Multi-source parallel bilinear co-attention fusion mechanical equipment life prediction method

Abstract

The invention relates to a multi-source parallel bilinear confocal fusion mechanical equipment life prediction method. The method comprises the following steps: firstly, preprocessing each type of monitoring signals; inputting the preprocessed data into a subnet, and carrying out deep feature extraction to obtain high-dimensional multi-source specific features; high-dimensional multi-source specific features are integrated through a multi-source feature fusion module, and a final predicted RUL prediction result is obtained through a regression prediction reasoning module; according to the method, heterogeneity of multi-source data is processed through a parallel subnet, feature extraction is enhanced by adopting a multi-wavelet kernel aggregation structure and a full-dimensional dynamic structure, and multi-source data features are effectively fused through a factorization bilinear pooling attention module; the comprehensive processing capability of the model on multi-source data is enhanced, and the accuracy and reliability of RUL prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of remaining service life prediction of mechanical equipment, and specifically relates to a multi-source parallel bilinear co-attention fusion method for predicting the service life of mechanical equipment. Background Art

[0002] With the rapid development of measurement engineering, information technology, and decision theory, the integration and intelligence of mechanical systems in key sectors such as energy, aerospace, and transportation have been significantly improved. While this progress has brought significant benefits, it has also led to increased interdependence between system components and an increase in influencing factors, which in turn increases the risk of mechanical equipment failure. To address this challenge, remaining useful life (RUL) prediction technology has emerged. By analyzing accumulated monitoring data, it can predict the future health status of mechanical equipment.

[0003] While traditional methods based on single-type data have shown some effectiveness in RUL prediction, they suffer from an inherent flaw: relying solely on a single type of sensor data for inference, they lack reliability and have a limited perspective. To address this issue, industrial sites are increasingly adopting multi-type sensor monitoring strategies, deploying additional sensor types and expanding the monitoring scope to capture more comprehensive and rich information about machine health status. Multi-type sensor monitoring technology provides rich and diverse data, which facilitates a deeper understanding of machine degradation patterns and provides a solid foundation for RUL prediction techniques based on multi-source data, enabling more accurate predictions of machine RUL.

[0004] Existing prediction methods based on multi-source data still face the following challenges: First, the inherent non-stationarity of multi-type sensor monitoring data is difficult to handle. This non-stationarity, characterized by its randomness and changing statistical properties, hinders the full utilization of fault-discriminating feature information. Under the assumption of stationarity, models cannot capture evolving patterns during long-term monitoring. Second, after training, most models degenerate into fixed mapping functions and are unable to flexibly adapt to real-time online data. This is particularly problematic in industrial environments, where machine operating conditions fluctuate over time. If the distribution of future monitoring data differs significantly from that of historical data, relying on fixed weights and parameters determined during the training phase will lead to a decrease in model accuracy and reliability. Third, existing multi-source feature fusion mechanisms are still imperfect. Fusion mechanisms rely on simple linear fusion techniques such as direct concatenation or element-by-element addition. These methods often fail to achieve satisfactory fusion results when dealing with multi-source data with significantly different distributions. Furthermore, these methods fail to fully capture the interactions between features of different data types and lack effective modeling of the complementarity and correlation information in multi-source data. Summary of the Invention

[0005] In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to provide a multi-source parallel bilinear co-attention fusion mechanical equipment life prediction method, which processes the heterogeneity of multi-source data through parallel subnetworks, adopts multi-wavelet kernel aggregation structure and full-dimensional dynamic structure to enhance feature extraction, and effectively fuses multi-source data features through factorized bilinear pooling attention module; it not only enhances the model's comprehensive processing ability for multi-source data, but also improves the accuracy and reliability of RUL prediction.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A multi-source parallel bilinear co-attention fusion method for mechanical equipment life prediction first preprocesses each type of monitoring signal to eliminate outliers; then the preprocessed data is input into a subnetwork for in-depth feature extraction to obtain high-dimensional multi-source specific features; then the high-dimensional multi-source specific features are integrated through a multi-source feature fusion module, and the final predicted RUL prediction result is obtained through a regression prediction and inference module.

