis of experimental data.
[0039] In this implementation, a self-learning method for characteristic threshold of rolling bearing vibration monitoring based on generative adversarial network is proposed. Figure 1 The present invention shows the process of the self-learning update method of the characteristic threshold of the bearing vibration monitoring, which includes the following steps.
[0040] Step S1: Vibration signal sample acquisition and preprocessing
[0041] The vibration acceleration sensor is used to collect the vibration acceleration signal of the bearing running state, the obtained vibration time domain signal is de-averaged, and then the sample is segmented to obtain several small samples x r , where each small sample x rMark the bearing running speed speed, load information load, and the corresponding sample collection timestamp time, and store the data samples in the database.
[0042] In order to explain the implementation process of the method in more detail, the present invention is based on the bearing test platform of the vibration engineering laboratory of Dalian University of Technology, and introduces the execution process of the method in detail. The bearing test bench is mainly composed of a DC variable frequency motor, a transmission shaft base, a coupling, a rotating disc, and a test bearing base. The bearing model supported by the test bearing base is NU205EM. In order to simulate the physical changes of the bearing components over time, the experimental scheme selected bearings in two states for vibration signal data collection. The first is a brand new bearing used to simulate the initial operation state of the bearing, and the other is a bearing with an inner ring wear of 0.1mm to simulate the bearing after long-term operation. These two types of bearings are installed on the bearing test bench to collect vibration signal data, and the collection scheme is shown in Table 1. Based on the experimental collection scheme in Table 1, four vibration signal samples with a length of 310K are collected using a vibration acceleration sensor, where sample numbers 1 and 3 are comparison groups under one working condition, and samples 2 and 4 are comparison groups under another working condition. First, the four long samples are processed by removing the mean, and then a sample segmentation scheme with a window length of 1024 points and a sliding distance of 512 points is selected to segment the four long samples, and four types of sample sets X are obtained. train ∈{X1,X2,X3,X4}, the first 600 samples are selected from each sample set, that is, the number of samples number(X train )=600, and the single sample length length(x r )=1024.
[0043] Table 1 Data acquisition experimental plan
[0044] Serial number Bearing condition Simulation Status Sampling rate Speed load Sampling time 1 New bearings Initial operation 10Ks / S 1200rpm 0 31S 2 New bearings Initial operation 10Ks / S 900rpm 0 31S 3 Wear 0.1mm Late run 10Ks / S 1200rpm 0 31S 4 Wear 0.1mm Late run 10Ks / S 900rpm 0 31S
[0045] Step S2: Building a self-learning neural network model
[0046] Based on deep learning generative adversarial neural network, the feature threshold self-learning neural network model is mainly composed of a generator network G and a discriminator D. The generator G contains three sets of one-dimensional convolutional layers, Decov1-Decov2-Decov3. The input of the generator G is a Gaussian white noise data of z = 100 dimensions. Through three layers of deconvolution layers, data upsampling is achieved, and the output is a sequence data x with a length of n = 1024. gThe discriminator D is a binary convolutional network, which contains two convolutional pooling layers and three fully connected layers, Cov1+Pool1-Cov2+Pool2-FC1-FC2-FC3. The input of the discriminator D is the sequence data x of length n=1024, and the output is converted into a matrix M(m1,m2) of length 2 by the Sigmoid function, where Where m1 represents the input x is x r The probability of m2 is x g probability.
[0047] Based on the pytorch deep learning framework, the self-learning neural network model network is built using the python language to implement model construction and subsequent model training.
[0048] Step S3: Self-learning neural network model training process
[0049] In step 2, the self-learning neural network model is written in Python. This step will use the four sample sets X preprocessed in step 1. train Self-learning neural network model training is performed separately.
[0050] The training of the self-learning neural network model is divided into two steps. The first step is to train the discriminator D of the self-learning neural network model. train The input is sent to D for recognition, and the output is converted into an interval value D(x) of [0,1] through Sigmoid. r ), using Cross-entropy to calculate [D(x r ),1], using the self-learning sample generator G to generate samples x from randomly distributed Gaussian noise of length 100 f , input discriminator x f , also use Cross-entropy to calculate [D(x f ),0], loss1 and loss2 are added and fed back to the discriminator, and the Adam optimization algorithm is used to update the gradient of the discriminator D.
