Device remaining service life prediction method based on dual-path architecture, program, device and storage medium

Abstract

The invention belongs to the technical field of aero-turbofan engine residual life prediction, and particularly relates to an equipment residual service life prediction method based on a double-path architecture, a program, equipment and a storage medium. According to the method, a turbofan engine residual life prediction model based on a dual-path interaction framework fusing multi-scale time-feature convolution and a dynamic weight adaptive module is constructed, a complex relationship between a time dimension and a feature dimension is captured through a dual-path structure, and then a plurality of convolution kernels of different scales are introduced through 1DCNN. According to the method, multi-scale feature extraction is performed, feature representation of time and feature dimensions with higher expressive power is generated through splicing and segmentation operations, and finally, the prediction performance and flexibility of the model are enhanced through dynamic interaction and adaptive weight adjustment between the time and feature dimensions. According to the method, multi-scale information extraction in time and feature dimensions is realized, more information interaction is reserved, and the accuracy of RUL prediction is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of remaining service life prediction of aviation turbofan engines, and specifically relates to a method, program, device and storage medium for predicting remaining service life of equipment based on a dual-path architecture. Background Art

[0002] During the long-term high-load operation of aircraft engines, the degradation of key components is serious, which can easily cause equipment failures and lead to huge economic losses. Traditional maintenance strategies, such as corrective maintenance, time-based maintenance, and condition-based maintenance, can no longer meet existing needs. By utilizing the remaining useful life (RUL) prediction, one of the core technologies in prognostics and health management (PHM), combined with big data, the Internet of Things and other technologies, equipment-related data can be obtained and analyzed in real time, and the equipment status can be evaluated, thus providing a basis for optimizing maintenance strategies and reducing operating costs.

[0003] With the deployment of sensors and the improvement of computing power, deep learning can use neural networks to automatically mine deeper features in historical data to achieve more accurate RUL prediction. At present, the most widely used models are recurrent neural networks (RNN) and convolutional neural networks (CNN). RNN can discover temporal features, and CNN can discover different scale features. In addition, in recent years, MLP (multi-layer perceptron) has great potential in time series prediction. Through its powerful representation learning ability, it can effectively capture complex time dependencies and patterns. For example, the MLP-Mixer architecture proposed by Tolstikhin et al. in the paper "Mlp-mixer: An all-mlp architecture for vision" was originally designed for image processing. However, in the RUL prediction task, MLP-Mixer can capture the timing information of device degradation by mixing local features at each time step and mixing global features across time steps. This method does not rely on convolution or attention mechanisms and provides a new way to process time series data. Although these studies have improved the accuracy of the model in RUL prediction, there are still some limitations. First, these methods mostly focus on feature extraction in the temporal dimension, ignoring the diversity of feature dimensions. Second, these studies mostly focus on the extraction of local features and specific scales, lacking in-depth exploration of the correlation between cross-scale features. Summary of the invention

[0004] The object of the present invention is to provide a method, program, device and storage medium for predicting the remaining useful life of equipment based on a dual-path architecture.

[0005] A method for predicting the remaining useful life of equipment based on a dual-path architecture comprises the following steps:

[0006] Step 1: Construct a training set, where each sample in the training set includes sensor data of a turbofan engine and its remaining life;

[0007] Step 2: training a prediction model using a training set; the prediction model includes a dual-path interaction network, a multi-scale time-feature convolution fusion module, and a dynamic adaptive weight module;

[0008] The dual-path interactive network architecture sets up a dual parallel path structure to extract features in two dimensions: time and feature. The input turbofan engine sensor data is passed through a parameter-sharing fully connected layer to generate two identical features x1 and x2. Features x1 and x2 T Nonlinear transformation is introduced through activation function, followed by linear transformation through fully connected layer, and finally processed through layer normalization to obtain preliminary extracted features y1 in time dimension and preliminary extracted features y2 in feature dimension; x1 and y1 are dynamically weighted fused and then added element by element with y2 for feature fusion, and the fused result is processed through layer normalization to obtain stable feature output f1 in time dimension; x2 and y2 are dynamically weighted fused and then added element by element with y1 for feature fusion, and the fused result is processed through layer normalization to obtain stable feature output f2 in time dimension;