[0008] A multi-source parallel bilinear co-attention fusion mechanical equipment life prediction method includes the following steps:

[0009] Step 1: Obtain a multi-source dataset of the entire life cycle of mechanical equipment Where M is the number of sequence-tag pairs, M is the number of monitoring signal types, are the channels for each signal type, is the number of sampling data points in each measurement for each signal type; multi-source dataset for the entire life of mechanical equipment It consists of time series measurements from multiple sensors, where the sampling time interval is kept constant;

[0010] Step 2: Apply the Hampel filter to detect and remove outliers in each signal type. The Hampel filter uses the median absolute deviation (MAD) within a sliding window to identify outliers and is applied to the input sequence. The process of each sliding window of size p is as follows:

[0011]

[0012] Where Mdn(·) is the median of the input sequence, is the MAD value of the sequence, and λ is the predefined anomaly detection threshold;

[0013] Step 3: Parallel computing is used to process multi-source data streams. Each signal type is subjected to targeted feature extraction by a corresponding sub-network. Each sub-network is composed of a basic unit called a spatiotemporal representation encoding (STRE) module. Multiple STRE modules are stacked to gradually refine features. The STRE module achieves two goals: spatial structure mining and temporal dependency capture. Spatial structure mining enhances the model's ability to distinguish potential fault characteristics by analyzing the spatial relationships of signal data. Temporal dependency tracking reveals subtle differences between similar signatures in complex scenarios by monitoring the changes in signal data over time. Within each STRE module, these functions are achieved through dynamic multiwavelet aggregation encoding (DMAE) layers and multi-scale convolutional gated recurrent unit (MCGRU) layers.

[0014] Step 4: Integrate the wavelet function as a kernel into the convolutional neural network (CNN) to handle non-stationarity and perform multi-domain analysis; for a small batch input N i The j-th output channel C out The calculation is expressed as:

[0015]

[0016] Where W represents the weight of the wavelet kernel, b is the bias term, * d represents the dilated convolution with coefficient d, ψ(·) is the wavelet basis function, C in and C out are the input and output channels respectively, X is the input data, m is the sequence number of the input data, α is the size factor in the wavelet transform, β is the time shift factor, and t represents the time step;

[0017] The wavelet kernel convolution is extended to multi-wavelet kernel aggregation to improve the ability to analyze non-stationary data; multiple wavelet kernels with different basis functions are integrated in one layer, and their outputs are aggregated to enhance the ability to capture discriminative information and give full play to the complementary advantages of different types of wavelets. The calculation process is as follows:

[0018]

[0019] Where, ω u (X) is the dynamic weight of the u-th aggregation kernel, which is calculated using the attention mechanism conditioned on the input X. U is the total number of aggregation kernels. ⊙ and * represent the Hadamard product and convolution operators, respectively. and denote the weight and bias of the u-th kernel respectively;

[0020] The dynamic properties are extended to all dimensions of the kernel space to construct the DMAE layer of full-dimensional dynamic structure; specifically, for U The kernel of , where g is the number of group connections, the DMAE layer introduces different attention scalars ω = (ωou ,ω wu ,ω su ,ω cu ) to dynamically adjust the importance of specific cues within the kernel, these scalars are assigned to different dimensions: ω wu Acting on the convolution kernel, are three extended attention scalars calculated on the group input dimension, kernel size dimension, and output dimension respectively. Then the extended operation process is:

[0021]

[0022] Step 5: Construct the MCGRU layer, capture temporal features through convolution operations, and add residual connections to learn hierarchical representations; reset gate R t and update gate Z t Control the flow of information, processing input and previous hidden state:

[0023]

[0024] Where, is the input, h t-1 is the hidden state of the previous step;

[0025] For the candidate hidden state h t-1 , a hierarchical design is adopted to capture a larger context, and the hierarchical design is expressed as follows:

[0026]

[0027] Where, ψ t is a convolution transformation, MSP(·) represents multi-scale pooling, which aims to capture representations of different resolutions; update gate Z t Control the previous hidden state h by element-wise multiplication t-1 Hide the current state The impact of

