Rolling Bearing Remaining Life Prediction Method Based on Parallel Feature Extraction

Abstract

The present invention discloses a method for predicting the remaining life of a rolling bearing based on parallel feature extraction. First, the vibration signals of the rolling bearing in its full life cycle are collected, and the time-domain feature sequences and synthetic time-frequency diagrams at each acquisition moment during the fault period are extracted therefrom, and the corresponding remaining life percentage labels are determined, thereby constituting training samples. A parallel life prediction network is constructed, and the training samples are used to train the parallel life prediction network. When it is necessary to predict the life of the rolling bearing, the vibration signals of the rolling bearing running to the current moment are collected, the time-domain feature sequences and synthetic time-frequency diagrams at the current moment are extracted, and they are input into the trained parallel life prediction network to obtain the remaining life prediction result. The present invention comprehensively considers the time-frequency domain feature information of the rolling bearing vibration signals and the timing information in the time-domain feature sequences, extracts features in parallel, and improves the performance of predicting the remaining life of the rolling bearing.

Description

Technical Field

[0001] The present invention belongs to the technical field of the remaining life of rolling bearings. More specifically, it relates to a method for predicting the remaining life of rolling bearings based on parallel feature extraction. Background Art

[0002] As one of the most common mechanical components, rolling bearings play a crucial role in almost all rotating mechanical equipment. During the operation of the equipment, the bearings are responsible for bearing and transmitting loads and are constantly subjected to alternating stresses. Therefore, bearings have become one of the most vulnerable mechanical components. According to statistics, among common rotating machinery failure cases, 45%-55% are caused by rolling bearings. During the operation of the equipment, due to factors such as improper installation, overspeed, overload, and poor lubrication, different types of failures such as metal peeling, working surface erosion, and cage fragmentation of rolling bearings will occur, resulting in a decrease in the rotation accuracy of the bearings, an increase in rotational resistance, vibration, and noise. Even the rolling elements are blocked or completely jammed, causing the overall failure of the mechanical equipment.

[0003] Existing methods for predicting the remaining life of equipment can be roughly divided into methods based on physical models and data-driven methods. The life prediction method based on a physical model involves the failure mechanism and damage law of the equipment and requires a large amount of physical knowledge and expert experience to construct a mathematical model describing the degradation process to predict the remaining life of the equipment. However, the complex systems of modern large-scale equipment have greatly increased the difficulty of modeling, making the method based on a physical model have limited applications. The data-driven life prediction method has a lower dependence on failure mechanisms or expert experience and mainly relies on a large amount of historical monitoring data. By using statistical models or machine learning methods to establish the correlation between monitoring data and the degradation process, the remaining life of the equipment can be predicted, which is easier to implement in engineering applications.

[0004] In recent years, with the rapid development of computer technology and artificial intelligence technology, deep learning can process massive amounts of data and extract implicit feature information from it, gradually becoming the mainstream method in the field of equipment life prediction. In the context of intelligent manufacturing, existing remaining life prediction technologies based on deep learning have problems such as weak feature information extraction capabilities of prediction models, complex networks caused by the mixing of feature information and temporal information, and large interference of previous biases on current predictions, which need to be studied and solved. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for predicting the remaining life of rolling bearings based on parallel feature extraction, which comprehensively considers the time-frequency domain feature information of the vibration signal of rolling bearings and the temporal information in the temporal feature sequence, extracts features in parallel, and improves the performance of predicting the remaining life of rolling bearings.

[0006] To achieve the above-mentioned invention purpose, the rolling bearing remaining life prediction method based on parallel feature extraction of the present invention includes the following steps:

[0007] S1: Obtain the full life cycle vibration signals of several rolling bearings according to actual needs. Each full life cycle vibration signal contains the acquisition signals at several acquisition moments, and the data length of each acquisition signal is X;

[0008] S2: Generate training samples respectively according to the full life cycle vibration signals of each rolling bearing to obtain a training sample set. The specific method for generating training samples is as follows:

[0009] S2.1: Denote the number of acquisition moments included in the full life cycle vibration signal as T, and perform initial fault point identification on the full life cycle vibration signal. Denote the initial fault point as T 0 , and take the time before the initial fault point T 0 as the healthy period, and the initial fault point and after as the fault period. The corresponding sample numbers are T H and T F , T H +T F =T;

