Bearing life prediction method based on improved polarization self-attention optimization

Abstract

This invention provides a bearing life prediction method based on improved polarization self-attention optimization, belonging to the field of feature extraction and prediction. The method includes the following steps: S1, constructing a gated recurrent unit model based on a polarization self-attention mechanism; S2, constructing a gated recurrent unit model optimized based on a time-related polarization self-attention mechanism; S3, optimizing and training the parameters of the gated recurrent unit model, and then testing the model by inputting the test set data of the bearing dataset into the trained gated recurrent unit model for deep feature extraction; S4, based on the deep feature extraction, constructing a deep interactive model to achieve accurate estimation and uncertainty quantification of the bearing's remaining life. The method of this invention can extract deep degradation features from long-term series data and enhance the ability to capture early information of the sequence by introducing a time weighting factor, thus accurately predicting the bearing's remaining life.

Description

Technical Field

[0001] This invention relates to the field of feature extraction and prediction, and specifically to a bearing life prediction method based on improved polarization self-attention optimization. Background Technology

[0002] Bearings, as critical support components of rotating machinery, are widely used in the transmission systems and power mechanisms of the aforementioned equipment, serving as fundamental elements to ensure the smooth and reliable operation of the equipment. These components endure alternating loads and complex operating conditions over extended periods, and their life-cycle operational status has a decisive impact on the continuous operation capability and safety redundancy of high-end equipment. Therefore, conducting fault prediction and health management (PHM) for bearings, especially accurately predicting their remaining service life, has become a key technological support for ensuring the safe operation of high-end equipment and reducing life-cycle maintenance costs, possessing significant engineering application value and industrial practical significance.

[0003] Currently, data-driven RUL prediction methods have been widely used in this field due to their advantages of not relying on complex physical models and automatically mining implicit features in data, with deep learning models showing particularly outstanding performance. Recurrent Neural Networks (RNNs) and their improved variants—Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRUs)—have become the mainstream technical means for processing equipment lifecycle monitoring data and realizing RUL prediction due to their ability to remember and process temporal information.

[0004] However, existing deep learning models based on LSTM and GRU still have significant technical limitations in bearing RUL prediction scenarios. Bearings, as long-life precision components, exhibit ultra-long time-series monitoring data. When processing such ultra-long time-series data, LSTM and GRU models are prone to gradient vanishing or decaying during backpropagation, making it difficult for the models to accurately capture the long-distance dependencies between early operating states and late-stage degradation states in the time-series. This deficiency leads to the forgetting or loss of crucial early state information, resulting in insufficient accuracy in the extracted equipment performance degradation features, ultimately causing the RUL prediction accuracy to fail to meet the stringent reliability requirements of high-end equipment. Therefore, existing deep learning models still have considerable room for improvement in terms of deep feature extraction capabilities and the accuracy of capturing long-distance dependencies in ultra-long time-series monitoring data. Summary of the Invention

[0005] To address the shortcomings of the prior art, the present invention aims to provide a bearing life prediction method based on improved polarization self-attention optimization, which is used to extract deep degradation features from long-term series data and enhances the ability to capture early information of the sequence by introducing a time weight factor.

[0006] Specifically, the present invention provides a bearing life prediction method based on improved polarization self-attention optimization, which includes the following steps: S1. Construct a gated recurrent unit model based on the polarization self-attention mechanism. The attention directions of the polarization self-attention layer include channel attention branches and spatial attention branches. S2. Construct a gated recurrent unit model optimized based on a time-dependent polarization self-attention mechanism for bearing depth feature extraction, specifically including the following sub-steps: S21. In the weight calculation stages of the channel attention branch and the spatial attention branch, a time weight factor is introduced respectively, and the channel time weight coefficient is... and space-time weighting coefficients The expression is as follows: ; in, , These are the time weighting factors for the channel attention branch and the spatial attention branch, respectively, where t is time; S22. In the gated cyclic unit model, assign time-related weighting coefficients to the data at different time steps. , We obtained a time-dependent polarization self-attention mechanism. in, This refers to the channel attention branch after adding channel time weighting coefficients. The input feature vector, This refers to the spatial attention branch after adding spatial-temporal weighting coefficients; S3. Optimize and train the gated recurrent unit model to make the model parameters optimal, and then test the model. Input the test set data of the bearing dataset into the trained gated recurrent unit model to complete the test. Then input the real-time collected bearing data into the trained gated recurrent unit model and output the bearing depth features. S4. Construct a deep interaction model between a gated cyclic unit model and an enhanced generalized Wiener process. Optimize the model through uncertainty quantification and objective function. Input the bearing depth features into the optimized deep interaction model to predict the remaining service life of the bearing and output the remaining service life of the bearing.