[0028]

[0029] Final representation The 1×1 convolution of the converted hidden state and the current input is combined, that is:

[0030]

[0031] Step 6: Design the factorized bilinear pooling attention (FBPA) fusion module to reduce the complexity. The FBPA fusion module adopts a pairwise fusion strategy for two signal types. and The fusion process, feature tensor and is first vectorized into and in and The bilinear fusion operation computes the pairwise interactions between a and b as follows:

[0032]

[0033] In the formula is the projection weight matrix, ∥ represents the connection operation, and o is the output dimension; W i Decomposed into two low-rank matrices and The formula is as follows:

[0034]

[0035] In the formula ρ is the decomposition rank, is a full 1 vector; according to the above derivation, to obtain the output fusion representation z of size o, it is necessary to learn two 3D weight tensors and Then by copying on o channels and To reshape these tensors, we can use the following:

[0036]

[0037] Then the matrix and The Hadamard product of is used to calculate the fused representation z, and then a sum pooling operation is performed on non-overlapping windows of size ρ:

[0038]

[0039] Extending the above second-order bilinear fusion to high-order fusion, the fusion process is as follows:

[0040]

[0041] In the formula is the intermediate feature vector before the k-th process summation and pooling operation, Corresponding to the (k-1)th process, k∈{1,…,K} is the index of the stacking process; the final fusion output z is obtained by connecting K processes:

[0042]

[0043] Step 7: Set up the co-attention mechanism in the FBPA module to learn the joint attention weights for each pair of signal types; and First, use the above FBP algorithm to fuse the original multi-source data features to obtain common features Then, the common features are projected into the multi-head computation space using a 1×1 convolutional layer and a ReLU activation function transformation block; Apply the Softmax layer on the channel dimension as follows:

[0044]

[0045] Where h is the number of heads, i = (0, ..., c-1); in each head, the original multi-source data features are combined with the transformed common features, and then summed over channel c to obtain weighted features and as follows:

[0046]

[0047] right and The fusion output of is finally obtained through the above FBP algorithm, that is, Finally, based on the above pairwise fusion operation, the multi-source fusion output is obtained as follows:

[0048]

[0049] where F1,…,F M is the raw monitoring data of each signal type, M is the number of detected signal types, Indicates the fusion operation using the FBPA module;

[0050] Step 8: Perform regression prediction inference and represent the fusion F z Enter the regression prediction inference module to obtain the RUL prediction value at each time step:

[0051]

[0052] where f r is the regression function, l is the number of FC layers in the module, and the first layer is defined as Each FC layer takes the output of the previous layer as input, with weights and bias The final layer f r It contains only one neuron and uses the Sigmoid function output to monitor the health status of the machine.

[0053] The signal type fusion process in step 6 can be seamlessly extended to more signal types.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] The present invention adopts a multi-source parallel bilinear co-attention fusion framework, processes the heterogeneity of multi-source data through multi-subnet parallel processing, adopts multi-wavelet kernel aggregation structure and full-dimensional dynamic structure to enhance feature extraction, and effectively fuses multi-source data features through factorized bilinear pooling attention module, thereby improving the accuracy and robustness of RUL prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Flowchart of an embodiment of the present invention.

[0057] Figure 2 This is the parallel sub-network data processing flow of an embodiment of the present invention.

[0058] Figure 3 This is a structural diagram of STRE according to an embodiment of the present invention.

[0059] Figure 4 This is the entire calculation flow chart of the FBPA module in an embodiment of the present invention.

