A bearing residual life prediction method based on ECAN tet-TCN and self-attention-biGRU mechanism

Abstract

The application discloses a bearing residual life prediction method based on ECAN Tet-TCN and a Self-Attention-BiGRU mechanism, mainly including two parts: a training stage and a test stage. In the training stage, according to the data detected under different working conditions of different mechanical equipment, the data is divided into training set vibration signal data and test set vibration signal data, and then a prediction model is constructed. Secondly, the vibration signal data of the training set and the test set are batch standardized and processed. The characteristics are input into the model, the RUL of the bearing is predicted and verified, and then the training set signal and the test set signal are one-to-one mapped through the prediction model after the test stage processing, and the bearing life prediction is completed. Compared with the existing method, the method is helpful to enhance the robustness of the model and improve the accuracy of the bearing RUL prediction. At the same time, the self-attention mechanism is added after the BiGRU layer, the attention to irrelevant information is reduced, the dependence on external information is reduced, and the model information overload problem is solved.

Description

Technical Field

[0001] This invention belongs to the field of bearing remaining life prediction, and particularly relates to a bearing remaining life prediction method based on ECANTet-TCN and Self-Attention-BiGRU mechanisms. Background Technology

[0002] As a key component of rotating machinery, the health of bearings directly affects the overall working condition of the machine. Bearing failure can severely impact the safety of the mechanical system and cause significant economic losses. However, rolling bearings often operate in harsh environments such as aircraft engines, making them prone to failure and potentially leading to engine malfunction. Therefore, the application of Remaining Useful Life (RUL) prediction technology in bearings has always been a key research area.

[0003] However, due to the influence of various factors, the degradation process of bearings is complex and difficult to predict accurately. Traditional methods for predicting the remaining life of bearings often rely on simple statistical models, which involve cumbersome information extraction and fail to fully tap into the potential information in the data, resulting in low accuracy in bearing life prediction. Summary of the Invention

[0004] In view of this, the present invention provides a bearing remaining life prediction method based on ECANTet-TCN and Self-Attention-BiGRU mechanisms, which can solve the problems of difficult data feature extraction and low prediction accuracy in existing bearing life prediction models.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This invention discloses a bearing remaining life prediction method based on ECANTet-TCN and Self-Attention-BiGRU mechanisms, comprising the following steps:

[0007] 1. Based on the data detected under different working conditions of different mechanical equipment, it is divided into training set vibration signal data and test set vibration signal data;

[0008] 2. Construct a prediction model based on the ECANTet-TCN and Self-Attention-BiGRU mechanisms;

[0009] 3. Perform batch standardization processing on the vibration signal data input into the training and test sets;

[0010] 4. Use ECANTet-TCN to mine data features from standardized data, while enhancing useful features and suppressing useless features;

[0011] 5. Utilize multi-layer TCN and attention mechanisms to mine deep temporal features of data;

[0012] 6. Input the features after multiple filterings into the Self-Attention BiGRU network model to predict the RUL of the bearing, and then add a Dropout layer to build the model and test it on the test set.

[0013] 7. Collect vibration signals that meet the requirements of the testing phase as the basis for life prediction.

[0014] 8. Perform batch standardization processing on the vibration signals of the test set;

[0015] 9. Use ECANTet-TCN to extract features from the batch of standardized test set vibration signals and establish a feature extraction set;

[0016] 10. The model established by ECANTet-TCN and Self-Attention-BiGRU mechanism maps the training set signals and test set signals one by one to complete the prediction of bearings.

[0017] 11. The methods based on ECANTet-TCN network and Self-Attention-BiGRU are evaluated using three evaluation metrics: Var, Mon, and Corr functions.

[0018] The beneficial effects of this invention are as follows:

[0019] Compared with existing methods, the bearing remaining life prediction method of the present invention constructs a prediction model based on ECANTet-TCN and Self-Attention-BiGRU mechanisms. ECANTet-TCN is used to mine data features from standardized data, while enhancing useful features and suppressing useless features. Multi-layer TCN and attention mechanisms are used to fuse and mine deep temporal features of the data. The features after multiple screenings are input into the Self-Attention BiGRU network model to predict the bearing's RUL, which helps to enhance the robustness of the model and improve the accuracy of bearing RUL prediction. At the same time, a self-attention mechanism is added after the BiGRU layer to reduce attention to irrelevant information, reduce dependence on external information, and solve the problem of model information overload.