[0007] Preferably, step S1 includes the following steps: S11. Gated recurrent unit network is selected as the basic structure for sequence modeling; S12. Based on the gated recurrent unit network model, a polarized self-attention layer is introduced to perform the first round of feature extraction on the input data. S13. Set the attention direction of the polarized self-attention layer, including channel branch attention and spatial branch attention; S14. In channel branch attention calculation, the polarization self-attention layer reduces the dimension of the spatial branch; in spatial branch attention calculation, it reduces the dimension of the channel branch. S15. The polarization self-attention layer uses the Softmax operation at the minimum feature tensor in the attention module, and then uses the Sigmoid function to dynamically map the features. S16. Set the output combination mode of channel branch attention and spatial branch attention to parallel layout mode; S17. Define the overall structure of the gated recurrent unit model, which includes a data processing layer, a feature extraction layer, and a prediction output layer.

[0008] Preferably, the data processing layer is responsible for receiving the preprocessed raw return oil flow dataset and constructing a sample set using the sliding window method; the feature extraction layer receives the sample set data, and the data is first processed by parallel channel and spatial self-attention through the PSA layer. Then, the features are fed into the first GRU layer for sequence modeling. Each subsequent GRU layer further abstracts and extracts features from the output of the previous layer; the model is mapped through a fully connected layer to output the return oil flow and remaining lifetime value predicted by the model.

[0009] Preferably, step S3 includes the following steps: S31. Select different channel and spatial-temporal weighting factors, set them to 0.1, 0.2, ... 0.8, 0.9 respectively, and perform permutations and combinations to conduct model training and experimental analysis in order to determine the optimal time weighting parameter size; S32. Extract all data after training is completed to obtain the degradation curves of the training set Bearing1_1, Bearing1_2, Bearing1_3 and the test set Bearing1_4, and smooth the extracted feature curves respectively. S33. Set up three comparative models for ablation experiments: ablation model GRU, ablation model PGRU, and ablation model TPGRU; S34. Calculate the three indices of the feature curves extracted based on the GRU, PGRU and TPGRU models respectively, and compare and analyze the results. S35. Based on the divided training and test sets, train and test the two contrast models: contrast model PCA and contrast model deep sparse autoencoder, and compare the extracted feature curves with the feature curves extracted by the TPGRU model. S36. Three HI indicators—monotonicity, robustness, and trend—are selected to evaluate and test the depth feature curves extracted based on PCA and multilayer sparse autoencoders.

[0010] Preferably, step S4 includes the following sub-steps: S41. Input the extracted bearing depth features as a health indicator into the enhanced generalized Wiener process model, and at the same time establish a degradation model and calculate the probability density function of the remaining service life. S42. Construct an objective function based on prediction error and probability density function. Backpropagate the gradient of the objective function through an optimization algorithm. Adjust the network parameters of the gated recurrent unit model to obtain an optimized deep interaction model between the gated recurrent unit model and the enhanced generalized Wiener process. The optimized deep interaction model can extract deeper features that are more conducive to remaining lifetime prediction and re-output the prediction results. S43. Input the bearing data into the optimized deep interactive model, and use the optimized deep interactive model to output the predicted value and uncertainty range of the remaining service life of the bearing.

[0011] Preferably, in step S14, the expression for channel branch attention is as follows: ; ; in, For channel branch attention, , Number of channels; The input feature tensor; , and This is the weight matrix; These are the intermediate parameters for channel convolution; and Reshape two tensors; For the SoftMax function, This refers to the dot product operation of matrices; and and The number of internal passages between them is ; For the Sigmoid function, The total number of feature elements involved in the calculation. These represent the values ​​at the m-th and j-th positions of the output, respectively. The output of the channel branch is as follows: ; in, For the output tensor of the channel-only branch, This is a channel-direction multiplication operation.

[0012] Preferably, in step S14, the expression for spatial branch attention is as follows: ; ; in, Attention to spatial branches, , , For dimensions; For tensor reshaping operators; This is a global pooling function; The output of the spatial branch is as follows: ; in, The output tensor of the spatial branch; This is a spatial direction multiplication operation.

[0013] Preferably, in step S16, the parallel layout is combined as follows: ; in, Combined polarized attention for parallel layout.

[0014] Preferably, in step S41, the bearing depth degradation characteristic curve extracted in step S3 is used as a health indicator. A degradation model is established based on the generalized Wiener process, and the expression of the degradation model is: ; in, As a health indicator, The drift coefficient, For model parameters, Where is the diffusion coefficient. For standard Brownian motion, The initial health index value of the bearing; Remaining lifespan The probability density function is expressed as: ; in, This is the bearing failure threshold. This is the drift coefficient.