[0010] S2.2: Determine N time domain features according to the actual situation. Extract N time domain features t i,n , T H +1≤i≤T for each acquisition signal in the fault period of the full life cycle vibration signal; use the following formula for normalization to obtain the normalized time domain feature f i,n :

[0011]

[0012] where, represents the nth time domain feature value of the initial fault point T 0 ;

[0013] Construct the time domain feature vector F i,n =[f i ,f i,1 ,…,f i,2 ,…,f i,N from the N normalized time domain features f

[0014] at each acquisition moment in the fault period; for the jth acquisition moment in the fault period, T H +L≤j≤T, construct a time domain feature sequence P j =[F j-L+1 ,F j-L+2 ,…,F j of length L from the time domain feature vectors of this acquisition moment and the previous L-1 acquisition moments;

[0015] S2.3: For the vibration signals in the whole life cycle, randomly select an acquisition moment during the healthy period, and generate the time-frequency diagram of its acquisition signal as the time-frequency diagram M in the normal state H , and then generate the time-frequency diagram of the j-th acquisition signal during the fault period. Respectively, splice the time-frequency diagram of the j-th acquisition signal during the fault period with the time-frequency diagram M H in the normal state to obtain the synthesized time-frequency diagram M j ;

[0016] S2.4: For the j-th acquisition moment during the fault period in the vibration signals of the whole life cycle, use the following formula to calculate its corresponding remaining life percentage label label j :

[0017]

[0018] S2.5: Take the time-domain feature sequence P j = [F j-L+1 , F j-L+2 , …, F j and the synthesized time-frequency diagram M j of the j-th acquisition moment during the fault period as the input, and the remaining life percentage label label j as the output to form a training sample;

[0019] S3: Construct a parallel life prediction network, including a time series information module, a feature information module, a feature splicing module, and a prediction module, where:

[0020] The time series information module is used to receive the time-domain feature sequence and perform feature extraction to obtain the time series information vector W T and send it to the feature splicing module;

[0021] The feature information module is used to receive the synthesized time-frequency diagram for feature extraction to obtain the feature information vector Q TF and send it to the feature splicing module;

[0022] The feature splicing module is used to splice the received time series information vector Q T and the obtained feature information vector Q TF and send the obtained spliced information vector to the prediction module;

[0023] The prediction module is used to process according to the spliced information vector and output the predicted remaining life percentage;

[0024] S4: Use the time-domain feature sequence and the synthesized time-frequency diagram in each training sample obtained in step S2 as the input, and the corresponding remaining life percentage label as the expected output to train the parallel life prediction network;

[0025] S5: When life prediction of a rolling bearing is required, collect the vibration signal of the rolling bearing at the current moment, and identify the initial fault point of the vibration signal. If the initial fault point does not exist, or the initial fault point exists and the time difference between the current moment and the initial fault point is less than L, then the remaining life percentage at the current moment is 1. If the initial fault point exists and the time difference between the current moment and the initial fault point is greater than or equal to L, then the following method is used for remaining life prediction:

[0026] Extract N time-domain features from the signal collected at the current moment and the signals collected at the previous L - 1 collection moments to form a time-domain feature sequence P'. Then, randomly select a collection moment from the healthy period before the initial fault point, generate the time-frequency diagram of its collected signal as the normal state time-frequency diagram, and then generate the time-frequency diagram of the signal collected at the current moment, and splice it with the normal time-frequency diagram to obtain a composite time-frequency diagram M'. Input the time-domain feature sequence P' and the composite time-frequency diagram M' into the parallel life prediction network trained in step S4 to obtain the remaining life percentage of the rolling bearing.

[0027] The rolling bearing remaining life prediction method based on parallel feature extraction in the present invention first collects the vibration signals of the rolling bearing in the full life cycle, extracts the time-domain feature sequence and the composite time-frequency diagram of each collection moment in the fault period from them, and determines the corresponding remaining life percentage label, thereby constituting a training sample; constructs a parallel life prediction network, and uses the training sample to train the parallel life prediction network; when life prediction of a rolling bearing is required, collect the vibration signal of the rolling bearing running to the current moment, extract the time-domain feature sequence and the composite time-frequency diagram of the current moment, and input them into the trained parallel life prediction network to obtain the remaining life prediction result.