[0051] Next, the generator G is iteratively trained. First, a Gaussian white noise data with a length of 100 is constructed, and then input into the generator G for upsampling to obtain a sequence x with a length of 1024. f , x f Input to the discriminator D for identification, and map the output result to a value D(x) in the interval [0,1] through Sigmoid. f ), and calculate [D(x f),1] Model loss loss, use Adam function to pass loss to generator G for gradient descent, and update the parameters of G.
[0052] Furthermore, the training of the two models G and D in this process is an iterative training of the sample self-learning neural network model. According to the specific sample training situation, the model iterative training end condition is set or the total number of model training iterations is set.
[0053] Furthermore, the four groups of bearing vibration data samples preprocessed in step 2, where 1 and 3, 2 and 4 are control groups, repeat the self-learning neural network model training method in this step for these four data sets. Each model is iterated and trained a total of 4000 times, with a learning rate of 0.001. Finally, four self-learning neural network models are obtained, that is, four self-learning sample generators G are obtained.
[0054] Step 4: Self-learning sample generation
[0055] Based on the four self-learning neural network model sample generators G trained in step 3, construct Gaussian white noise data with a length of 100 and input them into the four generators G respectively to obtain four groups of sequence samples with a length of 1024. This is the sample x obtained by model self-learning. l1 ,x l2 ,x l3 ,x l4 , generate samples as shown in the attached Figure 3 shown.
[0056] Step 5: Feature threshold calculation and update
[0057] Calculate the 4 groups of generated samples x in step 4 respectively l1 ,x l2 ,x l3 ,x l4 The RMS value and kurtosis value R l ,K l , and get 8 eigenvalues R l1 ,K l1 ,R l2 ,K l2 R l3 ,K l3 R l4 ,K l4 , each combination R l ,K l As the vibration signal characteristic value under the corresponding working condition, the calculated characteristic value is multiplied by 1.5 as the final self-learning characteristic threshold VT1, VT2 of the sample, and a total of eight characteristic thresholds VT are obtained. 11 ,VT 12 ,VT 21 ,VT 22,VT 31 ,VT 32 ,VT 41 ,VT 42 Each group of two characteristic thresholds VT1 and VT2 is the characteristic threshold of the vibration monitoring signal of the bearing equipment under the corresponding working conditions obtained by the model self-learning, and it is transmitted to the corresponding monitoring system for use.
[0058] Step 6: Feature threshold self-learning and self-updating
[0059] Set the update cycle of the self-learning neural network model, for example, 10 days. The bearing vibration signal samples accumulated in the past 10 days are used to perform self-learning updates on the model. Repeat steps 1 to 5 to obtain the latest vibration signal characteristic threshold under the corresponding working conditions, and send it to the monitoring end to monitor the bearing operation status. When the next 10 days of monitoring data is full, the model is updated to obtain the vibration characteristic threshold to achieve dynamic adjustment of the vibration characteristic threshold.
[0060] Furthermore, four vibration signal sample sets {X1, X2, X3, X4} have been introduced in step 1, where X1 and X3 are the vibration signal data sets of the initial and later stages of the bearing operation under the same working conditions. For comparison, the initial vibration characteristic threshold monitoring of this working condition is based on VT 11 ,VT 12 As the benchmark, in the later stage, after sample self-learning, VT 31 ,VT 32 In the second working condition, i.e., the data sample set {X2, X4}, the vibration characteristic threshold monitoring will be from the initial VT 21 ,VT 22 To VT 41 ,VT 42 to update.
[0061] Furthermore, four groups of sample data {X1, X2, X3, X4} were actually tested and the characteristic thresholds were obtained. The results are shown in Table 2. It can be seen that the signal characteristic thresholds have changed, which has further confirmed the effectiveness of the method.
[0062] Table 2. Feature thresholds of self-learning neural network model
[0063] Serial number Bearing condition Speed load <![CDATA[R l ]]> <![CDATA[K l ]]> <![CDATA[VT1]]> <![CDATA[VT2]]> 1 New bearings 1200rpm 0 0.078 3.178 0.117 4.767 2 New bearings 900rpm 0 0.058 3.372 0.087 5.058 3 Wear 0.1mm 1200rpm 0 0.243 7.322 0.3645 10.983 4 Wear 0.1mm 900rpm 0 0.219 8.617 0.3285 12.9255
[0064] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are only used to illustrate the technical solutions of the present invention and cannot be understood as limitations of the present invention. Ordinary technicians in the field can modify and replace the above embodiments within the scope of the present invention without departing from the principles and purpose of the present invention.