[0009] The multi-scale time-feature convolution fusion module processes f1 and f2 in parallel to generate an output feature f containing multi-scale time information. output1 and the output feature f containing multi-scale feature information output2 ;

[0010] The dynamic adaptive weight module is used to output1 and f output2 After dynamic weighted fusion, the fusion feature f is obtained combine ; f combine The vector is converted into a one-dimensional vector through the Flatten operation and then passes through the linear transformation layer to generate the final prediction value f final , i.e. the predicted value of the remaining life of the turbofan engine;

[0011] Step 3: Obtain the sensor data of the turbofan engine to be predicted, input it into the trained prediction model, and obtain the remaining life prediction result of the turbofan engine.

[0012] Furthermore, when constructing the training set in step 1, the sensor data of the turbofan engine is first collected, the available sensor data is selected from the collected data, and all sensor values ​​are normalized to [0, 1] using Min-Max normalization; a fixed time window size and sliding step size are set, and a new data sample is obtained each time a slide is made in the time dimension of the data; the entire life cycle of the engine is divided into a healthy stage and a degraded stage, and the RUL in the healthy stage maintains a relatively high constant value RUL max , and then the RUL label is standardized. The specific formula is:

[0013]

[0014] Among them, RUL norm is the standardized RUL label, with a value range of [0,1]; RUL origin is the original URL label. Further, the specific data processing process in the dual-path interactive network architecture is expressed as:

[0015] x1=W×input+b

[0016] x2=x1

[0017] y1=LayerNorm(W1×GELU(x1))+b1

[0018] y2=LayerNorm(W2×GELU(x2 T ))+b2

[0019] f1=LayerNorm(α1x1+β1y1+y2)

[0020] f2=LayerNorm(α2x2+β2y2+y1)

[0021] Among them, input is the sensor data of the input turbofan engine; W is the shared weight matrix; b is the shared bias parameter; W1, W2, b1, b2 are the weight matrix and bias parameter of the time dimension and the weight matrix and bias parameter of the feature dimension respectively; LayerNorm(·) is layer normalization; GELU(·) is the GELU activation function; α1, β1, α2, β2 are the learnable parameters for the time dimension and feature dimension in the dynamic adaptive weight module.

[0022] Furthermore, the data processing method for f1 and f2 in the multi-scale time-feature convolution fusion module is the same.

[0023] Furthermore, the data processing method for f1 in the multi-scale time-feature convolution fusion module is specifically as follows:

[0024] First, 1DCNN is used to process f1, and it is parallelized by 1x1 convolution to extract local detail features, 3x3 convolution to extract medium-scale features, and 5x5 convolution to extract global features. Then, the output features f of these three different scale convolutions are combined. conv1 、f conv2 、f conv3 To integrate, the pooling operation is used to reduce the feature dimension to 1. The specific data processing process is expressed as:

[0025] f conv1 =Conv1D(f1,w1,b3)

[0026] f conv2 =Conv1D(f1,w2,b4)

[0027] f conv3 =Conv1D(f1,w3,b5)

[0028] f concat =Concat(f conv1 ,f conv2 ,f conv3 )

[0029] f pool =AdaptiveAvgpool1D(f concat ; 1)

[0030] Among them, Conv1D(·) is a one-dimensional convolution layer, Concat(·) is a Concat integration layer, AdaptiveAvgpool1D(·) is a one-dimensional adaptive average pooling layer; w1, w2, w3 represent the convolution kernels of three different scales of 1x1, 3x3, and 5x5 respectively; b3, b4, b5 represent the bias vectors of three different scales of 1x1, 3x3, and 5x5 respectively;

