A dual-stream temporal convolution-enhanced interactive bearing life prediction method

Abstract

The present invention discloses a dual-stream time convolution enhanced interactive bearing life prediction method, which includes collecting original bearing vibration signal data, processing and reconstructing the data; integrating the local time series modeling capability of TCN with multi-scale expansion attention to construct a time stream feature extraction module, and dynamically adjusting the expansion rate through spectral entropy, combining sparse multi-head attention to capture multi-scale local features; integrating frequency-aware position coding with hierarchical sparse attention mechanism to construct a spatial stream feature extraction module; designing a bidirectional cross-layer attention coordination mechanism, interacting the features of the time stream feature extraction module and the spatial stream feature extraction module, enhancing the degradation characterization capability through hierarchical feature alignment and dynamic weight adjustment, and constructing a dual-stream feature extraction layer; inputting the reconstructed signal into the dual-stream feature extraction layer to extract features; inputting the extracted features into a multi-layer perceptron for RUL prediction; realizing local-global feature complementarity; and improving prediction stability.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical life prediction, and in particular to a dual-stream time convolution enhanced interactive bearing life prediction method. Background Art

[0002] The remaining useful life (RUL) of a bearing refers to the period of time between its current state and complete failure during which the bearing can continue to operate safely and effectively. RUL prediction can prevent equipment failure at critical moments, ensuring safe and efficient production. Traditional bearing remaining life prediction methods are primarily based on statistical features, physical models, or machine learning algorithms, predicting the remaining useful life by analyzing and modeling the bearing's performance degradation indicators. Bearing vibration signals contain high-frequency transient noise and low-frequency degradation trends, requiring the simultaneous capture of local mutations and global periodic patterns. However, existing methods separate local and global feature extraction, resulting in incomplete feature extraction capabilities, low feature fusion efficiency, and an inability to consider the changing characteristics of long time series, leading to limited prediction capabilities. Summary of the Invention

[0003] The purpose of the present invention is to provide a dual-stream time convolution enhanced interactive bearing life prediction method to solve the problems raised in the background technology.

[0004] To solve the above technical problems, the present invention adopts the following technical solution: a dual-stream time convolution enhanced interactive bearing life prediction method, comprising:

[0005] Collect original bearing vibration signal data, process and reconstruct the data;

[0006] Construct a two-stream feature extraction layer and input the reconstructed signal into the two-stream feature extraction layer to extract features;

[0007] Input the mentioned features into the multi-layer perceptron for RUL prediction;

[0008] Building a two-stream feature extraction layer includes:

[0009] The local temporal sequence modeling capability of TCN is integrated with multi-scale dilated attention to construct a temporal stream feature extraction module. The dilation rate is dynamically adjusted through spectral entropy, and sparse multi-head attention is combined to capture multi-scale local features.

[0010] The spatial stream feature extraction module is constructed by integrating frequency-aware position encoding and hierarchical sparse attention mechanism;

[0011] A bidirectional cross-layer attention coordination mechanism is designed to interact the features of the temporal stream feature extraction module and the spatial stream feature extraction module, and enhance the degradation representation capability through hierarchical feature alignment and dynamic weight adjustment.

[0012] Preferably, the original signal sample point data is processed through a sliding window to obtain time features; the sliding window is used for data segmentation preprocessing, so that the model can extract the most valuable information from the multivariate time series, reduce the amount of data, and extract key information; the original time series is represented as , where Ti represents the length of the time series and the length of the sliding window is , moving with a step length of Ls, the length of the overlapping area of ​​the two sliding windows is ,by The signal is intercepted in units for processing, and the input data for the estimation model is generated by applying a sliding time window to the sequence data; this process can be expressed as , T is the partition function, is the time window size.

[0013] Preferably, the time stream feature extraction module includes:

[0014] Design a parallel multi-branch dilated convolution layer, each layer contains three parallel dilated convolutions to extract features at different time scales. Use convolution kernels with different dilation rates in parallel in the same TCN layer to capture local patterns of different granularities and then fuse the features.

[0015] A multi-scale dilated attention mechanism is introduced to divide the multi-branch dilated convolution output of the TCN layer into N groups with C channels. Each group generates an independent query, key, and value matrix. Each group corresponds to an attention head of the MSDA, and each head uses a different dilation rate to capture contextual information at different time scales and achieve feature granularity alignment.

[0016] For each attention head’s query position, the expansion rate is used within the sliding window. Sparse selection of key-value pairs;

[0017] Compute local context features by scaling dot product attention:

[0018] ;

[0019] in, is the dimension of each attention head;

[0020] The output of each attention head is spliced ​​along the channel dimension, and the multi-scale features are aggregated through the linear transformation layer to retain the multi-scale context information and output :

[0021] .