[0020] Secondly, ECANet demonstrates strong capabilities in mining deep-level data features and reconstructing data characteristics, while the TCN network can better uncover the inherent temporal features of degradation trends and the correspondence between training features and the entire lifetime, ensuring prediction accuracy. It is necessary to combine the two into an ECANet-TCN model to more effectively improve data feature analysis and lifetime prediction capabilities. The TCN network is chosen to utilize dilated convolutions to address the long-range dependency problem commonly found in temporal models. Attached Figure Description

[0021] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0022] Figure 1 This is a flowchart illustrating the prediction process of the present invention;

[0023] Figure 2 The figures show the prediction results of the embodiments and comparative models of the present invention. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. Unless otherwise specified, the techniques used in the embodiments are conventional means well known to those skilled in the art.

[0025] It should be noted that, unless otherwise stated, the technical or scientific terms used in this invention should be understood in their ordinary sense by those skilled in the art. The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0026] like Figure 1 The diagram shows a flowchart of a bearing remaining life prediction method based on ECANTet-TCN and Self-Attention-BiGRU mechanisms provided by this invention. The method includes the following steps:

[0027] Step 1: Based on the data detected under different working conditions of different mechanical equipment, divide it into training set vibration signal data and test set vibration signal data;

[0028] Step 2: Construct a prediction model based on the ECANTet-TCN and Self-Attention-BiGRU mechanisms;

[0029] Step 3: Perform batch standardization processing on the vibration signal data input from the training set and the test set;

[0030] Step 4: Use ECANTet-TCN to mine data features from the standardized data, while enhancing useful features and suppressing useless features;

[0031] Step 5: Utilize multi-layer TCN and attention mechanisms to fuse and mine deep temporal features of the data;

[0032] Step 6: Input the features after multiple filterings into the Self-Attention BiGRU network model to predict the RUL of the bearing, and then add a Dropout layer to build the model and test it on the test set.

[0033] Step 7: Collect vibration signals that meet the requirements of the testing phase as the basis for life prediction.

[0034] Step 8: Perform batch standardization processing on the vibration signals of the test set;

[0035] Step 9: Extract features from the batch-standardized test set vibration signals using ECANTet-TCN to establish a feature extraction set;

[0036] Step 10: The model established by ECANTet-TCN and Self-Attention-BiGRU mechanism maps the training set signals and test set signals one by one to complete the prediction of the bearing.

[0037] Step 11: The methods based on the ECANTet-TCN network and Self-Attention-BiGRU are evaluated using three evaluation metrics: Var, Mon, and Corr functions.

[0038] The specific implementation methods of the above steps are described in detail below:

[0039] In step 1, the data is collected using an accelerometer to collect vibration signals in the horizontal and vertical directions. The sampling frequency is 25.6 kHz, and data is recorded every 10 seconds for a duration of 0.1 seconds. The vibration data collected each time are summarized. Cross-validation is used in the experiment to divide the data into test data and training data based on different test data, which are then superimposed on the input model for training.

[0040] In step 2, ECANet first uses global average pooling to aggregate the input features U∈R. W×H×C The spatial information of each channel. The expression for the average pooling operation is:

[0041]

[0042] Then, a one-dimensional convolution pair P with a kernel size of K is used. GA (U) Perform convolution calculations to quickly extract the feature relationships of K local information. Use the sigmoid function to calculate the activation values ​​of the one-dimensional convolution output, obtaining the weight values ​​ω∈R representing the local relationships and importance of the feature channels. 1×1×C The expressions for Sigmoid and ω are:

[0043]

[0044] ω=Sigmoid{C1D k [P GA (U)]}

[0045] Finally, in order to re-encode each channel feature of U, ω and U are multiplied one by one to obtain weighted features. Important features are assigned larger weights to achieve enhancement, while ineffective features are assigned smaller weights to achieve autonomous inhibition.

[0046] TCN utilizes dilated convolution to address the long-range dependency problem commonly found in temporal models. Dilated convolution has a larger receptive field compared to traditional convolution. The expression for calculating F using dilated convolution is:

[0047]

[0048] In the formula: F(x) s (x) represents the input x at time s during the dilated convolution calculation. s The network output is correct; k is the kernel size; d is the dilation coefficient; sd i Let i be the sequence corresponding to the elements in the convolution kernel, i∈(0,1,…,k-1).