[0015] Preferably, in step S42, the objective function Defined as: ; in, and These are weighting coefficients used to determine the accuracy and probability coverage of equilibrium point prediction. The root mean square error, This represents the probability that the actual remaining lifespan falls within the neighborhood of the predicted value. This represents the model's predicted remaining bearing life. This represents the actual remaining service life of the bearing.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention proposes a gated recurrent unit (TPGRU) model optimized based on an improved polarization self-attention mechanism. Addressing the ultra-long time-series characteristics of bearing life-cycle monitoring data, a polarization self-attention layer is introduced into the gated recurrent unit network. This attention layer, through parallel channel and spatial branching, mines the deep correlation between the features of multi-source sensors on the bearing and the time-series degradation patterns, effectively enhancing the ability to capture early, subtle degradation features of the bearing. This significantly improves the model's ability to extract key degradation information from long-series data, providing high-quality health indicators for accurate prediction of the remaining service life of the bearing.

[0017] (2) The bearing life prediction method of the present invention introduces a time weight factor in the polarization self-attention mechanism layer. This factor compensates for the loss of early degradation information in long-term prediction by giving higher weight to earlier data in the time series, thereby significantly enhancing the model’s ability to capture long-term series dependencies and more accurately predicting the remaining service life of bearings.

[0018] (3) The present invention further constructs an interactive model of gated cyclic unit model and enhanced generalized Wiener process. Through uncertainty quantification and objective function optimization, it realizes accurate prediction of bearing remaining service life and effectively solves the problem that a single deep learning model cannot quantify and predict uncertainty.

[0019] (4) The depth feature curve extracted by the present invention performs better in terms of monotonicity Mon, robustness Rob and trend Tre, and can output the confidence interval of the prediction results, providing more reliable technical support for the health management and maintenance decision of bearings. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the bearing life prediction method based on improved polarization self-attention optimization according to the present invention. Figure 2 This is a schematic diagram of the TPSA principle of the bearing life prediction method based on improved polarization self-attention optimization of the present invention. Figure 3 This is the PGRU model structure framework of the present invention based on the improved polarization self-attention optimization bearing lifetime prediction method; Figure 4 This is a schematic diagram of the parallel layout polarization self-attention mechanism of the present invention; Figure 5 This is a graph showing the test results under different combinations of time weighting factors according to the present invention; Figure 6 This is a feature curve extracted based on the GRU method in an embodiment of the present invention; Figure 7 This is a feature curve extracted based on the PGRU method in an embodiment of the present invention; Figure 8 This is a feature curve extracted based on the TPGRU method in an embodiment of the present invention; Figure 9 This is a feature curve extracted based on the PCA method in an embodiment of the present invention; Figure 10 This is a feature curve extracted based on the deep sparse autoencoder method in an embodiment of the present invention. Detailed Implementation

[0021] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0022] like Figure 1 and Figure 2 The diagram shows the preferred embodiment of the present invention, a bearing lifetime prediction method based on an improved polarization self-attention mechanism optimized gated cyclic unit (TPGRU), the specific steps of which are as follows: S1. Construct a gated recurrent unit model based on a polarization self-attention mechanism to predict bearing life. The attention directions of the polarization self-attention layer include channel attention branches and spatial attention branches. Specifically, this includes the following sub-steps: S11. Gated recurrent unit network is selected as the basic structure for sequence modeling.

[0023] S12. Based on the gated recurrent unit network model, a polarization self-attention layer is introduced to perform the first round of feature extraction on the input bearing monitoring data, thereby enhancing the key degradation information in different sensor channels or different time segments.

[0024] S13. Set the attention direction of the polarization self-attention layer, including channel branch attention and spatial branch attention. The channel branch corresponds to the multi-source feature dimension of different measurement points and different feature types of bearing monitoring data, and the spatial branch corresponds to the time series dimension of different sampling times.

[0025] S14. In channel branch attention calculation, the polarized self-attention layer reduces the dimension of the spatial branch; in spatial branch attention calculation, it reduces the dimension of the channel branch.

[0026] S15. The polarization self-attention layer uses the Softmax operation at the minimum feature tensor in the attention module, and then uses the Sigmoid function to dynamically map the features.

[0027] S16. Set the output combination of channel branch attention and spatial branch attention to a parallel layout, and mine bearing degradation information from multiple feature dimensions and time dimension to improve the comprehensiveness and diversity of feature extraction.

[0028] S17. The overall structure of the gated loop unit model is defined as including a data processing layer, a feature extraction layer, and a prediction output layer. The data processing layer is used to receive the preprocessed bearing monitoring data and construct a sample set. The feature extraction layer is used to extract deep degradation features. The prediction output layer is used to map and obtain the predicted value of the remaining service life.

[0029] In this embodiment, a gated recurrent unit (GRU) network is selected as the basic structure for sequence modeling. Based on the GRU model, a polarized self-attention (PSA) layer is introduced to enhance the ability to capture long-distance dependencies, strengthen the attention to important information, and suppress the attention to irrelevant information.

[0030] PSA's attention direction has two branches: channel branch attention and spatial branch attention. In channel branch attention calculation, the spatial branch dimension is reduced; in spatial branch attention calculation, the channel branch dimension is reduced.