[0022] Preferably, dynamically adjusting the expansion rate based on the spectral entropy of the input feature includes:

[0023] Calculate the spectral entropy H(f) of the current input signal x;

[0024] Through lightweight MLP, H(f) is mapped to the expansion rate adjustment coefficient Δd, so that the convolution kernel is more focused on the key local interval;

[0025] ;

[0026] The corrected expansion rate is: .

[0027] Preferably, in the process of constructing the spatial stream feature extraction module:

[0028] For the input sequence Perform FFT to extract the frequency component with the largest amplitude , adding frequency-aware relative position bias to the attention score:

[0029] Frequency modulation position encoding: ;

[0030] Relative position offset: ;

[0031] in, is a learnable parameter, t is the time step index, is the learnable scaling factor;

[0032] Introduce a time decay mask to increase attention to recent time steps:

[0033] ;

[0034] in, is the learnable decay coefficient, i and j are the time indexes of query and key respectively;

[0035] Corrected attention calculation: .

[0036] Preferably, in the process of constructing the spatial stream feature extraction module, the following steps are further included:

[0037] Divide the input sequence into M segments {X1, X2, ..., XM}, each segment length L = T / M;

[0038] Perform standard self-attention within each segment, computing the amount ;

[0039] Perform cross-fragment attention on the fragment representative vector, the computational And fuse local and global features through gating mechanism:

[0040] ;

[0041] in .

[0042] Preferably, the feature interaction includes:

[0043] The features extracted by the temporal stream feature extraction module are used as the query, and the features extracted by the spatial stream feature extraction module are used as the key / value to construct detail-enhanced attention:

[0044] ;

[0045] Dynamically adjust the contribution of local anomalies in global features through attention weights to generate fusion features , injecting the transient impact features captured by the time stream feature extraction module into the global representation of the spatial stream feature extraction module:

[0046] ;

[0047] The global trend of the spatial stream feature extraction module is used as the query, and the multi-scale features of the temporal stream feature extraction module are used as the key / value to calculate the attention weight:

[0048] ;

[0049] Modify the weight distribution of local features through global trends: ;

[0050] Adopting a learnable gating mechanism to achieve dynamic fusion of bidirectional features:

[0051] ;

[0052] Perform hierarchical feature interaction in L network layers and fuse the features of each layer Splicing along the channel dimension, the fusion features generated in each layer are , B is the batch size, T is the time step, C L is the number of channels in the Lth layer, " indicates splicing along the channel dimension, and constructing multi-scale representation by splicing along the channel dimension:

[0053] .

[0054] Preferably, each of the network layers performs feature interaction and reduces the amount of computation by sharing some parameters.

[0055] Preferably, a dynamic Tanh normalization layer is used to replace the traditional LayerNorm, specifically:

[0056] Replace each normalization layer with a DyT component, leaving other components unchanged:

[0057] ;

[0058] in, To dynamically adjust the factor, the input scale is automatically optimized by gradient descent, and the default initialization is 0.5; and The feature calibration parameters are channel-level affine parameters, initialized to all 1s and all 0s;

[0059] The temporal convolution layer extracts local temporal features and dynamically adjusts the convolution output:

[0060] ;

[0061] Insert DyT before and after the attention calculation to stabilize the attention weight distribution:

[0062] ;

[0063] By adjusting the fusion scale of cross-layer features through DyT, the stability of gradient flow is enhanced, and the output flow is:

[0064] .

[0065] Design multi-scale dynamic weighted loss constraints: In the time stream feature extraction module, multi-scale time series features are extracted through different expansion factors, and prediction results are generated at each scale. ; The spatial stream feature extraction module outputs the global prediction results , and use the mean square error as the loss function, extract the self-attention matrix of the spatial stream feature module Extract the importance weights of each time scale, realize the dynamic allocation of attention weights, and ensure , k is the number of expansion factors; the error of each scale of the time stream feature extraction module is determined by the attention weight Weighted; According to the global dependency of the input sequence, the loss weight of each time scale is dynamically allocated to alleviate the multi-scale fusion deviation; including 、 , to conduct information interaction, the loss of the spatial stream feature extraction module directly guides the loss weighting of the temporal stream feature extraction module, realizing explicit collaboration between branches;

[0066] The specific formula is:

[0067] ;

[0068] ;

[0069] ;

[0070] d∈{d1,…,dk}; is the attention weight A(d) of the d-th scale: the attention matrix corresponding to the d-th time scale;

[0071] Introduce regularization constraints and add feature alignment loss Control the spatial consistency of the hidden layer features of the temporal stream feature extraction module and the spatial stream feature extraction module; and enhance information fusion through the hidden layer representation L2 distance constraint; and adjust the convolution kernel weight of the temporal stream feature extraction module. L1 regularization is applied to achieve sparsity constraints; the attention matrix A is constrained to be close to the identity matrix through the Frobenius norm. This can promote modal fusion, forcing the two branches to learn consistent feature representations, avoiding feature space fragmentation due to architectural differences, and enhancing the model's robustness to input perturbations.