[0028] The present invention has the following beneficial effects:

[0029] 1) In the present invention, the time-frequency domain feature information of the rolling bearing vibration signal and the time sequence information in the time-domain feature sequence are comprehensively considered. Through the effective fusion of the two, the interference of the previous deviation to the current prediction is reduced, thereby improving the accuracy of life prediction.

[0030] 2) Aiming at the problem that directly extracting time sequence information from feature information leads to a large network complexity, the present invention proposes a method of parallel feature extraction, extracts time sequence information from time-domain features, and then combines it with time-frequency domain feature information for life prediction, simplifies the structure of the prediction model, and improves the prediction efficiency. Description of the Drawings

[0031] Figure 1 is a flowchart of the specific implementation manner of the rolling bearing remaining life prediction method based on parallel feature extraction;

[0032] Figure 2It is the flow chart for generating training samples in the present invention;

[0033] Figure 3 It is the structural diagram of the parallel life prediction network in the present invention;

[0034] Figure 4 It is the example diagram for identifying the initial fault point in this embodiment;

[0035] Figure 5 It is the example diagram for synthesizing the time-frequency diagram in this embodiment;

[0036] Figure 6 It is the alternative time-domain characteristic curve diagram of the vibration signal of a certain full life cycle in this embodiment;

[0037] Figure 7 It is the obtained time-domain characteristic curve diagram after screening;

[0038] Figure 8 It is the normalized time-domain characteristic curve diagram;

[0039] Figure 9 It is the comparison diagram of the prediction results and the true labels of the bearings in three test sets in this embodiment. Specific implementation manners

[0040] The following describes the specific implementation manners of the present invention with reference to the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be particularly noted that in the following description, when the detailed descriptions of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0041] Embodiment

[0042] Figure 1 It is the flow chart of the specific implementation manner of the rolling bearing remaining life prediction method based on parallel feature extraction. As Figure 1 shown, the specific steps of the rolling bearing remaining life prediction method based on parallel feature extraction in the present invention include:

[0043] S101: Obtain the vibration signals of the rolling bearings:

[0044] Obtain the full life cycle vibration signals of several rolling bearings under the preset working conditions according to actual needs. Each full life cycle vibration signal contains the acquisition signals at several acquisition moments, and the data length of each acquisition signal is X. In this embodiment, the vibration signal of the rolling bearing adopts the horizontal vibration acceleration sensor signal.

[0045] S102: Generate training samples:

[0046] Generate training samples respectively according to the full life cycle vibration signals of each rolling bearing to obtain a training sample set. Figure 2This is the flow chart for generating training samples in the present invention. As Figure 2 shown, the specific steps for generating training samples in the present invention include:

[0047] S201: Identify the initial fault point:

[0048] Record the number of acquisition times included in the vibration signal of the entire life cycle as T, and identify the initial fault point (i.e., the initial sample serial number corresponding to when the rolling bearing starts to fail) for the vibration signal of the entire life cycle. Denote the initial fault point as T 0 , and take the time before the initial fault point T 0 as the healthy period, and the initial fault point and after as the fault period. The corresponding number of samples are T H and T F , T H +T F =T.

[0049] In this embodiment, the specific method for identifying the initial fault point is: Calculate the root mean square value with a length of T based on the vibration signal of the entire life cycle to form a root mean square value curve, and use the 3-sigma criterion to identify the initial fault point T 0 .

[0050] S202: Extract time domain features:

[0051] Determine N time domain features according to the actual situation. For each acquisition signal in the fault period of the vibration signal of the entire life cycle, extract N time domain features t i,n , T H +1≤i≤T. Normalize using the following formula to obtain the normalized time domain feature f i,n :

[0052]

[0053] where represents the nth time domain feature value of the initial fault point T 0 .

[0054] For each acquisition time N in the fault period, the normalized time domain features f i,n constitute a time domain feature vector F i =[f i,1 ,f i,2 ,…,f i,N .

[0055] For the jth acquisition time in the fault period, T H +L≤j≤T, construct a time domain feature sequence P with a length of L from the time domain feature vectors of this acquisition time and the previous L-1 acquisition times j =[F j-L+1 ,Fj-L+2 , …, F j .