[0031] The first step of PSA is to achieve the effect of polarization filtering by completely collapsing the features of one dimension.

[0032] The second step is to use the Softmax operation at the minimum feature tensor in the attention module to ensure that important features are given higher attention weights; then the Sigmoid function is used to dynamically map the features to retain more effective information in the output.

[0033] In step S14, the expression for channel branch attention is as follows: ; ; in, For channel branch attention, , Number of channels; The input feature tensor; , and This is the weight matrix; These are the intermediate parameters for channel convolution; and Reshape two tensors; For the SoftMax function, This refers to the dot product operation of matrices; and and The number of internal passages between them is ; For the Sigmoid function, The total number of feature elements involved in the calculation. These represent the values ​​at the m-th and j-th positions of the output, respectively.

[0034] The output of the channel branch is as follows: ; in, For the output tensor of the channel-only branch, For channel-direction multiplication operations, , For dimensions.

[0035] In step S14, the expression for spatial branch attention is as follows: ; ; in, Attention to spatial branches, , , For dimensions; For tensor reshaping operators; This is a global pooling function; The output of the spatial branch is as follows: ; in, The output tensor of the spatial branch; This is a spatial direction multiplication operation.

[0036] There are two ways to combine the outputs of channel branch attention and spatial branch attention: parallel layout combination and serial layout combination.

[0037] The parallel layout combinations are as follows: ; in, Combined polarized attention for parallel layout.

[0038] The serial layout combinations are as follows: ; in, For serial layout, combine polarized attention.

[0039] Comparing the two layout methods, the parallel layout has relatively higher computational efficiency and can simultaneously mine multi-dimensional features of the data, enhancing the comprehensiveness and diversity of the model's feature extraction of flow data. In scenarios where data features are relatively uniformly distributed and there is no clearly dominant feature dimension, it can fully leverage the advantages of dual-branch parallelism to efficiently capture potential features of early-state data, providing rich evidence for remaining service life prediction. Especially when processing large-scale return oil flow data, it can reduce training time costs while maintaining good performance. Therefore, the parallel layout method is chosen. The principle of the parallel layout polarized self-attention mechanism is as follows: Figure 4 .

[0040] The GRU model based on PSA is named the PGRU model. The PGRU model structure framework is as follows: Figure 3 The data input to the model first undergoes preliminary feature extraction through the PSA layer, and then deep feature extraction is completed through the subsequent GRU layer.

[0041] The overall structure of the PGRU model mainly consists of three parts: a data processing layer, a feature extraction layer, and a prediction output layer.

[0042] The first data processing layer receives the raw return oil flow dataset after data preprocessing and constructs a sample set by dividing the time window using the sliding window method.

[0043] The second feature extraction layer receives input data from the data processing layer. First, it performs parallel processing of channel self-attention and spatial self-attention mechanisms in the PSA layer to perform the first round of feature extraction. The data is then fed into the first GRU layer, where the GRU layer models the sequence and further extracts temporal features and other information from the sequence. Each subsequent GRU layer further abstracts and extracts features from the output of the previous layer, thereby learning more complex time series patterns. Finally, the data is mapped through a fully connected layer to output the model's predicted return flow rate and RUL value.

[0044] S2. Construct a gated recurrent unit model optimized based on a time-dependent polarization self-attention mechanism for bearing depth feature extraction, specifically including the following sub-steps: S21. In the weight calculation stages of the channel attention branch and the spatial attention branch, a time weight factor is introduced respectively, and the channel time weight coefficient is... and space-time weighting coefficients The expression is as follows: ; in, , These are the time weight factors corresponding to the channel attention branch and the spatial attention branch, respectively, with values ​​ranging from 0 to 1, where t is time.

[0045] S22. In the gated cyclic unit model, assign time-related weighting coefficients to the data at different time steps. , We obtained a time-dependent polarization self-attention mechanism. in, This refers to the channel attention branch after adding channel time weighting coefficients. The input feature vector, This is the spatial attention branch after adding spatial-temporal weighting coefficients.

[0046] Specifically, input data Time series data with time weighting coefficients as shown in the following formula are added. , .

[0047] ; After the above calculations, the expression for the channel-only branch of the TPSA method is as follows: ; ; The TPSA method only outputs the channel branch. as follows: ; Spatial branch only The expression is as follows: ; ; Output of only spatial branch as follows: ; The expression for parallel placement combined polarization attention in the TPSA method is as follows: ; The GRU model improved based on the TPSA mechanism is named the TPGRU model. The principle of the TPSA mechanism improved based on the time weight coefficient is as follows: Figure 1 .