[0031] The pooled feature f pool Through the first fully connected layer, and through the ReLU activation function for nonlinear mapping, the features are converted to the new feature space; then, through the second fully connected layer, the time dimension features are expanded to 3 times the window size, and then the Split operation is performed to re-split the comprehensive features containing multi-scale information into time dimension features corresponding to the three original convolution scales of 1x1, 3x3, and 5x5; then, the Sigmoid activation function is applied to the segmented features to generate feature weights of the corresponding scales, and multiplied with the corresponding original convolution output to complete the feature weighting operation. Finally, the weighted features are added to generate the final output feature f containing multi-scale time information. output1 , the specific data processing process is expressed as:

[0032] fsplit1 ,f split2 ,f split3 =Split(W 2d (ReLU(W 1d f pool +b 1d ))+b 2d )

[0033] f weight1 =f conv1 ·Sigmod(f split1 )

[0034] f weight2 =f conv2 ·Sigmod(f split2 )

[0035] f weight3 =f conv3 ·Sigmod(f split3 )

[0036] f output1 =f weight1 +f weight2 +f weight3

[0037] Among them, W 1d and b 1d are the weight matrix and bias parameters of the first fully connected layer respectively; W 2d and b 2d are the weight matrix and bias parameters of the second fully connected layer respectively; ReLU(·) is the ReLU activation function; Sigmod(·) is the Sigmoid activation function.

[0038] Furthermore, the specific data processing process in the dynamic adaptive weight module is expressed as:

[0039] f combine =αf output1 +βf output2

[0040] f final =W0(Flatten(f combine )+b0)

[0041] Among them, α and β are dynamic weight variables; W0 and b0 are the weight matrix and bias parameters of the linear layer; Flatten(·) is the Flatten layer.

[0042] A computer device / equipment / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-mentioned method for predicting the remaining useful life of equipment based on a dual-path architecture.

[0043] A computer-readable storage medium stores a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned method for predicting the remaining useful life of equipment based on a dual-path architecture.

[0044] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned method for predicting the remaining useful life of equipment based on a dual-path architecture.

[0045] The beneficial effects of the present invention are:

[0046] The present invention constructs a turbofan engine remaining life prediction model based on a dual-path interaction framework that integrates multi-scale time-feature convolution and dynamic weight adaptive modules. The dual-path structure captures the complex relationship between the time dimension and the feature dimension, and then introduces multiple convolution kernels of different scales through 1DCNN to perform multi-scale feature extraction, and generates more expressive feature representations of the time and feature dimensions through splicing and segmentation operations. Finally, the dynamic interaction and adaptive weight adjustment between the time and feature dimensions enhance the prediction performance and flexibility of the model. The present invention realizes multi-scale information extraction in the time and feature dimensions and retains more information interaction, which significantly improves the accuracy of RUL prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall flow chart of the present invention.

[0048] Figure 2 The effect diagram of standardizing URL labels.

[0049] Figure 3 This is the overall architecture diagram of the prediction model in the present invention.

[0050] Figure 4 It is a structural diagram of the dual-path interaction network in the present invention.

[0051] Figure 5 This is a structural diagram of the multi-scale time-feature convolution fusion module in the present invention.

[0052] Figure 6 It is a structural diagram of the dynamic adaptive module in the present invention.

[0053] Figure 7 It is a diagram showing the effect of the present invention on RUL prediction on the C-MAPSS data set. DETAILED DESCRIPTION

[0054] The present invention is further described below in conjunction with the accompanying drawings.

[0055] The present invention provides a method for predicting the remaining useful life of equipment based on a dual-path architecture. The dual-path interactive framework based on the fusion of multi-scale time-feature convolution and dynamic weight adaptive modules can capture the complex relationship between the time dimension and the feature dimension, enhance the information interaction in the feature extraction process, thereby improving the quality of feature representation and the accuracy of prediction. Multi-scale feature extraction and adaptive weight adjustment are performed through the multi-scale time-feature convolution fusion module and the dynamic weight adaptive module, thereby enhancing the prediction performance and flexibility of the model, thereby significantly improving the accuracy of RUL prediction.