[0049] The self-attention mechanism involves two steps. The first step is to calculate the attention weights between any two vectors in the input sequence. The second step is to calculate the weighted average of the input sequence based on these attention weights. The specific calculation method is as follows:

[0050] Q = XW q

[0051] K = XW k

[0052] V = XW ν

[0053]

[0054] In the formula: Q is the query matrix; K is the key matrix; V is the value matrix; dim is the dimension of Q, K and V.

[0055] The Bidirectional Gated Cyclic Unit (BiGRU) captures information through backpropagation, which can better capture the potential timing relationships between signals.

[0056] In step 3, the input data is standardized, mapping it to the range of 0 to 1 to ensure data stability. This method uses the Z-score standardization criterion to process the selected features. The Z-score standardization criterion formula is shown below:

[0057] Z=(X-μ) / σ

[0058] In the formula: Z is the Z-score, X is the original data, μ is the mean of the original data, and σ is the standard deviation of the original data.

[0059] In steps 4, 5, 6, and 7, the steps are executed in sequence, and calculations are performed according to the formula in step 2.

[0060] In step 8, follow the same procedure as in step 3.

[0061] In step 9, the steps for feature extraction are as follows: 1) Feature selection: time-domain features of the vibration signal (such as mean, variance, peak-to-peak value), frequency-domain features (spectral peak value), time-frequency features (wavelet transform coefficients), and other relevant features (temperature, pressure), etc.; 2) Feature extraction: implemented using wavelet transform method; 3) Feature engineering: based on feature extraction, standardization and normalization processing are performed; 4) Feature combination: different features may have certain correlations, and they are combined through feature cross and feature merging to generate new features; 5) Feature screening: the feature set is screened using feature selection algorithm to remove features that have little impact on the prediction effect.

[0062] In step 10, proceed as required.

[0063] In step 11, three evaluation feature indicators—correlation (Corr(f,t)), monotonicity (Mon(f)), and variance (Var(f))—are used to screen features that can effectively reflect the degradation process.

[0064] Before evaluating the feature set, the center-shifting method is used to treat feature f as a stochastic process, dividing it into a trend component, f T f represents the average trend and the random component. R The residual is represented by the following formula:

[0065] f(t k )=f T (t k )+f R (t k )

[0066] In the formula: f(t) k (t) represents time t k The degenerative characteristics.

[0067] The above three evaluation indicators can be derived from the following formula:

[0068]

[0069] In the formula: k is the total number of observations; h(t) is the step function.

[0070] The evaluation and selection methods are as follows:

[0071] Ce=0.2Var(f)+0.5Mon(f)+0.3Corr(f,t)

[0072] The weight value before each feature evaluation indicates the importance of that indicator. The weight values ​​range from (0,1). During the process of load degradation, cumulative damage occurs; secondly, in order to measure the dispersion of degradation features, the variance of degradation features should also be considered. Therefore, 0.2, 0.5, and 0.3 are selected as the weight values ​​for Var, Mon, and Corr, respectively.

[0073] The prediction process in this invention mainly includes two parts: a training phase and a testing phase. In the training phase, firstly, based on the data detected by different mechanical equipment under different operating conditions, the data is divided into training set vibration signal data and test set vibration signal data. Then, a prediction model is constructed based on the ECANTet-TCN and Self-Attention-BiGRU mechanisms. Secondly, the vibration signal data input to the training and test sets are batch standardized. Thirdly, ECANTet-TCN is used to mine data features from the standardized data, while enhancing useful features and suppressing useless features. Fourthly, multi-layer TCN and attention mechanisms are used to fuse and mine deep temporal features of the data. Finally, the features after multiple rounds of filtering are input into Self-Attention. A BiGRU network model is used to predict the bearing's roughness and lifespan (RUL). A Dropout layer is then added to build the model, which is then validated on a test set. During the testing phase, firstly, vibration signals consistent with the testing requirements are collected as the basis for lifespan prediction. Secondly, the vibration signals from the test set are batch-standardized. Next, ECANTet-TCN is used to extract features from the batch-standardized test set vibration signals, establishing a feature extraction set. Then, a model built using ECANTet-TCN and Self-Attention-BiGRU mechanisms maps the training set signals to the test set signals to complete the bearing lifespan prediction. Finally, Var, Mon, and Corr functions are used to evaluate the results based on the ECANTet-TCN network and Self-Attention-BiGRU. Compared with existing methods, this method helps enhance the model's robustness and improve the accuracy of bearing RUL prediction. Furthermore, the addition of a self-attention mechanism after the BiGRU layer reduces attention to irrelevant information, lowers dependence on external information, and solves the problem of model information overload.