[0056] In order to make the time-domain features better represent the operating characteristics of rolling bearings, in this embodiment, a screening method for time-domain features is designed. The specific method is as follows:

[0057] Set N' alternative time-domain features, where N' > N, and obtain the evaluation score Score of each alternative time-domain feature respectively. The specific method is as follows: Arbitrarily select a vibration signal in the full life cycle, and extract the feature value a of the n'-th alternative time-domain feature of each acquisition signal in the vibration signal of the full life cycle n′ , 1 ≤ k ≤ T. The eigenvalue difference difference between the healthy period and the faulty period of the n'-th alternative time-domain feature is calculated using the following formula k,n′ : n′

[0058]

[0059] The fluctuation quantization value fluctutation of the n'-th alternative time-domain feature is calculated using the following formula n′ :

[0060] fluctutation n′ = σ H,n′ / σ F,n′

[0061] where σ H,n′ , σ F,n′ respectively represent the standard deviations of the eigenvalue of the n'-th alternative time-domain feature in the healthy period and the faulty period.

[0062] The trend quantization value trend of the n'-th alternative time-domain feature is calculated using the following formula n′ :

[0063]

[0064] where σ B represents the standard deviation of the sequence B.

[0065] Then, the evaluation score Score of the n'-th alternative time-domain feature is calculated using the following formula n′ :

[0066] Score n′ = difference n′ - 3 * fluctutation n′ + 2 * trend n′

[0067] ​Then, arrange the N' alternative time-domain features of the full life-cycle vibration signal in descending order according to the evaluation scores, and select the first N alternative time-domain features as the finally used time-domain features.

[0068] S203: Generate a synthetic time-frequency diagram:

[0069] For the full life-cycle vibration signal, randomly select an acquisition moment during the healthy period, and generate the time-frequency diagram of its acquisition signal as the time-frequency diagram M in the normal state. H , and then generate the time-frequency diagram of the j-th acquisition signal in the fault period. Respectively, splice the time-frequency diagram of the j-th acquisition signal in the fault period with the time-frequency diagram M in the normal state. H to obtain the synthetic time-frequency diagram M. j .

[0070] S204: Determine the remaining life percentage label:

[0071] For the j-th acquisition moment in the fault period of the full life-cycle vibration signal, calculate its corresponding remaining life percentage label label using the following formula. j :

[0072]

[0073] S205: Construct a training sample:

[0074] Take the time-domain feature sequence P j = [F j-L+1 , F j-L+2 , …, F j of the j-th acquisition moment in the fault period and the synthetic time-frequency diagram M j as the input, and the remaining life percentage label label j as the output to form a training sample.

[0075] S103: Construct a parallel life prediction network:

[0076] In order to realize the prediction of the remaining life of the rolling bearing, a parallel life prediction network that comprehensively uses time-frequency domain features and time-domain features is constructed in the present invention. Figure 3 is the structure diagram of the parallel life prediction network in the present invention. As Figure 3 shown, the parallel life prediction network in the present invention includes a timing information module, a feature information module, a feature splicing module, and a prediction module. Next, each module will be described in detail.

[0077] The timing information module is used to receive the time-domain feature sequence and perform feature extraction to obtain the timing information vector W T and send it to the feature splicing module. As Figure 3As shown in the figure, in this embodiment, the timing information module includes a first linear layer, an LSTM (Long Short Term Memory) network, and a second linear layer, where:

[0078] The first linear layer is used to process the time-domain feature sequence by using the Relu activation function and send the processed feature sequence to the LSTM network.

[0079] The LSTM network is used to extract the timing features from the received feature sequence and send them to the second linear layer. The LSTM network is a commonly used neural network, and its specific structure and working process will not be elaborated here.

[0080] The second linear layer is used to process the timing features by using the Relu activation function to obtain the timing information vector Q T and send it to the feature concatenation module.

[0081] The feature information module is used to receive the synthesized time-frequency map for feature extraction to obtain the feature information vector Q TF and send it to the feature concatenation module. As Figure 3 shown, in this embodiment, the feature information module includes a first convolution module, a second convolution module, a feature stacking module, a multi-head attention module, and a linear layer, where:

[0082] The first convolution module is used to extract features from the synthesized time-frequency map by using convolution operations to obtain a feature G of size W×H×A 1 , where W×H represents the spatial scale of the feature G 1 and A represents the number of channels of the feature G 1 , and send the feature G 1 to the feature stacking module.