[0072] The specific formula is:

[0073] ;

[0074] ;

[0075] ;

[0076] in, 、 is the hidden layer feature vector of the TCN and spatial stream feature extraction module, 、 is the regularization weight coefficient;

[0077] Combined multi-scale dynamic weighted loss constraints and regularization constraints to optimize the two-stream feature extraction layer;

[0078] The joint constraints are:

[0079]

[0080] is a hyperparameter that balances local (temporal stream feature extraction module) and global (spatial stream feature extraction module) predictions, and η is a hyperparameter that controls the strength of cross-modal alignment.

[0081] Preferably, a hierarchical parameter update mechanism is designed to optimize the training of the dual-stream feature extraction layer, including:

[0082] Group weight decay is used for the temporal stream feature extraction module, with δ=0.01 for the convolution layer and δ=0.001 for the normalization layer;

[0083] For the spatial stream feature extraction module, the learning rate is 0.0001 and δ=0.005;

[0084] Global gradient clipping, setting the threshold to 10.0, adaptive mixed precision training, using FP16 for the temporal stream feature extraction module and keeping FP32 for the spatial stream feature extraction module:

[0085] .

[0086] Beneficial effects: The present invention constructs a dual-stream feature extraction layer, using the time stream feature extraction module as a local temporal feature extractor, which is parallel or cross-stacked with the global attention mechanism of the spatial stream feature extraction module to achieve local-global feature complementarity; wherein, multi-scale convolution kernels are introduced in the time stream feature extraction module to capture time dependencies of different granularities; multi-scale features are adaptively weighted through the attention mechanism to solve the problem of the sensitivity of a single convolution scale to complex temporal changes and improve the robustness of the model; frequency-aware position coding and hierarchical sparse attention mechanism are integrated in the spatial stream feature extraction module to enhance the ability to analyze the global degradation trend modeling of vibration signal machines.

[0087] The present invention introduces a cross-layer attention mechanism to form a two-way interactive architecture, and improves the structure through MoH, breaking the limitations of traditional unidirectional feature fusion. It can dynamically combine the local vibration details of the temporal stream feature extraction module with the global degradation trend of the spatial stream feature extraction module through attention weights, realize two-way guidance of local details and global semantics, and greatly improve computational efficiency; improve anti-noise ability, suppress false detection of a single module, avoid problems such as TCN being sensitive to burst noise and the spatial stream feature extraction module being insensitive to local mutations, and achieve robust fusion through weight adjustment; in addition, cross-layer attention can better capture signal characteristics at each stage, the temporal stream feature extraction module captures early weak anomalies, and the spatial stream feature extraction module judges the global trend tilt, thereby improving prediction stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0089] In the attached figure:

[0090] Figure 1 This is a flow chart of the dual-stream time convolution enhanced interactive bearing life prediction method of the present invention;

[0091] Figure 2 It is a schematic diagram of the MSDA principle in the time stream feature extraction module of the present invention;

[0092] Figure 3 Schematic diagram of the working principle of the bidirectional interactive attention mechanism of the present invention;

[0093] Figure 4 Schematic diagram of the signal sliding window processing process of the present invention;

[0094] Figure 5 This is the RUL prediction curve of the bearings in the validation set B1_5;

[0095] Figure 6 This is the RUL prediction curve of the bearing in the validation set B2_4;

[0096] Figure 7 This is the RUL prediction curve of the bearings in the validation set B3_5. DETAILED DESCRIPTION

[0097] The following describes the embodiments of the present invention in conjunction with the accompanying drawings. The terms used in the embodiments of the present invention are only used to explain the specific embodiments of the present invention and are not intended to limit the present invention. The following describes the embodiments of the present application in conjunction with the accompanying drawings.

[0098] refer to Figures 1-4 As shown, a dual-stream time convolution enhanced interactive bearing life prediction method includes:

[0099] Collect original bearing vibration signal data, process and reconstruct the data;

[0100] The original signal sample point data is processed through a sliding window to obtain time features; the sliding window is used for data segmentation preprocessing, which enables the model to extract the most valuable information from the multivariate time series, reduce the amount of data, and extract key information; the original time series is represented as , where Ti represents the length of the time series and the length of the sliding window is , moving with a step length of Ls, the length of the overlapping area of ​​the two sliding windows is ,by The signal is intercepted in units for processing, and the input data for the estimation model is generated by applying a sliding time window to the sequence data; this process can be expressed as , T is the partition function, is the time window size;

[0101] Construct a two-stream feature extraction layer and input the reconstructed signal into the two-stream feature extraction layer to extract features;

[0102] Input the mentioned features into the multi-layer perceptron for RUL prediction;

[0103] The dual-stream feature extraction layer includes the following steps:

[0104] (1) We integrate the local temporal modeling capability of TCN with Multi-Scale Dilated Attention (MSDA) to construct a temporal stream feature extraction module (MSDA-TCN). We dynamically adjust the dilation rate through spectral entropy and combine sparse multi-head attention to capture multi-scale local features. Specifically:

[0105] We design parallel multi-branch dilated convolution layers to implement multi-scale mixed dilated convolution. Each layer contains three parallel dilated convolutions to extract features at different time scales. Convolution kernels with different dilation rates are used in parallel within the same TCN layer to capture local patterns of different granularities and then fuse the features. The dilation rate is dynamically adjusted based on the spectral entropy of the input features. The specific process is as follows:

[0106] Calculate the spectral entropy of the current input signal x ;

[0107] in It is the frequency domain component after FFT. The spectrum entropy reflects the complexity of the signal. High entropy value corresponds to high-frequency transient characteristics, and low entropy value corresponds to low-frequency trend characteristics.

[0108] Through lightweight MLP, H(f) is mapped to the expansion rate adjustment coefficient Δd∈[−0.5,0.5] to make the convolution kernel more focused on the key local interval:

[0109] ;

[0110] The corrected expansion rate is: ;

[0111] in, It is the pre-set baseline dilation rate for each independent attention head (or channel group) in the DilateFormer Multi-Scale Dilated Attention Mechanism (MSDA).

[0112] A multi-scale dilated attention mechanism (DilateFormer) is introduced to divide the multi-branch dilated convolution output of the TCN layer into N groups with C channels. Each group generates an independent query, key, and value matrix. Each group corresponds to an attention head of the MSDA, and each attention head uses a different dilation rate to capture contextual information at different time scales and achieve feature granularity alignment.

[0113] For each attention head’s query position (t,c), the expansion rate is used within the sliding window. Sparse selection of key-value pairs; the positions of keys and values ​​start from the query position (t, c) and are calculated according to the expansion step size. Sample uniformly within the window.

[0114] The key value position set of the i-th head is:

[0115] ;

[0116] The time step position of the selected key and value.

[0117] Reduce computational complexity by sparse sampling (from O(T 2 ) is reduced to O(N×w 2 T)) where w is the window size.

[0118] For each attention head, we generate query, key, and value matrices, and calculate local context features by scaling the dot product attention:

[0119] ;

[0120] in, is the dimension of each attention head, Represents a set of real numbers, used to describe the matrix dimension; T is the number of rows, and each time step corresponds to a query / key / value vector.

[0121] The output of each attention head is spliced ​​along the channel dimension, and the multi-scale features are aggregated through the linear transformation layer to retain the multi-scale context information and output :

[0122] ;

[0123] The temporal feature extraction module (MSDA-TCN module) enables the model to use a smaller expansion rate in high-entropy segments (complex vibrations) to capture details, and a larger expansion rate to model trends in low-entropy segments (smooth operation), effectively capturing local features. The TCN expansion rate prediction MLP and the MSDA attention head share a weight matrix, enabling parameter reuse, thereby reducing the number of parameters and improving operational efficiency.

[0124] (2) By integrating frequency-aware relative position encoding (FE) with a hierarchical sparse attention mechanism, a spatial stream feature extraction module (FE-iTransformer) is constructed to enhance the ability to analyze the global degradation trend modeling of vibration signal machines. Specifically, the module is: (Time Decay Mask)

[0125] For the input sequence Perform FFT to extract the frequency component with the largest amplitude , adding frequency-aware relative position bias to the attention score:

[0126] Frequency modulation position encoding: ;

[0127] Relative position offset: ;

[0128] in, is a learnable parameter, t is the time step index, is the learnable scaling factor;

[0129] Introducing the Time Decay Mask to increase the focus on recent time steps and improve the accuracy of later predictions:

[0130] ;

[0131] in, is the learnable decay coefficient, i and j are the time indexes of query and key respectively;

[0132] Corrected attention calculation: .

[0133] In order to reduce the computational overhead of long sequences (such as T>1000), a hierarchical sparse attention is designed: the input sequence is divided into M segments , length of each segment L=T / M;

[0134] Perform standard self-attention within each segment, computing the amount ;

[0135] Perform cross-segment attention on the segment representative vector (mean of each segment), and calculate the amount And fuse local and global features through gating mechanism:

[0136] ;

[0137] in .

[0138] in, is the Hadamard product, σ(⋅) is the Sigmoid activation function, is the learnable weight matrix.

[0139] (3) Design a bidirectional cross-layer attention coordination mechanism to interact the features of the temporal stream feature extraction module and the spatial stream feature extraction module, and enhance the degradation representation capability through hierarchical feature alignment and dynamic weight adjustment.