[0074] The following is an example of an experiment:

[0075] In the experimental setup, the collected data was a bearing life-cycle dataset, including vibration data under variable speed conditions (0-8000 rpm) and constant radial load (100 N). The sampling frequency was 20 kHz, and data was collected every 1 hour for 2 minutes each time. Three sets were selected as the training set and one set as the test set. To demonstrate the superiority of this method, it was compared with traditional ARIMA models, LSTM networks, TCN models, and BiGRU models without self-attention mechanisms. The evaluation metrics used were: root mean square error (RMSE), mean absolute percentage error (MAPE), and the proposed Var (variance), Mon (monotonicity), and Corr (correlation) comprehensive scores.

[0076] In the data preprocessing, the raw vibration signal was divided into time windows of 10 seconds each. Within each window, 12-dimensional features were extracted, including time domain (mean, peak-to-peak value, kurtosis), frequency domain (centroid of the spectrum, root mean square frequency), and time-frequency domain (wavelet packet energy entropy). Z-score normalization was then applied. Labels were generated based on the bearing failure threshold (vibration amplitude exceeding 20 m / s²). 2 The Remaining Lifetime (RUL) is defined as the percentage of time from the current time to the time of failure, and normalized to the interval [0,1].

[0077] During model training, the ECANet-TCN layer dilation coefficient d = 2, the convolution kernel size k = 3, the BiGRU hidden layer dimension was 64, the number of self-attention heads was 4, the dropout rate was 0.3, the batch size was 32, the initial learning rate was 0.001, the Adam optimizer was used, 5-fold cross-validation was used, early stopping (patience = 10) was used to prevent overfitting, and the maximum number of training epochs was 200.

[0078] The comparative test data are as follows:

[0079] Model RMSE MAPE (%) Var Mon Corr ARIMA 0.142 18.7 0.62 0.58 0.71 LSTM 0.096 12.3 0.75 0.69 0.82 TCN 0.085 10.1 0.78 0.73 0.85 BiGRU 0.079 9.8 0.81 0.76 0.87 This method 0.063 7.2 0.89 0.84 0.93

[0080] According to the data in the table above, the method in this paper reduces RMSE and MAPE by 18.9% and 26.5% respectively compared with the best comparative model (BiGRU), and the combined scores of Var, Mon, and Corr are significantly improved, indicating that it can more accurately capture the degradation trend and reduce prediction bias.

[0081] Taking a bearing in the test set as an example, its actual remaining life is 85 hours. The prediction result of this method is as follows: Figure 2 As shown, the predicted RUL is 82 hours (error 3.5%), and the degradation curve is smooth, highly consistent with the actual trend (Corr = 0.93). In contrast, the TCN model predicts a value of 76 hours (error 10.6%), and exhibits abnormal fluctuations in the mid-to-late stages (Corr = 0.81). The proposed ECANet-TCN and Self-Attention-BiGRU fusion model demonstrates higher prediction accuracy and robustness under various operating conditions.