[0060] Figure 5 The RUL prediction results of the embodiment of the present invention and the predicted mean and corresponding intervals of the proposed framework under each working condition, where (a) is the working condition of condition 1, (b) is the working condition of condition 2, (c) is the working condition of condition 3, and (d) is the working condition of condition 4. DETAILED DESCRIPTION

[0061] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0062] Reference Figure 1 and Figure 2 A multi-source parallel bilinear co-attention fusion mechanical equipment life prediction method includes the following steps:

[0063] Step 1: Obtain a multi-source dataset of the entire life cycle of mechanical equipment Where N is the number of sequence-tag pairs, M is the number of monitored signal types, are the channels for each signal type, is the number of sampling data points in each measurement for each signal type; multi-source dataset for the entire life of mechanical equipment It consists of time series measurements from multiple sensors, where the sampling time interval remains constant and each signal type reflects the machine's operating status from a different perspective;

[0064] Step 2: Actual industrial data often contains outliers. These outliers are measurements that significantly deviate from the normal pattern and may have an adverse effect on the accuracy of model predictions. To address this issue, we first apply a Hampel filter to detect and remove outliers in each signal type. The Hampel filter uses the median absolute deviation (MAD) within a sliding window to identify outliers and is applied to the input sequence. The process of each sliding window of size p is as follows:

[0065]

[0066] Where Mdn(·) is the median of the input sequence, is the MAD value of the sequence, and λ is the predefined anomaly detection threshold;

[0067] Step 3: Given the inherent heterogeneity of different sensors in terms of characteristics and sensitivity, the framework uses parallel computing to process multi-source data streams to avoid data aliasing and interference, ensuring effective mining and feature extraction of specific information from multi-source data; each signal type is targeted by a corresponding sub-network for feature extraction, and each sub-network is composed of a basic unit spatiotemporal representation encoding (STRE) module. Multiple STRE modules are stacked to gradually refine features; these STRE modules are designed specifically for monitoring the intrinsic properties of signal data, aiming to achieve two core goals: spatial structure mining and temporal dependency capture; spatial structure mining enhances the model's ability to distinguish potential fault features by analyzing the spatial relationship of signal data; temporal dependency tracking reveals subtle differences in similar signatures in complex scenarios by monitoring changes in signal data over time; these two goals are crucial for identifying overall trends and degradation patterns that are not obvious in static snapshots; Figure 3 As shown in Figure 2, in each STRE module, the above functions are implemented through the dynamic multi-wavelet aggregation encoding (DMAE) layer and the multi-scale convolutional gated recurrent unit (MCGRU) layer;

[0068] Step 4: Multi-source monitoring data reflects multiple random processes that change over time. These processes have changing statistical characteristics, revealing their non-stationary nature. This poses a challenge to the model in capturing the dynamic changes in the health status of the machine. Therefore, the method integrates the wavelet function as the kernel into the convolutional neural network (CNN) to handle non-stationarity and perform multi-domain analysis to improve performance. For the small batch input N i The j-th output channel C out The calculation is expressed as:

[0069]

[0070] Where W represents the weight of the wavelet kernel, b is the bias term, * drepresents the dilated convolution with coefficient d, ψ(·) is the wavelet basis function, C in and C out are the input and output channels respectively, X is the input data, m is the sequence number of the input data, α is the size factor in the wavelet transform, β is the time shift factor, and t represents the time step.

[0071] Since the effect of wavelet analysis is affected by the characteristics of sensor data and prediction requirements, and the selection of appropriate wavelet basis functions is time-consuming and subjective, the wavelet kernel convolution is extended to multi-wavelet kernel aggregation to improve the ability to analyze non-stationary data. This method integrates multiple wavelet kernels with different basis functions in one layer, aggregates their outputs to enhance the ability to capture discriminant information, and gives full play to the complementary advantages of different types of wavelets. The calculation process is as follows:

[0072]

[0073] Where, ω u (X) is the dynamic weight of the u-th aggregation kernel, which is calculated using the attention mechanism conditioned on the input X. U is the total number of aggregation kernels. ⊙ and * represent the Hadamard product and convolution operators, respectively. and denote the weight and bias of the u-th kernel respectively;

[0074] The dynamic properties are extended to all dimensions of the kernel space to construct the DMAE layer of full-dimensional dynamic structure; specifically, for U The kernel of , where g is the number of group connections, the DMAE layer introduces different attention scalars ω = (ω ou ,ω wu ,ω su ,ω cu ) to dynamically adjust the importance of specific cues within the kernel, these scalars are assigned to different dimensions: ω wu Acting on the convolution kernel, are three extended attention scalars calculated on the group input dimension, kernel size dimension, and output dimension respectively. Then the extended operation process is:

[0075]

[0076] Step 5: To fully utilize the temporal dependency in the lifecycle monitoring signal data, construct the MCGRU layer, capture the temporal features through convolution operations, and add residual connections to learn hierarchical representations; reset gate R t and update gate Z t Control the flow of information, processing input and previous hidden state:

[0077]

[0078] Where, is the input, h t-1 is the hidden state of the previous step;

[0079] For the candidate hidden state h t-1 , a hierarchical design is adopted to capture a larger context, and the hierarchical design is expressed as follows:

[0080]

[0081] Where, ψ t is a convolution transformation, MSP(·) represents multi-scale pooling, which aims to capture representations of different resolutions; update gate Z t Control the previous hidden state h by element-wise multiplication t-1 Hide the current state The impact of

[0082]

[0083] Final representation The 1×1 convolution of the converted hidden state and the current input is combined, that is:

[0084]

[0085] Step 6: After the features corresponding to each type of signal data are extracted, they need to be fused for RUL prediction. Bilinear pooling can capture the complex interactions between features, but the computational complexity is high. Therefore, a factorized bilinear pooling attention (FBPA) fusion module is designed to reduce the complexity, such as Figure 4 As shown, the FBPA fusion module adopts a pairwise fusion strategy. To simplify the explanation, two signal types are used. and As an example, it should be noted that this fusion can be seamlessly extended to more signal types; the feature tensor and is first vectorized into and in and The bilinear fusion operation computes the pairwise interactions between a and b as follows:

[0086]

[0087] In the formula is the projection weight matrix, || represents the connection operation, and o is the output dimension; directly use the high-dimensional weight matrix To solve the problem of high computational cost of interaction, W i Decomposed into two low-rank matrices and The formula is as follows:

[0088]

[0089] In the formula ρ is the decomposition rank, is a full 1 vector; according to the above derivation, to obtain the output fusion representation z of size o, it is necessary to learn two 3D weight tensors and Then by copying on o channels and To reshape these tensors, we can use the following:

[0090]

[0091] Then the matrix and The Hadamard product of is used to calculate the fused representation z, and then a sum pooling operation is performed on non-overlapping windows of size ρ:

[0092]

[0093] In order to strengthen the interaction between multi-source features and capture common semantics, the above second-order bilinear fusion can be extended to high-order fusion. The fusion process is as follows:

[0094]

[0095] In the formula is the intermediate feature vector before the k-th process summation and pooling operation, Corresponding to the (k-1)th process, k∈{1,…,K} is the index of the stacking process; the final fusion output z is obtained by connecting K processes:

[0096]

[0097] A co-attention mechanism is set up in the FBPA module to learn the joint attention weights for each pair of signal types; and First, use the above FBP algorithm to fuse the original multi-source data features to obtain common features Then, the common features are projected into the multi-head computation space using a 1×1 convolutional layer and a ReLU activation function transformation block; Apply the Softmax layer on the channel dimension as follows:

[0098]

[0099] Where h is the number of heads, i = (0, ..., c-1); in each head, the original multi-source data features are combined with the transformed common features, and then summed over channel c to obtain weighted features and as follows:

[0100]

[0101] right and The fusion output of is finally obtained through the above FBP algorithm, that is, Finally, based on the above pairwise fusion operation, the multi-source fusion output is obtained as follows:

[0102]

[0103] where F1,…,F M is the raw monitoring data of each signal type, M is the number of detected signal types, Indicates the fusion operation using the FBPA module;

[0104] Step 8: Perform regression prediction inference and represent the fusion F z Enter the regression prediction inference module to obtain the RUL prediction value at each time step:

[0105]

[0106] where f r is the regression function, l is the number of FC layers in the module, and the first layer is defined as Each FC layer takes the output of the previous layer as input, with weights and bias The final layer f r It contains only one neuron and uses the Sigmoid function output to monitor the health status of the machine.