[0140] Feature interactions include:

[0141] Enhancement from local details to global semantics: Use cross-layer attention to hierarchically fuse the extracted features, use the features extracted by the temporal stream feature extraction module as the query, and the features extracted by the spatial stream feature extraction module as the key / value, to build detail enhancement attention:

[0142] ;

[0143] Dynamically adjust the contribution of local anomalies in global features through attention weights to generate fusion features , injecting transient impact features (such as early weak faults) captured by the time stream feature extraction module into the global representation of the spatial stream feature extraction module to improve the model's sensitivity to early faults:

[0144] ;

[0145] Correction of local features by global trends: Using the global trend of the spatial stream feature extraction module as the query and the multi-scale features of the temporal stream feature extraction module as the key / value, calculate the attention weight:

[0146] ;

[0147] Modify the weight distribution of local features through global trends: ;

[0148] Adopting a learnable gating mechanism to achieve dynamic fusion of bidirectional features:

[0149] ;

[0150] Perform hierarchical feature interaction in L network layers and fuse the features of each layer Splicing along the channel dimension, the fusion features generated in each layer are , B is the batch size, T is the time step, C L is the number of channels in the Lth layer, " indicates splicing along the channel dimension, and constructing multi-scale representation by splicing along the channel dimension:

[0151] .

[0152] After splicing, the features retain the time dimension T, and the channel dimension is expanded to the sum of the number of channels in each layer; it can retain features at different levels and enhance the model's expressiveness.

[0153] Among them, feature interaction is performed at each layer instead of only fusion at the last layer, and the amount of computation is reduced by sharing some parameters.

[0154] In another embodiment, to address the limited normalization effect of traditional LayerNorm when processing non-stationary degraded signals, the normalization layer is improved, and Dynamic Tanh (DyT) is adopted as an innovative normalization alternative. DyT significantly simplifies deep model design by combining a dynamic scaling factor with the tanh function. It eliminates the need to calculate input statistics (mean, variance) and dynamically adjusts the input range only through a learnable parameter α, significantly reducing computational overhead. It can also directly replace existing normalization layers without modifying the activation function or network structure, significantly reducing migration costs. Specifically:

[0155] Replace each normalization layer with a DyT component, leaving other components unchanged:

[0156] ;

[0157] in, To dynamically adjust the factor, the input scale is automatically optimized by gradient descent, and the default initialization is 0.5; and The feature calibration parameters are channel-level affine parameters, initialized to all 1s and all 0s;

[0158] DyT enhances the model's expressiveness in the following ways:

[0159] The temporal convolution layer extracts local temporal features, and DyT replaces the original normalization layer to dynamically adjust the convolution output:

[0160] ;

[0161] The multi-head attention mechanism captures long-range dependencies. DyT is inserted before and after the attention calculation to stabilize the attention weight distribution:

[0162] ;

[0163] By adjusting the fusion scale of cross-layer features through DyT, the stability of gradient flow is enhanced, and the output flow is:

[0164] .

[0165] In the task of predicting RUL, based on the above, MLP takes the degradation information extracted by the feature extraction layer as input; After flattening, it passes through two layers of MLP. MLP takes the fused features as input, performs approximate calculations through the hidden layer, and obtains the RUL estimated by the output layer; the ReLu function is used as the activation function; the dropout layer is introduced in the feature fusion layer to randomly delete the connections of some neurons with probability p to prevent the model from overfitting.

[0166] In another embodiment, a multi-objective joint optimization is designed, specifically:

[0167] Design multi-scale dynamic weighted loss constraints: In the time stream feature extraction module, multi-scale time series features are extracted through different expansion factors, and prediction results are generated at each scale. ; The spatial stream feature extraction module outputs the global prediction results , and use the mean square error as the loss function, extract the self-attention matrix of the spatial stream feature module Extract the importance weights of each time scale, realize the dynamic allocation of attention weights, and ensure , k is the number of expansion factors; the error of each scale of the time stream feature extraction module is determined by the attention weight Weighted; According to the global dependency of the input sequence, the loss weight of each time scale is dynamically allocated to alleviate the multi-scale fusion deviation; including 、 , to conduct information interaction, the loss of the spatial stream feature extraction module directly guides the loss weighting of the temporal stream feature extraction module, realizing explicit collaboration between branches;

[0168] The specific formula is:

[0169] ;

[0170] ;

[0171] ; ; is the attention weight A(d) of the d-th scale: the attention matrix corresponding to the d-th time scale;

[0172] Introduce regularization constraints and add feature alignment loss Control the spatial consistency of the hidden layer features of the temporal stream feature extraction module and the spatial stream feature extraction module; and enhance information fusion through the hidden layer representation L2 distance constraint; and adjust the convolution kernel weight of the temporal stream feature extraction module. L1 regularization is applied to achieve sparsity constraints; the attention matrix A is constrained to be close to the identity matrix through the Frobenius norm. This can promote modal fusion, forcing the two branches to learn consistent feature representations, avoiding feature space fragmentation due to architectural differences, and enhancing the model's robustness to input perturbations.