[0083] The second convolution module is used to extract features from the synthesized time-frequency map by using convolution operations with different scales from the first convolution module to obtain a feature G of size W×H×B 2 , where B represents the number of channels of the feature G 2 , and send the feature G 2 to the feature stacking module.

[0084] In this embodiment, both the first convolution module and the second convolution module adopt a 4-layer cascade structure, and each layer is composed of a convolution layer and a max pooling layer. Table 1 is the parameter configuration table of the first convolution module and the second convolution module in this embodiment.

[0085]

[0086] Table 1

[0087] The feature stacking module is used to stack the feature G 1 , the feature G2 Stack them in the channel dimension to obtain multi-scale features and send them to the multi-head attention module.

[0088] The multi-head attention module is used to process the multi-scale features by adopting the multi-head attention mechanism, and send the obtained feature vectors to the linear layer.

[0089] The linear layer is used to reduce the dimension of the received feature vectors to obtain the feature information vector Q TF and send it to the feature concatenation module.

[0090] From the above description, it can be seen that the feature information module of this embodiment adopts a multi-scale convolutional neural network - multi-head attention mechanism, comprehensively extracts features from the synthetic time-frequency diagram by using two different convolutional modules, and uses the multi-head attention mechanism to assign weights, which can obtain richer and more effective feature information, thereby enhancing the representation ability of the feature information vector for the vibration signal features, and further improving the remaining life accuracy.

[0091] The feature concatenation module is used to concatenate the received time series information vector Q T and the feature information vector Q TF and send the obtained concatenated information vector to the prediction module.

[0092] The prediction module is used to process according to the concatenated information vector and output the predicted remaining life percentage. As Figure 3 shown, in this embodiment, the prediction module includes a first linear layer, a second linear layer and a third linear layer, where:

[0093] The first linear layer is used to process the concatenated information vector by adopting the ReLU activation function, and send the obtained features to the second linear layer.

[0094] The second linear layer is used to process the received features by adopting the LeakyReLU activation function, and send the obtained features to the third linear layer.

[0095] The third linear layer is used to process the received features by adopting the Sigmoid activation function to obtain the predicted remaining life percentage.

[0096] S104: Train the parallel remaining life prediction network:

[0097] Use the time domain feature sequence and synthetic time-frequency diagram in each training sample obtained in step S102 as inputs, and the corresponding remaining life percentage label as the expected output to train the parallel remaining life prediction network.

[0098] S105: Predict the remaining life of the rolling bearing:

[0099] When it is necessary to predict the life of a rolling bearing, the vibration signal of the rolling bearing running to the current moment is collected, and the initial fault point of the vibration signal is identified. When the initial fault point does not exist, or the initial fault point exists and the time difference between the current moment and the initial fault point is less than L, the remaining life percentage at the current moment is 1. When the initial fault point exists and the time difference between the current moment and the initial fault point is greater than or equal to L, the following method is used for remaining life prediction:

[0100] Extract N time-domain features from the signal collected at the current moment and the signals collected at the previous L-1 acquisition moments to form a time-domain feature sequence P'. Then, randomly select an acquisition moment from the healthy period before the initial fault point, generate the time-frequency diagram of its acquisition signal as the normal state time-frequency diagram, and then generate the time-frequency diagram of the signal collected at the current moment, and splice it with the normal time-frequency diagram to obtain a composite time-frequency diagram M'. Input the time-domain feature sequence P' and the composite time-frequency diagram M' into the parallel life prediction network trained in step S104 to obtain the remaining life percentage of the rolling bearing.

[0101] Embodiment

[0102] To better illustrate the technical solution and technical effect of the present invention, a specific example is used to analyze and explain the working process and technical effect of the present invention. Bearing failure is a typical failure in rotating machinery. Therefore, in this embodiment, the bearing accelerated full-life vibration signal data of the XJTU-SY bearing dataset of Xi'an Jiaotong University is used for experimental testing.