[0048] S3. Optimize and train the gated recurrent unit (GRU) model to achieve optimal parameters. Then test the model by inputting the test set data from the bearing dataset into the trained GRU model. After testing, input the real-time collected bearing data into the trained GRU model and output the bearing depth features. This includes the following sub-steps: S31. Select different channel and spatial-temporal weighting factors, set them to 0.1, 0.2, ... 0.8, 0.9 respectively, and perform permutations and combinations to conduct model training and experimental analysis in order to determine the optimal time weighting parameter size.

[0049] S32. Extract all data after training is completed to obtain the degradation curves of the training set Bearing1_1, Bearing1_2, Bearing1_3 and the test set Bearing1_4, and smooth the extracted feature curves respectively.

[0050] S33. Set up three comparative models for ablation experiments: ablation model GRU, ablation model PGRU, and ablation model TPGRU.

[0051] S34. Calculate the three indices Mon, Rob, and Tre for the feature curves extracted based on the GRU, PGRU, and TPGRU models respectively, and compare and analyze the results.

[0052] S35. Train and test the gated recurrent unit model according to the divided training and test sets, and compare the extracted feature curves with the feature curves extracted by the TPGRU model.

[0053] In this embodiment, the parameters of the TPGRU model are first set. Different channel and spatial-temporal weight factors are selected and set to 0.1, 0.2, ... 0.8, 0.9 respectively, and then permuted and combined. Model training and experimental analysis are then performed, and the experimental results are as follows. Figure 5 As shown, the model trained with a combination of channel time weighting factor of 0.8 and space time weighting factor of 0.7 has the lowest root mean square error (RMSE).

[0054] Other experimental parameters are set as follows: (1) The optimizer is Adam algorithm and the learning rate is set to 0.005; (2) The number of GRU layers is 2; (3) The number of neurons in the GRU layers is 60; (4) The training cycle is 600; (5) The number of channels in the PSA layer is 3.

[0055] After the model training is completed, all the data is input into it to extract the degradation curves of the training set Bearing1_1, Bearing1_2, Bearing1_3 and the test set Bearing1_4, and then the extracted feature curves are smoothed respectively.

[0056] Then, to demonstrate the effectiveness of the TPGRU model improvement, an ablation experiment was designed and experimentally verified.

[0057] Three ablation models were set up: (1) Ablation model GRU: GRU modeling was performed, and feature extraction was performed solely based on it; (2) Ablation model PGRU: PSA mechanism was added to the GRU model to establish the PGRU model; (3) Ablation model TPGRU: Time weight factor was introduced on the basis of the PGRU model to establish the TPGRU model. The three ablation models were trained and tested according to the divided bearing dataset.

[0058] Three commonly used HI (Highly Indicative of Income) metrics are applied to further analyze the deep features extracted from the model, in order to quantitatively describe the practicality and effectiveness of the extracted deep features in the field of remaining useful life prediction. The calculation methods for the three commonly used HI metrics are as follows: Monotonicity (Mon): In actual production processes, the actual health condition of equipment often deteriorates gradually; therefore, the actual health curve (HI curve) of the equipment should exhibit monotonicity. The formula for calculating monotonicity is as follows: ; in, This represents a sequence of changes in the device's HI value over time. Indicates that the device is in HI value at any given time; Indicates the increment of the HI value; and These represent the number of positive and negative increments of the HI value at adjacent time points, respectively. The value of HI ranges from 0 to 1, and the closer the value is to 1, the better the monotonicity of HI.

[0059] Robustness (Ro): External environmental interference during data collection can severely affect the stability of health assessment results. The robustness calculation formula for the HI curve is as follows: ; in, Indicates in The average trend value of the time-time characteristic curve; The value of is between 0 and 1, and the closer its value is to 1, the better the robustness of the characteristic curve.

[0060] Trend (Trend, Tre): Typically, the characteristic curve of a device is related to its operating time. The formula for calculating the trend of the HI curve is shown in equation (3-36): ; in, The value ranges from 0 to 1, and the closer the value is to 1, the stronger the trend of the characteristic curve.

[0061] Feature curves extracted based on GRU, PGRU, and TPGRU models are as follows: Figure 6 , Figure 7 , Figure 8 As can be seen, the bearing characteristic curves constructed by the TPGRU model have good monotonicity and trend, and can describe the degradation process of the bearing throughout its entire life cycle. Compared with the characteristic curves extracted by directly using the GRU model and the PGRU-based method, the characteristic curves extracted by TPGRU are smoother and have better monotonicity and trend.

[0062] To further compare the merits of the characteristic curves, three indices for the characteristic curves were calculated: Mon, Rob, and Tre, and the results were compared and analyzed.

[0063] Table 1 shows the evaluation results of different models in the ablation experiment: the three indices of the TPGRU characteristic curve are all better than the GRU model and the PGRU model. Among them, the robustness index of Bearing1_1 and the trend index of Bearing1_4 reached 0.90 and 0.94 respectively, which are far superior to the other two models. The three indices of the PGRU characteristic curve are also better than the GRU model. Among them, the monotonicity index of the two models of Bearing1_3 is not much different, and the index of the TPGRU characteristic curve is better, which proves that TPGRU is more capable of dealing with complex data.