[0173] The specific formula is:

[0174] ;

[0175] ;

[0176] ;

[0177] in, 、 is the hidden layer feature vector of the TCN and spatial stream feature extraction module, 、 is the regularization weight coefficient;

[0178] Combined multi-scale dynamic weighted loss constraints and regularization constraints to optimize the two-stream feature extraction layer;

[0179] The joint constraints are: ;

[0180] is a hyperparameter that balances local (temporal stream feature extraction module) and global (spatial stream feature extraction module) predictions, and η is a hyperparameter that controls the strength of cross-modal alignment;

[0181] (3) Designing a hierarchical parameter update mechanism to optimize the training of the dual-stream feature extraction layer, improve computational efficiency, and avoid gradient vanishing and gradient exploding, including:

[0182] Group weight decay is used for the temporal stream feature extraction module, with δ=0.01 for the convolution layer and δ=0.001 for the normalization layer;

[0183] For the spatial stream feature extraction module, the learning rate is 0.0001 and δ=0.005;

[0184] Global gradient clipping, setting the threshold to 10.0, adaptive mixed precision training, using FP16 for the temporal stream feature extraction module and keeping FP32 for the spatial stream feature extraction module:

[0185] .

[0186] To evaluate the prediction effect of this method, three indicators, mean absolute error (MAE), root mean square error (RMSE), and the designed scoring function (Score), are used for performance evaluation. MAE and RMSE reflect the absolute deviation between the predicted value and the true value. RMSE is more sensitive to outliers (such as sudden failure signals). MAE and RMSE are calculated as follows: where RULt is the true RUL value of the rolling bearing at time t. is the predicted RUL value of the rolling bearing at time t; n is the number of samples in the test set:

[0187] ;

[0188] ;

[0189] ;

[0190] The test set and training set under three working conditions in the XJTU-SY dataset are divided into six tasks, as shown in Table 1:

[0191] Table 1 Training task division table

[0192]

[0193] Experimental results:

[0194]

[0195] refer to Figure 5-Figure 7 Shown are the RUL prediction curves of the bearings in the validation sets B1_5, B2_4, and B3_5 under three working conditions;

[0196] The RUL prediction curve, based on the First Detection Time (FDT), provides a more accurate estimate of the remaining operating life of a bearing. The closer the predicted curve (orange line) is to the true curve (blue line), the more accurate the prediction. This indicates that the predicted RUL closely tracks the true RUL. Early on, the tracking effect is better, with less fluctuation in the predicted results, staying close to the true life. Later, as the bearing nears failure, more degradation information is present, gradually deviating from the true value curve and experiencing increased fluctuations.

[0197] Ablation experiment:

[0198] To demonstrate the advanced performance of the method of the present invention and verify the role of the local-global dual-stream structure in the model, ablation tests were conducted. Six comparison methods were used on the combination of prediction networks. To verify the importance of local and global features in sequence modeling, the iTransformer and TCN were removed to obtain a single-stream structure. The six combinations are: TCN, iTransformer, TCN+iTransformer, TCN+MSDA, iTransformer+MoH, and TCAIT. The ablation test results are shown in Table 2. The comparison of MAE, RMSE error values ​​and accuracy of different models on different tasks is shown in Table 2:

[0199] Table 2 Summary of ablation experiment results

[0200]

[0201] Comparison of the MAE, RMSE, and Score results in all groups shows that the MAE (0.315–0.325), RMSE (0.362–0.383), and Score (0.978–0.982) of the TCAIT method of the present invention are all stable and optimal. When TCN and iTransformer are used alone, the model performance decreases significantly. Although the fluctuation range of MAE and RMSE is small, the Score of TCN (0.901–0.915) decreases by 9.3%, and the Score of iTransformer (0.817–0.842) decreases by 8.3%, indicating that the performance of a single model is greatly affected by the local feature distribution of the data, and that information from the spatial dimension can provide more information for RUL prediction, thereby improving the prediction accuracy of the model.

[0202] Adding cross-layer attention to the TCN reduced MAE by approximately 22.4%, RMSE by 17.9%, and the average accuracy score by approximately 4.2%. iTransformer also reduced MAE by an average of approximately 34.5%, RMSE by 29.6%, and the average accuracy score by approximately 4.6%. This demonstrates that the attention mechanism effectively improves feature utilization, enhancing the model's feature selection capabilities, temporal dependency modeling capabilities, and multi-scale feature extraction. Multi-scale TCN captures more detailed features of vibration signals, which is crucial for model performance, especially for improving RUL prediction performance under multiple operating conditions and fault modes.

[0203] The above results show that the method of the present invention can achieve satisfactory accuracy in RUL prediction. The scores of the TCAIT model in the six data sets are close to or exceed 0.98, indicating that it is suitable for the high-precision prediction requirements under complex working conditions of the XJTU-SY dataset.