[0056] First, the sensor data of the turbofan engine is collected, and the available sensor data is selected from the collected data and all sensor values ​​are normalized to [0,1] using Min-Max normalization; a fixed time window size and sliding step size are set, and a new data sample is obtained each time a slide is made in the time dimension of the data; the entire life cycle of the engine is divided into a healthy stage and a degraded stage, and the RUL in the healthy stage maintains a higher constant value RUL max , and then the URL label is standardized. The specific formula is as follows:

[0057]

[0058] Among them, RUL norm is the standardized RUL label, with a value range of [0,1]; RUL origin It is the original URL tag.

[0059] Build a turbofan engine remaining life prediction model based on a dual-path interaction framework that integrates multi-scale time-feature convolution and dynamic weight adaptive modules, such as Figure 3 As shown in the figure, it includes a dual-path interaction network architecture (DPIN), a multi-scale temporal-feature convolution fusion module (MTF-CFM) and a dynamic adaptive weight module (DWAM). The dual-path interaction network is responsible for the realization of feature extraction and interaction process, the multi-scale temporal-feature convolution fusion module is used to capture and integrate multi-scale spatiotemporal information, and the dynamic adaptive weight module improves the accuracy of RUL prediction by dynamically adjusting weights.

[0060] like Figure 4As shown in the figure, the dual-path interactive network architecture sets up a dual parallel path structure to extract features in two dimensions: time and features. The input sensor data of the turbofan engine generates two identical features x1 and x2 after passing through the parameter-sharing fully connected layer. The features x1 and x2 T The nonlinear transformation is introduced through the GELU activation function, followed by linear transformation through the fully connected layer, and finally processed through layer normalization to obtain the preliminary extracted features y1 in the time dimension and y2 in the feature dimension;

[0061] x1=W×input+b

[0062] x2=x1

[0063] y1=LayerNorm(W1×GELU(x1))+b1

[0064] y2=LayerNorm(W2×GELU(x2 T ))+b2

[0065] Among them, input is the sensor data of the input turbofan engine; W is the shared weight matrix; b is the shared bias parameter; W1, W2, b1, b2 are the weight matrix and bias parameter of the time dimension and the weight matrix and bias parameter of the feature dimension respectively; LayerNorm(·) is layer normalization; GELU(·) is the GELU activation function; α1, β1, α2, β2 are the learnable parameters for the time dimension and feature dimension in the dynamic adaptive weight module.

[0066] After performing dynamic weighted fusion on x1 and y1, the feature fusion is performed by adding them to y2 element by element. The fusion result is processed by layer normalization to obtain a stable time dimension feature output f1. After performing dynamic weighted fusion on x2 and y2, the feature fusion is performed by adding them to y1 element by element. The fusion result is processed by layer normalization to obtain a stable time dimension feature output f2.

[0067] f1=LayerNorm(α1x1+β1y1+y2)

[0068] f2=LayerNorm(α2x2+β2y2+y1)

[0069] Among them, α1, β1, α2, β2 are learnable parameters for the time dimension and feature dimension in the dynamic adaptive weight module.

[0070] like Figure 5As shown in the figure, the multi-scale time-feature convolution fusion module processes the obtained time dimension output f1 and feature dimension output f2 in parallel. Taking the time dimension feature as an example, f1 is first processed by 1DCNN, and then it is parallelly processed through 1x1 convolution to extract local detail features, 3x3 convolution to extract medium-scale features, and 5x5 convolution to extract global features. Subsequently, the output features f of these three different scale convolutions are combined. conv1 、f conv2 、f conv3 To integrate, the pooling operation is used to reduce the feature dimension to 1. The specific data processing process is expressed as:

[0071] f conv1 =Conv1D(f1,w1,b3)

[0072] f conv2 =Conv1D(f1,w2,b4)

[0073] f conv3 =Conv1D(f1,w3,b5)

[0074] f concat =Concat(f conv1 ,f conv2 ,f conv3 )

[0075] f pool =AdaptiveAvgpool1D(f concat ; 1)