[0064] Table 1. Ablation Experiment Evaluation Results Finally, comparative analysis experiments were conducted to analyze and verify the advantages of the TPGRU model over the PCA model and the deep sparse autoencoder model.

[0065] The PCA model and the deep sparse autoencoder model were trained and tested respectively based on the divided bearing dataset, and the extracted feature curves were compared with the feature curves extracted by the TPGRU model.

[0066] The depth feature curves extracted based on PCA and multilayer sparse autoencoders are shown below. Figure 9 , Figure 10 As shown, compared with the feature curves extracted by TPGRU, the depth feature curves extracted by the PCA method have worse monotonicity, trend, and smoothness; compared with them, the depth feature curves of the deep sparse autoencoder show obvious fluctuations, are not monotonic enough, and to some extent show a trend of rapid early degradation, which does not match the actual situation.

[0067] Furthermore, monotonicity, robustness, and trend were selected as three evaluation indicators to evaluate the depth feature curves extracted by the TPGRU model, PCA model, and deep sparse autoencoder model. Table 2 shows the comparative experimental evaluation results of the PCA model, deep sparse autoencoder model, and TPGRU model under the working condition of ball bearing subjected to a load of 4000 N and a speed of 1800 r / min.

[0068] Table 2 Comparative Experiment Evaluation Results S4. Construct a deep interaction model between the gated cyclic unit model and the enhanced generalized Wiener process. Optimize this model through uncertainty quantification and an objective function. Input the bearing's deep features into the optimized deep interaction model to predict and output the bearing's remaining service life. This includes the following sub-steps: S41. Input the extracted bearing depth features as health indicators into the enhanced generalized Wiener process model, and at the same time establish a degradation model and calculate the probability density function of the remaining service life.

[0069] S42. Construct an objective function based on prediction error and probability density function. Backpropagate the gradient of the objective function through an optimization algorithm. Adjust the network parameters of the gated recurrent unit model to obtain an optimized deep interaction model of the gated recurrent unit model and the enhanced generalized Wiener process. The optimized deep interaction model can extract deeper features that are more conducive to remaining lifetime prediction and re-output the prediction results.

[0070] S43. Input the bearing data into the optimized deep interactive model, and use the optimized deep interactive model to output the predicted value and uncertainty range of the remaining service life of the bearing.

[0071] In this embodiment, after completing the feature extraction and validation of the TPGRU model, to further achieve accurate prediction of the remaining bearing service life and quantify the uncertainty of the prediction results, this invention constructs a deep interaction model between the TPGRU model and the Enhanced Generalized Wiener Process (EGWP). First, the bearing deep degradation characteristic curve extracted in S3 is used as a health indicator. A degradation model is established based on the generalized Wiener process, and its expression is: ; in, As a health indicator, The drift coefficient, For model parameters, Where is the diffusion coefficient. This is standard Brownian motion. The failure threshold is set based on the bearing failure mechanism. By analyzing historical degradation data, the model parameters and Maximum likelihood estimation is performed to determine the specific form of the degradation model.

[0072] Based on the first-arrival time concept of the generalized Wiener process, the remaining service life of the bearing Based on the first-arrival time concept of the generalized Wiener process, the remaining service life of the bearing First time reaching the threshold The required time. At this point, the remaining lifespan. The probability density function can be analytically expressed as: .

[0073] This probability density function describes the probability distribution of the future failure time of the bearing given its current health condition, thus quantifying the uncertainty of the prediction results.

[0074] To achieve deep interaction between TPGRU and EGWP, this invention constructs a comprehensive objective function that combines the remaining useful life point estimate predicted by the TPGRU model with the probability density function calculated by the EGWP model. Objective function Defined as: ; in, and These are weighting coefficients used to determine the accuracy and probability coverage of equilibrium point prediction. The root mean square error, This represents the probability that the actual remaining lifespan falls within the neighborhood of the predicted value, reflecting the reliability of the prediction. This represents the model's predicted remaining bearing life. This represents the true remaining service life of the bearing. By minimizing this objective function, we can simultaneously optimize the accuracy of point prediction and the reliability of uncertainty quantification.

[0075] Subsequently, the objective function is used as an additional loss term in the TPGRU model, and the network parameters of the TPGRU model are updated using the Adam optimization algorithm. The specific process is as follows: First, the TPGRU model trained in S3 is used to perform forward propagation on the bearing test set data to obtain initial predicted values ​​of remaining service life points and depth feature curves; then, the extracted depth feature curves are input into the EGWP model to estimate the degradation model parameters and calculate the probability density function of the remaining service life at each time step; next, the comprehensive objective function value is calculated according to the above formula; then, the gradient of the objective function is backpropagated to each network layer of the TPGRU model to adjust the weight parameters; finally, the updated TPGRU model is used to re-predict the bearing test set, outputting the final estimated values ​​of the remaining service life points and the corresponding probability density function.