[0204] Comparative experiment:

[0205] The proposed method was compared with TCN-BiGRU, TCN-informer, MS-TCN++, AFPFM, and CRAformer. Each model completed Task A with the same hyperparameter settings. To mitigate the effects of randomness, the experiment was repeated three times for each task, and the MAE, RMSE, and Score values ​​were recorded as indicators. The experimental results are shown in Table 3:

[0206] Table 3 Summary of comparative experimental results

[0207]

[0208] Experimental analysis:

[0209] The comprehensive performance of the method of the present invention is the best among the six groups, with an average MAE of 0.27 and an average RMSE of 0.402, both of which are the lowest. The score is far ahead of other models, reaching 0.992. The TCAIT model dynamically integrates features from different levels and uses dilated convolution to capture long-term dependencies, solving the information attenuation problem caused by hierarchical stacking in traditional models and having high information transmission efficiency. Through the global optimization of iTransformer, a multi-dimensional time series independent encoding strategy is adopted to avoid feature confusion and enhance the expression ability of complex time patterns. Combined with lightweight design, it improves running speed and reduces computational overhead, making it suitable for deployment in industrial scenarios. Compared with CRAformer, the model with the highest score among the other five models, the method of the present invention has a MAE 26.3% lower and an RMSE 37.2% lower, and a score 1.85% higher, which proves the effectiveness and superiority of the method proposed in this article.

[0210] Computational efficiency analysis:

[0211] To verify the computational efficiency of the network constructed in this paper, the following experiment was designed: We maintained the same number of encoder layers as in TCAIT, building TCN-BiGRU, TCN-informer, MS-TCN++, AFPFM, and CRAformer networks. We compared the time taken to process time series of the same length under the same run cycle. Task A was selected for the experiment, running for 150 epochs. The training times are shown in Table 4:

[0212] Table 4 Training time of different models

[0213]

[0214] Table 4 shows that the network constructed by the present invention has the shortest runtime. Compared to the longest-running TCN-BiGRU, the computation time is reduced by 54.5%; compared to the TCN-informer, the computation time is reduced by 48.5%; compared to the MS-TCN++, the computation time is reduced by 52.7%; and compared to the AFPFM, the computation time is reduced by 49.4%. Even compared to the shortest-running CRAformer, the network constructed by the present invention still reduces the runtime by 39.6%. This demonstrates that the dual-stream mechanism combined with the attention mechanism to optimize the representation of input features can effectively improve the computational efficiency of bearing RUL prediction speed.

[0215] The above describes the embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. After knowing the contents described in the present invention, ordinary technicians in this technical field can make several equivalent changes and substitutions without departing from the principles of the present invention. These equivalent changes and substitutions should also be regarded as falling within the scope of protection of the present invention.

Claims

1. A dual-stream time convolution enhanced interactive bearing life prediction method, characterized in that: include: Collect original bearing vibration signal data, process and reconstruct the data; Construct a two-stream feature extraction layer and input the reconstructed signal into the two-stream feature extraction layer to extract features; Input the mentioned features into the multi-layer perceptron for RUL prediction; Building a two-stream feature extraction layer includes: The local temporal sequence modeling capability of TCN is integrated with multi-scale dilated attention to construct a temporal stream feature extraction module. The dilation rate is dynamically adjusted through spectral entropy, and sparse multi-head attention is combined to capture multi-scale local features. The spatial stream feature extraction module is constructed by integrating frequency-aware position encoding and hierarchical sparse attention mechanism; A bidirectional cross-layer attention coordination mechanism is designed to interact the features of the temporal stream feature extraction module and the spatial stream feature extraction module, enhancing the degradation representation capability through hierarchical feature alignment and dynamic weight adjustment. Dynamic Tanh normalization layer is used to replace the traditional LayerNorm, specifically: Replace each normalization layer with a DyT component, leaving other components unchanged: ； in, To dynamically adjust the factor, the input scale is automatically optimized by gradient descent, and the default initialization is 0.5; and The feature calibration parameters are channel-level affine parameters, initialized to all 1s and all 0s; The temporal convolution layer extracts local temporal features and dynamically adjusts the convolution output: ； Insert DyT before and after the attention calculation to stabilize the attention weight distribution: ； By adjusting the fusion scale of cross-layer features through DyT, the stability of gradient flow is enhanced, and the output flow is: ； The two-stream feature extraction layer is optimized, including: Design multi-scale dynamic weighted loss constraints: In the time stream feature extraction module, multi-scale time series features are extracted through different expansion factors, and prediction results are generated at each scale. ; The spatial stream feature extraction module outputs the global prediction results , and use the mean square error as the loss function, extract the self-attention matrix of the spatial stream feature module Extract the importance weights of each time scale, realize the dynamic allocation of attention weights, and ensure , k is the number of expansion factors; the error of each scale of the time stream feature extraction module is determined by the attention weight weighted; Introduce regularization constraints and add feature alignment loss Control the spatial consistency of the hidden layer features of the temporal stream feature extraction module and the spatial stream feature extraction module; and enhance information fusion through the hidden layer representation L2 distance constraint; and adjust the convolution kernel weight of the temporal stream feature extraction module. Apply L1 regularization to achieve sparsity constraints; constrain the attention matrix A to be close to the identity matrix through the Frobenius norm; The two-stream feature extraction layer is optimized by combining multi-scale dynamic weighted loss constraints and regularization constraints.