Claims

1. A bearing residual life prediction method based on ECAN Tet-TCN and Self-Attention-BiGRU mechanism, characterized in that, Comprise steps of: Step 1, acquire the data detected under different working conditions of different mechanical equipment, and divide the data into training set vibration signal data and test set vibration signal data; Step 2, a prediction model is constructed according to the ECANTet-TCN and Self-Attention-BiGRU mechanisms; Step 3, the batch standardization processing is performed on the vibration signal data input by the training set and the test set; Step 4, the ECANTet-TCN is used to mine data features and enhance useful features while suppressing useless features for the standardized data; Step 5, the data deep time sequence features are mined by using the multi-layer TCN and attention mechanism fusion; Step 6, the time sequence features filtered multiple times are input into the Self-Attention BiGRU network model to predict the RUL of the bearing, and then the Dropout layer is added to construct the model and test the test set; Step 7, collect the vibration signal conforming to the test phase as the basis for judging the life prediction; Step 8, the batch standardization processing is performed on the vibration signal data of the test set; Step 9, the batch standardized test set vibration signal is subjected to feature extraction by the ECANTet-TCN to establish a feature extraction set; Step 10, the model established based on the ECANTet-TCN and the Self-Attention-BiGRU mechanism is used to map the training set signal and the test set signal one by one to complete the prediction of the bearing life; In step 1, the acceleration sensor is used to collect vibration signals in the horizontal and vertical directions, the sampling frequency of the data is 25.6 kHz, the data is recorded once every 10 s, the collection time is 0.1 s, and the vibration data collected each time is summarized; In the experiment, the cross-validation method is used to divide the data, the test data and the training data are divided according to different test data, and the monitoring vibration data of a mechanical equipment under different working conditions is used to construct a feature set, and the training set and the test set are divided; The monitoring data of different working conditions of the same mechanical equipment or different mechanical equipment is used to construct a target domain, a training set and a test set; In step 2, ECANet first uses global average pooling to aggregate the spatial information of each channel of the input feature The average pooling operation expression is: wherein is the value of the element in the input matrix U at the ith row, jth column; W is the width of the matrix U; and H is the height of the matrix U. Then, a one-dimensional convolution pair with a kernel size of K is used. Convolution calculations are performed to quickly extract the feature relationships of K local information points. The sigmoid function is used to calculate the activation values ​​of the one-dimensional convolution output, obtaining weight values ​​representing the local relationships and importance of feature channels. , Sigmoid and The expression is: In the formula, CIDk is the kth operation of the channel interaction module; Finally, in order to re-encode each channel feature of , the and are multiplied to obtain the weighted feature , and the important features are given greater weight to achieve enhancement, and the invalid features are given smaller weight to achieve autonomous inhibition. In step 2, TCN uses dilated convolutions to deal with the long-range dependency problem common in sequential models. Dilated convolutions have a larger receptive field than traditional convolutions, and the computation of dilated convolutions The expression is: wherein: is the dilated convolution calculation process s at time the network output for k is the convolution kernel size; d is the dilation coefficient; is the sequence corresponding to the element in the convolution kernel, ; The calculation of the self-attention mechanism includes two steps: the first step calculates the attention weight between any two vectors in the input sequence; The second step calculates the weighted average value of the input sequence according to the attention weight, and the specific operation mode is as follows: wherein: is a query matrix; K is a key matrix; V is a value matrix; Dim is , K and V dimension of The bidirectional gated recurrent unit BiGRU captures information by back propagation to obtain the potential time sequence relationship between signals.

2. The bearing residual life prediction method based on the ECAN Tet-TCN and the Self-Attention-BiGRU mechanism according to claim 1, characterized in that, In step 3, the input data is standardized, the data is mapped to 0~1, and the selected features are processed and selected by Z-score standardization criteria, and the Z-score standardization criteria formula is as follows: In the formula, Z is the Z-score value, X is the original data, μ is the average value of the original data, and σ is the standard deviation of the original data.

3. The bearing residual life prediction method based on the ECAN Tet-TCN and the Self-Attention-BiGRU mechanism according to claim 1, characterized in that, In step 9, feature selection is performed first, and the time domain features, frequency domain features, time-frequency features and other related features of the vibration signal are selected; The wavelet transform method is used to realize feature extraction; on the basis of feature extraction, standardization and normalization processing are performed; different features have certain correlation, and they are combined to generate new features through feature intersection and feature merging; the feature set is screened through a feature selection algorithm to remove features that have little effect on prediction, and finally the feature set is generated.

4. The bearing residual life prediction method based on the ECAN Tet-TCN and the Self-Attention-BiGRU mechanism according to claim 1, characterized in that, The method further comprises: Step 11: The method based on the ECAN Tet-TCN network and the Self-Attention-BiGRU is evaluated by using three evaluation indexes of Var, Mon and Corr functions.

5. The bearing residual life prediction method based on the ECAN Tet-TCN and the Self-Attention-BiGRU mechanism according to claim 1, characterized in that, In step 11, before evaluating the feature set, the feature is considered as a random process, which is divided into a trend part, represents the average trend and the random part, represents the residual error, as shown in the following equation: In the formulae: is the time degradation characteristic; The above three evaluation indexes can be obtained by the following formula: In the formula, k is the total number of observations; h(t) is a step function; The evaluation and screening mode is as follows: The weight value of each feature before evaluation represents the importance of the index, and the weight value is in the range of (0, 1); in the process of bearing degradation, the cumulative damage is accumulated, and the variance of the degradation feature is considered, and 0.2, 0.5 and 0.3 are selected as the weight values of Var, Mon and Corr respectively.

Images