[0076] In another embodiment of the present invention, as an optional embodiment, the present invention can also provide a bearing life prediction system based on improved polarization self-attention optimization, which includes a gated cyclic unit model construction module, a gated cyclic unit model training module, and a bearing life prediction module. The gated cyclic unit model construction module is used to construct a gated cyclic unit model based on the polarization self-attention mechanism for bearing depth feature extraction. The gated cyclic unit model training module is used to optimize and train the parameters of the gated cyclic unit model, and then test the model after optimizing the model parameters. The test set data of the bearing dataset is input into the trained gated cyclic unit model to complete the test. After the test is completed, the bearing data collected in real time is input into the trained gated cyclic unit model, and the bearing depth features are output. The bearing life prediction module constructs a deep interaction model between the gated cyclic unit model and the enhanced generalized Wiener process, optimizes it through uncertainty quantification and objective function, inputs the bearing depth features into the optimized deep interaction model to predict the remaining service life of the bearing, and outputs the remaining service life of the bearing.

[0077] This invention designs a gated recurrent unit (TPGRU) model optimized based on an improved polarization self-attention mechanism and its interaction model with an enhanced generalized Wiener process (EGWP). By introducing a polarization self-attention layer and adding a time weight factor into the gated recurrent unit network, it effectively solves the problem of poor long-term series feature extraction in the RUL prediction process of the GRU model. Furthermore, by constructing a deep interaction mechanism between TPGRU and EGWP, it achieves accurate prediction and uncertainty quantification of bearing remaining service. Experimental verification on bearing datasets shows that the feature curves extracted by the proposed TPGRU model have significant advantages in monotonicity (Mon), robustness (Rob), and trend (Trend) indices. The TPGRU-EGWP interaction model significantly improves the RUL prediction accuracy and can output the confidence interval of the prediction results, providing more reliable technical support for bearing health management and maintenance decisions.

[0078] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A bearing life prediction method based on improved polarization self-attention optimization, characterized in that: It includes the following steps: S1. Construct a gated recurrent unit model based on polarization self-attention mechanism to predict bearing life. The attention direction of the polarization self-attention layer includes channel attention branch and spatial attention branch. S2. Construct a gated recurrent unit model optimized based on a time-dependent polarization self-attention mechanism for bearing depth feature extraction, specifically including the following sub-steps: S21. In the weight calculation stages of the channel attention branch and the spatial attention branch, a time weight factor is introduced respectively, and the channel time weight coefficient is... and space-time weighting coefficients The expression is as follows: ； in, , These are the time weighting factors for the channel attention branch and the spatial attention branch, respectively, where t is time; S22. In the gated cyclic unit model, assign time-related weighting coefficients to the data at different time steps. , The time-dependent polarization self-attention mechanism was obtained. ； in, This refers to the channel attention branch after adding channel time weighting coefficients. The input feature vector, This refers to the spatial attention branch after adding spatial-temporal weighting coefficients; S3. Optimize and train the gated recurrent unit model to make the model parameters optimal, and then test the model. Input the test set data of the bearing dataset into the trained gated recurrent unit model to complete the test. Then input the real-time collected bearing data into the trained gated recurrent unit model and output the bearing depth features. S4. Construct a deep interaction model between a gated cyclic unit model and an enhanced generalized Wiener process. Optimize the model through uncertainty quantification and objective function. Input the bearing depth features into the optimized deep interaction model to predict the remaining service life of the bearing and output the remaining service life of the bearing.

2. The bearing life prediction method based on improved polarization self-attention optimization according to claim 1, characterized in that: Step S1 includes the following steps: S11. Gated recurrent unit network is selected as the basic structure for sequence modeling; S12. Based on the gated recurrent unit network model, a polarization self-attention layer is introduced to perform the first round of feature extraction on the input bearing monitoring data, thereby enhancing the key degradation information in different sensor channels or different time segments. S13. Set the attention direction of the polarized self-attention layer to include channel branch attention and spatial branch attention. The channel branch corresponds to the multi-source feature dimension of different measurement points and different feature types of bearing monitoring data, and the spatial branch corresponds to the time series dimension of different sampling times. S14. In channel branch attention calculation, the polarization self-attention layer reduces the dimension of the spatial branch; in spatial branch attention calculation, it reduces the dimension of the channel branch. S15. The polarization self-attention layer uses the Softmax operation at the minimum feature tensor in the attention module, and then uses the Sigmoid function to dynamically map the features. S16. Set the output combination method of channel branch attention and spatial branch attention to a parallel layout, and mine bearing degradation information from multiple feature dimensions and time dimension to improve the comprehensiveness and diversity of feature extraction. S17. The overall structure of the gated loop unit model is defined as including a data processing layer, a feature extraction layer, and a prediction output layer. The data processing layer is used to receive the preprocessed bearing monitoring data and construct a sample set. The feature extraction layer is used to extract deep degradation features. The prediction output layer is used to map and obtain the predicted value of the remaining service life.