2. The dual-stream time convolution enhanced interactive bearing life prediction method according to claim 1 is characterized by: The time stream feature extraction module includes: Design a parallel multi-branch dilated convolution layer, each layer contains three parallel dilated convolutions to extract features at different time scales. Use convolution kernels with different dilation rates in parallel in the same TCN layer to capture local patterns of different granularities and then fuse the features. A multi-scale dilated attention mechanism is introduced to divide the multi-branch dilated convolution output of the TCN layer into N groups with C channels. Each group generates an independent query, key, and value matrix. Each group corresponds to an attention head of the MSDA, and each head uses a different dilation rate to capture contextual information at different time scales and achieve feature granularity alignment. For each attention head’s query position, the expansion rate is used within the sliding window. Sparse selection of key-value pairs; Compute local context features by scaling dot product attention: ； in, is the dimension of each attention head; The output of each attention head is spliced ​​along the channel dimension, and multi-scale features are aggregated through the linear transformation layer to retain the multi-scale context information and output : 。 3. The dual-stream time convolution enhanced interactive bearing life prediction method according to claim 2 is characterized by: Dynamically adjust the expansion rate based on the spectral entropy of the input features, including: Calculate the spectral entropy H(f) of the current input signal x; Through lightweight MLP, H(f) is mapped to the expansion rate adjustment coefficient Δd, so that the convolution kernel is more focused on the key local interval; ； The corrected expansion rate is: .

4. The dual-stream time convolution enhanced interactive bearing life prediction method according to claim 2 is characterized by: In the process of building the spatial stream feature extraction module: For the input sequence Perform FFT to extract the frequency component with the largest amplitude , adding frequency-aware relative position bias to the attention score: Frequency modulation position encoding: ; Relative position offset: ; in, is a learnable parameter, t is the time step index, is the learnable scaling factor; Introduce a time decay mask to increase attention to recent time steps: ； in, is the learnable decay coefficient, i and j are the time indexes of query and key respectively; Corrected attention calculation: .

5. The dual-stream time convolution enhanced interactive bearing life prediction method according to claim 4 is characterized by: The process of building the spatial stream feature extraction module also includes: Divide the input sequence into M segments {X1, X2, ..., XM}, each segment length L = T / M; Perform standard self-attention within each segment, computing the amount ; Perform cross-fragment attention on the fragment representative vector, the computational And fuse local and global features through gating mechanism: ； in , is the Sigmoid activation function, is the learnable weight matrix.

6. The dual-stream time convolution enhanced interactive bearing life prediction method according to claim 2, characterized in that: The feature interaction includes: The features extracted by the temporal stream feature extraction module are used as the query, and the features extracted by the spatial stream feature extraction module are used as the key / value to construct detail-enhanced attention: ； Dynamically adjust the contribution of local anomalies in global features through attention weights to generate fusion features , injecting the transient impact features captured by the time stream feature extraction module into the global representation of the spatial stream feature extraction module: ； The global trend of the spatial stream feature extraction module is used as the query, and the multi-scale features of the temporal stream feature extraction module are used as the key / value to calculate the attention weight: ； Modify the weight distribution of local features through global trends: ; Adopting a learnable gating mechanism to achieve dynamic fusion of bidirectional features: ； Perform hierarchical feature interaction in L network layers and fuse the features of each layer Splicing along the channel dimension, the fusion features generated in each layer are , B is the batch size, T is the time step, C L is the number of channels in the Lth layer, " indicates splicing along the channel dimension, and constructing multi-scale representation by splicing along the channel dimension: 。 7. The dual-stream time convolution enhanced interactive bearing life prediction method according to claim 6, characterized in that: Each of the network layers performs feature interaction and reduces the amount of computation by sharing some parameters.

8. The dual-stream time convolution enhanced interactive bearing life prediction method according to claim 1 is characterized by: Design a hierarchical parameter update mechanism to optimize the training of the two-stream feature extraction layer, including: The time stream feature extraction module uses group weight decay, convolutional layer =0.01, normalization layer =0.001; For the spatial stream feature extraction module, the learning rate is 0.0001. =0.005; Global gradient clipping, setting the threshold to 10.0, adaptive mixed precision training, using FP16 for the temporal stream feature extraction module and keeping FP32 for the spatial stream feature extraction module: 。

Images