[0076] Among them, Conv1D(·) is a one-dimensional convolution layer, Concat(·) is a Concat integration layer, AdaptiveAvgpool1D(·) is a one-dimensional adaptive average pooling layer; w1, w2, w3 represent the convolution kernels of three different scales of 1x1, 3x3, and 5x5 respectively; b3, b4, b5 represent the bias vectors of three different scales of 1x1, 3x3, and 5x5 respectively;

[0077] The pooled feature f pool Through the first fully connected layer, and through the ReLU activation function for nonlinear mapping, the features are converted to the new feature space; then, through the second fully connected layer, the time dimension features are expanded to 3 times the window size (3w), and then the Split operation is performed to re-split the comprehensive features containing multi-scale information into time dimension features corresponding to the three original convolution scales of 1x1, 3x3, and 5x5; then, the Sigmoid activation function is applied to the segmented features to generate feature weights of the corresponding scales, and multiplied with the corresponding original convolution output to complete the feature weighting operation. Finally, the weighted features are added to generate the final output feature f containing multi-scale time information.output1 , the specific data processing process is expressed as:

[0078] f split1 ,f split2 ,f split3 =Split(W 2d (ReLU(W 1d f pool +b 1d ))+b 2d )

[0079] f weight1 =f conv1 ·Sigmod(f split1 )

[0080] f weight2 =f conv2 ·Sigmod(f split2 )

[0081] f weight3 =f conv3 ·Sigmod(f split3 )

[0082] f output1 =f weight1 +f weight2 +f weight3

[0083] Among them, W 1d and b 1d are the weight matrix and bias parameters of the first fully connected layer respectively; W 2d and b 2d are the weight matrix and bias parameters of the second fully connected layer respectively; ReLU(·) is the ReLU activation function; Sigmod(·) is the Sigmoid activation function.

[0084] The multi-scale time-feature convolution fusion module is used to perform the same processing on the acquired feature dimension output f2 to generate the final output f containing multi-scale feature information. output2 .

[0085] like Figure 6 As shown in the figure, the dynamic adaptive weight module outputs the multi-scale temporal information feature f output1 And the output of multi-scale feature information f output2 After dynamic weighted fusion, the fused feature f is obtained combine

[0086] f combine =αf output1 +βf output2

[0087] Among them, α and β are dynamic weight variables;

[0088] The fused feature f combine The vector is converted into a one-dimensional vector through the Flatten operation and then passes through the linear transformation layer to generate the final RUL prediction value f final , i.e. the predicted value of the remaining life of the turbofan engine;

[0089] f final =W0(Flatten(f combine )+b0)

[0090] Among them, W0 and b0 are the weight matrix and bias parameter of the linear layer; Flatten(·) is the Flatten layer.

[0091] Finally, the sensor data of the turbofan engine to be predicted is input into the trained turbofan engine remaining life prediction model, and the predicted value of the remaining life is output.

[0092] Embodiment 1:

[0093] This embodiment uses the C-MAPSS dataset, which is widely used in the PHM field, to conduct experiments. The C-MAPSS dataset is a simulated engine degradation dataset specifically used for health management research of aircraft engines.

[0094] From the collected data, 14 available sensor data, 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20 and 21, are selected to construct a data set and all sensor values ​​are normalized to [0, 1] using Min-Max normalization; the time window size is set to 40 and the sliding step is set to 1. In the time dimension of the data, a new data sample is obtained each time the sliding is performed; the entire life cycle of the engine is divided into a healthy stage and a degraded stage. In the healthy stage, RUL is kept at a constant value RUL max =125, then the URL label is standardized. The standardized effect is as follows Figure 2 As shown, the specific formula is as follows:

[0095] The hyperparameter settings for training the turbofan engine remaining life prediction model are shown in Table 1.

[0096] Table 1 Experimental hyperparameter settings

[0097]

[0098] The preprocessed test set data is used to evaluate the remaining life prediction model of the turbofan engine. The evaluation indicators include RMSE, MAPE and Score. The specific formula is as follows:

[0099]

[0100]

[0101] in, With y i They represent the estimated RUL and true RUL of the i-th test sample respectively, and N represents the total number of test samples.