3. The bearing life prediction method based on improved polarization self-attention optimization according to claim 2, characterized in that: The data processing layer is responsible for receiving the preprocessed raw return oil flow dataset and constructing a sample set using the sliding window method. The feature extraction layer receives the sample set data. The data is first processed by the PSA layer in parallel channel and spatial self-attention. Then, the features are fed into the first GRU layer for sequence modeling. Each subsequent GRU layer further abstracts and extracts features from the output of the previous layer. The model is then mapped through a fully connected layer to output the return oil flow and remaining lifetime values ​​predicted by the model.

4. The bearing life prediction method based on improved polarization self-attention optimization according to claim 1, characterized in that: Step S3 includes the following steps: S31. Select different channel and spatial-temporal weighting factors, set them to 0.1, 0.2, ... 0.8, 0.9 respectively, and perform permutations and combinations to conduct model training and experimental analysis in order to determine the optimal time weighting parameter size; S32. Extract all data after training is completed to obtain the degradation curves of the training set Bearing1_1, Bearing1_2, Bearing1_3 and the test set Bearing1_4, and smooth the extracted feature curves respectively. S33. Set up three comparative models for ablation experiments: ablation model GRU, ablation model PGRU, and ablation model TPGRU; S34. Calculate the three indices of the feature curves extracted based on the GRU, PGRU and TPGRU models respectively, and compare and analyze the results. S35. Train and test the gated recurrent unit model according to the divided training and test sets, and compare the extracted feature curves with the feature curves extracted by the TPGRU model.

5. The bearing life prediction method based on improved polarization self-attention optimization according to claim 1, characterized in that: Step S4 includes the following sub-steps: S41. Input the extracted bearing depth features as a health indicator into the enhanced generalized Wiener process model, and at the same time establish a degradation model and calculate the probability density function of the remaining service life. S42. Construct an objective function based on prediction error and probability density function. Backpropagate the gradient of the objective function through an optimization algorithm. Adjust the network parameters of the gated recurrent unit model to obtain an optimized deep interaction model between the gated recurrent unit model and the enhanced generalized Wiener process. The optimized deep interaction model can extract deeper features that are more conducive to remaining lifetime prediction and re-output the prediction results. S43. Input the bearing data into the optimized deep interactive model, and use the optimized deep interactive model to output the predicted value and uncertainty range of the remaining service life of the bearing.

6. The bearing life prediction method based on improved polarization self-attention optimization according to claim 1, characterized in that: In step S14, the expression for channel branch attention is as follows: ； ； in, For channel branch attention, , Number of channels; The input feature tensor; , and This is the weight matrix; These are the intermediate parameters for channel convolution; and Reshape two tensors; For the SoftMax function, This refers to the dot product operation of matrices; and and The number of internal passages between them is ; For the Sigmoid function, The total number of feature elements involved in the calculation. These represent the values ​​at the m-th and j-th positions of the output, respectively. The output of the channel branch is as follows: ； in, For the output tensor of the channel-only branch, For channel-direction multiplication operations, , For dimensions.

7. The bearing life prediction method based on improved polarization self-attention optimization according to claim 6, characterized in that: In step S14, the expression for spatial branch attention is as follows: ； ； in, Attention to spatial branches, , , For dimensions; Remodeling operators for tensors; This is a global pooling function; The output of the spatial branch is as follows: ； in, The output tensor of the spatial branch; This is a spatial direction multiplication operation.

8. The bearing life prediction method based on improved polarization self-attention optimization according to claim 7, characterized in that: In step S16, the parallel layout combination is as follows: ； in, Combined polarized attention for parallel layout.

9. The bearing life prediction method based on improved polarization self-attention optimization according to claim 6, characterized in that: In step S41, the bearing depth degradation characteristic curve extracted in step S3 is used as a health indicator. A degradation model is established based on the generalized Wiener process, and the expression of the degradation model is: ； in, As a health indicator, The drift coefficient, For model parameters, The diffusion coefficient is... For standard Brownian motion, The initial health index value of the bearing; Remaining lifespan The probability density function is expressed as: ； in, This is the bearing failure threshold. This is the drift coefficient.

10. The bearing life prediction method based on improved polarization self-attention optimization according to claim 9, characterized in that: In step S42, the objective function Defined as: ； in, and These are weighting coefficients used to determine the accuracy and probability coverage of equilibrium point prediction. The root mean square error, This represents the probability that the actual remaining lifespan falls within the neighborhood of the predicted value. This represents the model's predicted remaining bearing life. This represents the actual remaining service life of the bearing.

Images