Claims
1. A self-learning method for characteristic threshold of bearing vibration monitoring based on generative adversarial network. It is characterized in that The steps include: Step S1: Self-learning feature value selection The root mean square and kurtosis values of vibration signal characteristics are selected as the design basis of characteristic thresholds for bearing condition monitoring; Step S2: Self-learning sample preprocessing The bearing operation status data within a period T is collected by using a vibration acceleration sensor, and the bearing operation status data samples are preprocessed by removing the mean value, and then segmented to obtain multiple short samples x with the same length n. r ; Each short sample marks the operating condition information, including the bearing operating speed, load and sampling time, and stores it in the database; from the database, select a bearing operating status data sample with a period T according to the same operating condition as the training sample set X of the self-learning neural network model train ; According to the operating conditions of the bearing, a periodic sample is collected for each operating condition and stored in the sample database to construct a training sample set of the self-learning neural network model under each operating condition; Step S3: Constructing a feature threshold self-learning neural network model Based on deep learning generative adversarial neural network, the self-learning neural network model is mainly composed of generator G and discriminator D; generator G contains three layers of one-dimensional convolutional layers, Decov1-Decov2-Decov3, and the input of generator G is a z-dimensional Gaussian white noise data that is different from the training sample set in S2. Through the three layers of convolutional layers, data upsampling is achieved, and the output is a data sample x with a length of n g ; The discriminator D is a binary convolutional network, which contains two convolutional pooling layers and three fully connected layers, Cov1+Pool1-Cov2+Pool2-FC1-FC2-FC3. The input of the discriminator D is a data sample x of length n, and the output is converted into a matrix M of length 2 by the Sigmoid function (m 1 ,m 2 ),in Where m 1 Represented as input x is x r The probability of m 2 This means that the input x is x g probability; Step S4: using the training sample set X train Training a self-learning neural network model Using the training sample set constructed in step 2, the self-learning neural network model is trained. The training objective function is shown in formula (1). The training is mainly divided into the training of the discriminator D and the generator G. First, the discriminator D is trained and the training sample set X train Input the discriminator D and use the cross entropy loss function shown in formula (2) to calculate the loss of the discriminator D 1 ; Then the generator G generates an equal amount of data samples x g And input the discriminator D, and use the cross entropy loss function of formula (2) to calculate the loss 2 , merge loss 1 and loss 2 Feedback to the discriminator D and perform gradient update through the Adam function; Then the generator G is trained and the generator G regenerates the data sample x′ g , and the data sample x g Input the label 1 into the discriminator D for identification, that is, calculate the data sample x g Belong to X train The probability of , using the cross entropy loss function of formula (2) to calculate the loss at this time, and then the loss will be passed to the generator G and the gradient of G will be updated through the Adam function; In formula (1), D(x) represents x r The probability of belonging to 1, and D(G(z)) represents x g The probability of belonging to 0; In formula (2), y i represents the label of sample i. For this method, sample x r is marked as 1, and the sample x r The sign of is 0, and p i It represents the probability that sample i is identified as 1; Then, the second iterative training is carried out, that is, the training process of the discriminator D and the generator G is repeated, and the parameters of the discriminator D are updated so that the discriminator D can accurately identify the real sample x r And generate samples x g ; Iteratively train the generator G to generate samples x g Identified as a true sample x by the discriminator D r The probability of generating samples x is continuously increased. g The sample distribution and true sample x r The sample distribution of is getting closer; and so on, after completing multiple self-learning neural network model training, the generator G and discriminator D of the epoch generation are obtained. At this time, the discriminator D cannot distinguish whether the input sample x is the self-learning generated sample x g and the real sample x r ; Step S5: Generate self-learning samples using the trained sample generator G Using the self-learning sample generator G trained in step 4, generate a set of generated samples x from the randomly distributed Gaussian white noise data g , which generates a sample x g That is, the sample is obtained through self-learning of the bearing vibration signal sample, and the vibration signal characteristic threshold will be calculated based on the generated sample; Step S6: Feature threshold calculation Generate sample x obtained in step 5 g As the characteristic threshold self-learning benchmark data under the corresponding working conditions, the sample x is calculated and generated using equations (4) and (5) g The root mean square value RMS and kurtosis value K are then multiplied by 1.2 times as the feature threshold VT of the self-learning sample of the final self-learning neural network model 1 and VT 2 ; Where n is the length of a single sample, and x(i) is the signal value; Step S7: Feature threshold update Select an update period T of the characteristic threshold, repeat steps S1 to S6, and use the vibration signal samples within the period T to learn and update the characteristic threshold of the bearing vibration signal.