[0102] In order to verify the effectiveness and superiority of the constructed model, the predicted value of the remaining life is compared with the relevant advanced research in this field. The output predicted value of the remaining life is shown in Table 2. -M means adding the multi-scale time-feature convolution fusion module (MTF-CFM) to verify the effectiveness of the module.

[0103] Table 2 Comparative experimental results

[0104]

[0105] From the data, it can be seen that the performance of the model proposed in the present invention on the four data sets is better than that of the current advanced methods, and the best performance is achieved in FD002 and FD004 with complex working conditions, indicating that the model of the present invention can more accurately predict RUL when dealing with complex working conditions. At the same time, the multi-scale time-feature convolution fusion module proposed in the present invention can stably improve the RUL prediction performance of the model in almost all cases, and has good applicability and application potential.

[0106] In addition, in order to more intuitively demonstrate the performance of the algorithm proposed in this invention on the C-MAPSS dataset, the real RUL and predicted RUL of each engine are visualized and analyzed. The prediction results of some engines are shown in the figure below. Figure 7 shown.

[0107] In summary, the present invention captures the complex relationship between the time dimension and the feature dimension through a dual-path structure, and then introduces multiple convolution kernels of different scales through a 1D convolutional neural network (1DCNN) to perform multi-scale feature extraction, and generates more expressive feature representations of the time and feature dimensions through splicing and segmentation operations. Finally, through the dynamic interaction and adaptive weight adjustment between the time and feature dimensions, the prediction performance and flexibility of the model are enhanced, thereby significantly improving the accuracy of RUL prediction.

[0108] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the remaining useful life of equipment based on a dual-path architecture, characterized in that: The following steps are involved: Step 1: Construct a training set, where each sample in the training set includes sensor data of a turbofan engine and its remaining life; Step 2: training a prediction model using a training set; the prediction model includes a dual-path interaction network, a multi-scale time-feature convolution fusion module, and a dynamic adaptive weight module; The dual-path interactive network architecture sets up a dual parallel path structure to extract features in two dimensions: time and feature. The input sensor data of the turbofan engine generates two identical features x1 and x2 after passing through the parameter-sharing fully connected layer. The features x1 and x2 T Nonlinear transformation is introduced through activation function, followed by linear transformation through fully connected layer, and finally processed through layer normalization to obtain preliminary extracted features y1 in time dimension and preliminary extracted features y2 in feature dimension; x1 and y1 are dynamically weighted fused and then added element by element with y2 for feature fusion, and the fused result is processed through layer normalization to obtain stable feature output f1 in time dimension; x2 and y2 are dynamically weighted fused and then added element by element with y1 for feature fusion, and the fused result is processed through layer normalization to obtain stable feature output f2 in time dimension; The multi-scale time-feature convolution fusion module processes f1 and f2 in parallel to generate an output feature f containing multi-scale time information. output1 and the output feature f containing multi-scale feature information output2 ; The dynamic adaptive weight module is used to output1 and f output2 After dynamic weighted fusion, the fusion feature f is obtained combine ; f combine The vector is converted into a one-dimensional vector through the Flatten operation and then passes through the linear transformation layer to generate the final prediction value f final , i.e. the predicted value of the remaining life of the turbofan engine; Step 3: Obtain the sensor data of the turbofan engine to be predicted, input it into the trained prediction model, and obtain the remaining life prediction result of the turbofan engine.

2. The method for predicting the remaining useful life of equipment based on a dual-path architecture according to claim 1, characterized in that: When constructing the training set in step 1, first collect the sensor data of the turbofan engine, select the available sensor data from the collected data, and use Min-Max normalization to normalize all sensor values ​​to between [0,1]; set a fixed time window size and sliding step size, and obtain a new data sample each time sliding in the time dimension of the data; divide the entire life cycle of the engine into a healthy stage and a degraded stage, and the RUL in the healthy stage maintains a higher constant value RUL max , and then the RUL label is standardized. The specific formula is: Among them, RUL norm is the standardized RUL label, with a value range of [0,1]; RUL origin It is the original URL tag.