[0103] In the bearing accelerated life test of Xi'an Jiaotong University used in this embodiment, a total of 3 types of working conditions are designed, and 5 rolling bearings are subjected to accelerated life tests under each working condition. Table 2 shows the specific test working conditions based on the bearing data of Xi'an Jiaotong University in this embodiment.

[0104]

[0105] Table 2

[0106] In this embodiment, the vibration acceleration signal dataset is collected through a unidirectional acceleration sensor, the sampling frequency is 25.6 kHz, the sampling interval is 1 min, and the sampling duration each time is 1.28 s. It contains the full-cycle life vibration signals of 15 rolling bearings under 3 working conditions, and clearly marks information such as the total number of samples, basic rated life, actual life, and fault location of each failed bearing.

[0107] Generate training samples according to the full-life cycle vibration signals of each rolling bearing. First, identify the initial fault point. Figure 4 This is an example diagram of initial fault point identification in this embodiment. Table 3 is the initial fault point information table of each rolling bearing in this embodiment.

[0108]

[0109] Table 3

[0110] As shown in Table 3, due to their particularity, the initial fault points of the two rolling bearings are not given, and the signals of these two rolling bearings are discarded in the subsequent process.

[0111] Next, the wavelet transform is used to construct the time-frequency diagram, and the synthetic time-frequency diagram of each acquisition moment in the fault period is generated. Figure 5 It is an example diagram of the synthetic time-frequency diagram in this embodiment.

[0112] For the time-domain features, 16 alternative time-domain features are set in this embodiment. Table 4 is the information table of the alternative time-domain features in this embodiment.

[0113]

[0114] Table 4

[0115] The evaluation scores of 16 alternative time-domain features are obtained according to the rolling bearing Bearing1_1 in the training set. Table 5 is the evaluation score table of the alternative time-domain features in this embodiment.

[0116]

[0117] Table 5

[0118] As shown in Table 5, the top 6 features with the highest comprehensive evaluation scores are selected as the final time-domain features used, namely root mean square, variance, standard deviation, absolute mean, root mean square mean, and average power. Figure 6 It is the curve diagram of the alternative time-domain features of a certain vibration signal in the whole life cycle in this embodiment. Figure 7 It is the curve diagram of the time-domain features obtained by screening. Figure 8 It is the curve diagram of the time-domain features after normalization.

[0119] The parallel feature extraction network model in this embodiment is built by Pytorch, the initial learning rate is 0.00001, the optimization algorithm selects the Adam algorithm, and the loss function selects MSE (Mean Square Error). The parallel feature extraction network model is trained with the training set, and after training is completed, the test set is input into the parallel feature extraction network model for remaining life prediction. Figure 9 It is the comparison diagram between the predicted labels and the true labels of some rolling bearings in the test set obtained by the present invention in this embodiment. As Figure 9 shown, the predicted labels obtained by the present invention are relatively close to the true labels.

[0120] For the purpose of comparing technical effects, in this embodiment, two remaining useful life prediction methods, namely the RUL prediction model based on LSTM and CNN (CNN-LSTM) and the RUL prediction model based on MSCNN and BiGRU (MSCNN-BiGRU), are used as comparative methods to conduct comparative experiments with the present invention. In the comparative experiments, the data of rolling bearings Bearing1_1, Bearing1_2, and Bearing2_1 are used as training samples to train the corresponding prediction models, and other rolling bearings are used as test bearings to input the trained prediction models for predicting the remaining useful life percentage. The evaluation indicators selected are RMSE (Root Mean Square Error) and MAE (Mean Absolute Error). Table 6 is a comparison table of the diagnostic results of the present invention and the comparative methods in this embodiment.

[0121]

[0122] Table 6

[0123] As shown in Table 6, the remaining useful life prediction error of the present invention is generally smaller than that of the two comparative models. The average error index is between 0.184 and 0.156, which is about 23% lower than the error of the RUL prediction model based on LSTM and CNN, and about 15% lower than the error of the RUL prediction model based on MSCNN and BiGRU. It can be seen that the present invention is effective in improving the performance of predicting the remaining useful life of rolling bearings.