[0107] The signal data used in this example comes from a gearbox fatigue life test. During the gear degradation process, three types of signal data were collected: vibration, sound, and current. The signal sampling frequency was 12.8kHz, the sampling interval was 2 minutes, and the sampling duration was 2.56 seconds. The speed and torque of the test gears on the test bench were controlled by a control cabinet. The validation test included 11 gear data under four operating conditions, as shown in Table 1.

[0108] Table 1 Gear vibration signal sample set

[0109]

[0110] Reference Figure 5 ,from Figure 5It can be seen that this embodiment is able to make effective predictions for gears under different working conditions and provide corresponding prediction intervals. Since the degradation process of gears is much slower than that of other machines (such as milling cutters), and the life span ranges from 530 to 4495 cycles, a model is needed that can integrate multi-source data information, fully grasp the health status of the gears, and effectively manage long-term memory storage and transmission. Despite the long life of the gears, this embodiment can still show satisfactory convergence and the ability to fit the actual RUL (real situation). It is worth noting that the predicted values ​​are usually lower than the actual situation. This conservative deviation is intentional to avoid excessive delays in predictions, which leads to an increased risk of unexpected failures. In the initial stage of degradation, due to limited monitoring data, the predictions will fluctuate greatly. However, as the data accumulates, more valuable information can be captured, and the internal parameters can be adjusted to achieve more accurate RUL predictions.

Claims

1. A multi-source parallel bilinear co-attention fusion mechanical equipment life prediction method, characterized by: First, each type of monitoring signal is preprocessed to eliminate outliers. Then, the preprocessed data is input into the subnet for in-depth feature extraction to obtain high-dimensional multi-source specific features. Then, the high-dimensional multi-source specific features are integrated through a multi-source feature fusion module, and the final predicted RUL prediction result is obtained through a regression prediction and reasoning module.