[0124] Although the above describes the illustrative specific embodiments of the present invention for the convenience of those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

Claims

1. A method for predicting the remaining useful life of rolling bearings based on parallel feature extraction, characterized in that, it includes the following steps: S1: Obtain the vibration signals of several rolling bearings during their full life cycles according to actual needs. Each vibration signal during the full life cycle contains acquisition signals at several acquisition moments, and the data length of each acquisition signal is X; S2: Generate training samples respectively according to the vibration signals of each rolling bearing during its full life cycle to obtain a training sample set. The specific method for generating training samples is as follows: S2.1: Denote the number of acquisition times included in the vibration signal of the entire life cycle as T, and perform initial fault point identification on the vibration signal of the entire life cycle. Denote the initial fault point as T 0 , and take the time before the initial fault point T 0 as the healthy period, and the initial fault point and after as the fault period. The corresponding sample numbers are T H and T F , T H + T F = T; S2.2: Determine N time-domain features according to the actual situation, and extract N time-domain features t for each acquired signal in the fault cycle of the vibration signal in the whole life cycle i,n , T H +1 ≤ i ≤ T; The screening method for N time-domain features is as follows: Set N' alternative time-domain features, where N' > N, and obtain the evaluation score Score for each alternative time-domain feature respectively n′ , and the specific method is as follows: Arbitrarily select a vibration signal in the whole life cycle, and extract the eigenvalue a of the n'-th alternative time-domain feature of each acquisition signal in the vibration signal of the whole life cycle k,n′ , 1 ≤ k ≤ T, and calculate the eigenvalue difference difference between the healthy period and the faulty period of the n'-th alternative time-domain feature using the following formula n′ : The fluctuation quantization value fluctutation of the n'-th alternative time-domain feature is calculated using the following formula n′ :[[]]END]] fluctutation n′ = σ H,n′ / σ F,n′ Among them, σ H,n′ and σ F,n′ respectively represent the standard deviations of the feature values of the n'-th alternative time-domain feature in the healthy period and the faulty period; The trend quantization value trend of the n'-th alternative time-domain feature is calculated using the following formula n′ :[[]]END]] Among them, σ B represents the standard deviation of sequence B; Then, the evaluation score Score of the n'-th alternative time-domain feature is calculated using the following formula n′ :[[]] Score n′ = difference n′ - 3 * fluctuation n′ + 2 * trend n′ Then, arrange the N' alternative time-domain features of the vibration signal during the full life cycle in descending order according to the evaluation scores, and select the first N alternative time-domain features as the finally used time-domain features; Normalization is performed using the following formula to obtain the normalized time-domain feature f i,n : Among them, t T0,n represents the nth time-domain eigenvalue of the initial fault point T 0 ; Normalize the N time-domain features f at each acquisition moment during the fault cycle i,n to form a time-domain feature vector F i = [f i,1 , f i,2 , …, f i,N ; For the j-th acquisition moment during the fault period, T H +L ≤ j ≤ T, construct a time-domain feature sequence P of length L from the time-domain feature vectors of this acquisition moment and the previous L - 1 acquisition moments j =[F j-L+1 ,F j-L+2 ,…,F j ; S2.3: For the vibration signals in the whole life cycle, randomly select an acquisition moment during the healthy period, and generate the time-frequency diagram of its acquisition signal as the normal state time-frequency diagram M H , then generate the time-frequency diagram of the j-th acquisition signal during the fault period, and splice the time-frequency diagram of the j-th acquisition signal during the fault period and the normal state time-frequency diagram M H respectively to obtain the composite time-frequency diagram M j ; S2.4: For the j-th acquisition moment in the fault period of the vibration signal in the whole life cycle, calculate its corresponding remaining life percentage label label using the following formula j : S2.5: Use the time-domain feature sequence P at the j-th acquisition moment during the fault period j = [F j-L+1 , F j-L+2 , …, F j , and the synthetic time-frequency diagram M j as the input, and the remaining life percentage label label j as the output to form a training sample; S3: Construct a parallel remaining useful life prediction network, including a time series information module, a feature information module, a feature splicing module, and a prediction module, where: The timing information module is used to receive the time-domain feature sequence and perform feature extraction to obtain the timing information vector W T and send it to the feature splicing module; The feature information module is used to receive the synthesized time-frequency map for feature extraction to obtain the feature information vector Q TF and send it to the feature splicing module; The feature splicing module is used to splice the received timing information vector Q T and the obtained feature information vector Q TF to perform splicing, and send the obtained spliced information vector to the prediction module; The prediction module is used to process according to the spliced information vector and output the predicted remaining useful life percentage; S4: Use the time-domain feature sequence and the synthesized time-frequency diagram in each training sample obtained in step S2 as inputs, and the corresponding remaining useful life percentage label as the expected output to train the parallel remaining useful life prediction network; S5: When it is necessary to predict the remaining useful life of a rolling bearing, collect the vibration signal of the rolling bearing at the current moment, and identify the initial fault point of the vibration signal. If the initial fault point does not exist, or the initial fault point exists and the time difference between the current moment and the initial fault point is less than L, then the remaining useful life percentage at the current moment is 1. If the initial fault point exists and the time difference between the current moment and the initial fault point is greater than or equal to L, then the following method is used for remaining useful life prediction: Extract N time-domain features from the acquisition signal at the current moment and the acquisition signals at the previous L-1 acquisition moments to form a time-domain feature sequence P'. Then randomly select an acquisition moment from the healthy period before the initial fault point, generate the time-frequency diagram of its acquisition signal as the normal state time-frequency diagram, and then generate the time-frequency diagram of the acquisition signal at the current moment, and splice it with the normal time-frequency diagram to obtain a synthesized time-frequency diagram M'. Input the time-domain feature sequence P' and the synthesized time-frequency diagram M' into the parallel remaining useful life prediction network trained in step S4 to obtain the remaining useful life percentage of the rolling bearing.

2. The method for predicting the remaining useful life of rolling bearings according to claim 1, characterized in that, the vibration signal of the rolling bearing uses the signal of a horizontal vibration acceleration sensor.

3. The method for predicting the remaining useful life of rolling bearings according to claim 1, characterized in that, The specific method for identifying the initial fault point is as follows: Calculate the root mean square value with a length of T from the vibration signal in the whole life cycle to form a root mean square value curve, and use the 3-sigma criterion to identify the initial fault point T 0 .

4. The method for predicting the remaining useful life of rolling bearings according to claim 1, characterized in that, in step S3, the time series information module includes a first linear layer, an LSTM network, and a second linear layer, where: The first linear layer is used to process the time-domain feature sequence by using the Relu activation function and send the processed feature sequence to the LSTM network; The LSTM network is used to extract the time series features from the received feature sequence and send them to the second linear layer; The second linear layer is used to process the temporal features with the Relu activation function to obtain the temporal information vector Q T and send it to the feature splicing module.

5. The method for predicting the remaining useful life of rolling bearings according to claim 1, characterized in that, In the step S3, the feature information module includes a first convolution module, a second convolution module, a feature stacking module, a multi-head attention module, and a linear layer, where: The first convolution module is used to extract features from the synthesized time-frequency map through convolution operations, obtaining a feature G of size W×H×A 1 , where W×H represents the spatial scale of the feature G 1 , and A represents the number of channels of the feature G 1 . The feature G 1 is sent to the feature stacking module; The second convolution module is used to perform feature extraction on the synthesized time-frequency map by using a convolution operation with a different scale from that of the first convolution module, and obtain a feature G of size W×H×B 2 , where B represents the number of channels of the feature G 2 . The feature G 2 is sent to the feature stacking module; The feature stacking module is used to stack feature G 1 and feature G 2 in the channel dimension to obtain multi-scale features and send them to the multi-head attention module; The multi-head attention module is used to process the multi-scale features by using the multi-head attention mechanism and send the obtained feature vectors to the linear layer; The linear layer is used to reduce the dimension of the received feature vector to obtain the feature information vector Q TF and send it to the feature splicing module.

6. The rolling bearing remaining life prediction method according to claim 1, characterized in that in the step S3, the prediction module includes a first linear layer, a second linear layer, and a third linear layer, where: The first linear layer is used to process the concatenated information vector by using the ReLU activation function and send the obtained features to the second linear layer; The second linear layer is used to process the received features by using the LeakyReLU activation function and send the obtained features to the third linear layer; The third linear layer is used to process the received features by using the Sigmoid activation function to obtain the predicted remaining life percentage.

Images