2. A multi-source parallel bilinear co-attention fusion mechanical equipment life prediction method, characterized in that: The following steps are involved: Step 1: Obtain a multi-source dataset of the entire life cycle of mechanical equipment Where N is the number of sequence-tag pairs, M is the number of monitored signal types, are the channels for each signal type, is the number of sampled data points in each measurement for each signal type; multi-source data set for the entire life of mechanical equipment It consists of time-series measurements from multiple sensors, where the sampling time interval is kept constant; Step 2: Apply the Hampel filter to detect and remove outliers in each signal type. The Hampel filter uses the median absolute deviation (MAD) within a sliding window to identify outliers and is applied to the input sequence. The process of each sliding window of size p is as follows: Where Mdn(·) is the median of the input sequence, is the MAD value of the sequence, and λ is the predefined anomaly detection threshold; Step 3: Use parallel computing to process multi-source data streams. Each signal type is extracted by a corresponding sub-network. Each sub-network is composed of a basic unit spatiotemporal representation encoding (STRE) module. Multiple STRE modules are stacked to gradually refine the features. The STRE module achieves two goals: spatial structure mining and temporal dependency capture. Spatial structure mining enhances the model's ability to distinguish potential fault features by analyzing the spatial relationship of signal data. Time-dependent tracking reveals subtle differences between similar signatures in complex scenes by monitoring how signal data changes over time. This is accomplished in each STRE module through a Dynamic Multi-wavelet Aggregate Encoding (DMAE) layer and a Multi-Scale Convolutional Gated Recurrent Unit (MCGRU) layer. Step 4: Integrate the wavelet function as a kernel into a convolutional neural network (CNN) to handle non-stationarity and perform multi-domain analysis; for a small batch input N i The jth output channel C out The calculation is expressed as: Where W represents the weight of the wavelet kernel, b is the bias term, * d represents the dilated convolution with coefficient d, ψ(·) is the wavelet basis function, C in and C out are the input and output channels respectively, X is the input data, m is the input data sequence number, α is the size factor in wavelet transform, β is the time shift factor, and t represents the time step; The wavelet kernel convolution is extended to multi-wavelet kernel aggregation to improve the ability to analyze non-stationary data; multiple wavelet kernels with different basis functions are integrated in one layer, and their outputs are aggregated to enhance the ability to capture discriminant information and give full play to the complementary advantages of different types of wavelets. The calculation process is as follows: In the formula, ω u (X) is the dynamic weight of the u-th aggregation kernel, which is calculated using the attention mechanism conditioned on the input X. U is the total number of aggregation kernels. ⊙ and * represent the Hadamard product and convolution operators, respectively. and Respectively represent the weight and bias of the u-th kernel; The dynamic properties are extended to all dimensions of the kernel space to construct the DMAE layer of the full-dimensional dynamic structure; specifically, for U The kernel of , where g is the number of group connections, the DMAE layer introduces different attention scalars ω = (ω ou ,ω wu ,ω su ,ω cu ) to dynamically adjust the importance of specific cues within the kernel. These scalars are assigned to different dimensions: ω wu Acting on the convolution kernel, are three extended attention scalars calculated on the group input dimension, kernel size dimension, and output dimension respectively. Then the extended operation process is: Step 5: Construct the MCGRU layer, capture temporal features through convolution operations, and add residual connections to learn hierarchical representations; reset gate R t and update gate Z t Control the flow of information, processing input and previous hidden state: In the formula, is the input, h t-1 is the hidden state of the previous step; For the candidate hidden state h t-1 , a hierarchical design is adopted to capture a larger context, and the hierarchical design is expressed as follows: In the formula, ψ t is a convolutional transformation, MSP(·) represents multi-scale pooling, which aims to capture representations of different resolutions; update gate Z t Control the previous hidden state h by element-wise multiplication t-1 Hide the current state The impact of Final characterization The 1×1 convolution combines the converted hidden state with the current input, namely: Step 6: Design a factorized bilinear pooling attention (FBPA) fusion module to reduce complexity. The FBPA fusion module adopts a pairwise fusion strategy for two signal types. and The fusion process, feature tensor and First it is vectorized into and in and The bilinear fusion operation computes the pairwise interactions between a and b as follows: In the formula is the projection weight matrix, || represents the connection operation, and o is the output dimension; W i Decomposed into two low-rank matrices and The formula is as follows: In the formula ρ is the decomposition rank, is a vector of all 1s; according to the above derivation, to obtain an output fusion representation z of size o, two 3D weight tensors need to be learned and Then by copying on o channels and To reshape these tensors, we can use: Then the matrix and The Hadamard product of is used to calculate the fused representation z, and then a sum pooling operation is performed on non-overlapping windows of size ρ: The above second-order bilinear fusion is extended to high-order fusion, and the fusion process is as follows: In the formula is the intermediate feature vector before the k-th process summation and pooling operation, Corresponding to the (k-1)th process, k∈{1,…,K} is the index of the stacking process; the final fusion output z is obtained by connecting K processes: Step 7: Set up the co-attention mechanism in the FBPA module to learn the joint attention weights for each pair of signal types; and First, use the above FBP algorithm to fuse the original multi-source data features to obtain common features Then, the common features are projected into the multi-head computation space using a 1×1 convolutional layer and a transformation block with ReLU activation function. Apply the Softmax layer on the channel dimension of as follows: Where h is the number of heads, i = (0, ..., c-1); in each head, the original multi-source data features are combined with the transformed common features, and then summed over channel c to obtain weighted features and as follows: right and The fusion output of is finally obtained through the above FBP algorithm, that is, Finally, based on the above pairwise fusion operation, the multi-source fusion output is obtained as follows: where F1,…,F M is the raw monitoring data of each signal type, M is the number of detected signal types, Indicates the fusion operation using the FBPA module; Step 8: Perform regression prediction inference and represent the fusion F z Enter the regression prediction inference module to get the RUL prediction value at each time step: where f r is the regression function, l is the number of FC layers in the module, and the first layer is defined as Each FC layer takes the output of the previous layer as input, with weights and bias The final layer f r It contains only one neuron and uses the Sigmoid function output to monitor the health status of the machine.

3. The method according to claim 2, characterized in that: The signal type fusion process in step 6 can be extended to more signal types.

Images