3. The method for predicting the remaining useful life of equipment based on a dual-path architecture according to claim 1, characterized in that: The specific data processing process in the dual-path interactive network architecture is expressed as follows: x1=W×input+b x2=x1 y1=LayerNorm(W1×GELU(x1))+b1 <h2 style=";text-align:left;direction:ltr">y2 = LayerNorm(W2×GELU(x2<h2 style=";text-align:left;direction:ltr"> T <h2 style=";text-align:left;direction:ltr"> ))+b2 f1=LayerNorm(α1x1+β1y1+y2) f2=LayerNorm(α2x2+β2y2+y1) Among them, input is the sensor data of the input turbofan engine; W is the shared weight matrix; b is the shared bias parameter; W1, W2, b1, b2 are the weight matrix and bias parameter of the time dimension and the weight matrix and bias parameter of the feature dimension respectively; LayerNorm(·) is layer normalization; GELU(·) is the GELU activation function; α1, β1, α2, β2 are the learnable parameters for the time dimension and feature dimension in the dynamic adaptive weight module.

4. The method for predicting the remaining useful life of equipment based on a dual-path architecture according to claim 3, characterized in that: The data processing method for f1 and f2 in the multi-scale time-feature convolution fusion module is the same.

5. The method for predicting the remaining useful life of equipment based on a dual-path architecture according to claim 4, characterized in that: The data processing method for f1 in the multi-scale time-feature convolution fusion module is specifically as follows: First, 1DCNN is used to process f1, and it is parallelized by 1x1 convolution to extract local detail features, 3x3 convolution to extract medium-scale features, and 5x5 convolution to extract global features. Then, the output features f of these three different scale convolutions are combined. conv1 、f conv2 、f conv3 To integrate, the pooling operation is used to reduce the feature dimension to 1. The specific data processing process is expressed as: f conv1 =Conv1D(f1,w1,b3) f conv2 =Conv1D(f1,w2,b4) <h2 style=";text-align:left;direction:ltr">f<h2 style=";text-align:left;direction:ltr"> conv3 <h2 style=";text-align:left;direction:ltr"> =Conv1D(f1,w3,b5) f concat =Concat(f conv1 ,f conv2 ,f conv3 ) f pool =AdaptiveAvgpool1D(f concat ;1) Among them, Conv1D(·) is a one-dimensional convolution layer, Concat(·) is a Concat integration layer, AdaptiveAvgpool1D(·) is a one-dimensional adaptive average pooling layer; w1, w2, w3 represent the convolution kernels of three different scales of 1x1, 3x3, and 5x5 respectively; b3, b4, b5 represent the bias vectors of three different scales of 1x1, 3x3, and 5x5 respectively; The pooled feature f pool Through the first fully connected layer, and through the ReLU activation function for nonlinear mapping, the features are converted to the new feature space; then, through the second fully connected layer, the time dimension features are expanded to 3 times the window size, and then the Split operation is performed to re-split the comprehensive features containing multi-scale information into time dimension features corresponding to the three original convolution scales of 1x1, 3x3, and 5x5; then, the Sigmoid activation function is applied to the segmented features to generate feature weights of the corresponding scales, and multiplied with the corresponding original convolution output to complete the feature weighting operation. Finally, the weighted features are added to generate the final output feature f containing multi-scale time information. output1 , the specific data processing process is expressed as: f split1 ,f split2 ,f split3 =Split(W 2d (ReLU(W 1d f pool +b 1d ))+b 2d ) f weight1 =f conv1 ·Sigmod(f split1 ) f weight2 =f conv2 ·Sigmod(f split2 ) f weight3 =f conv3 ·Sigmod(f split3 ) f output1 =f weight1 +f weight2 +f weight3 Among them, W 1d and b 1d are the weight matrix and bias parameters of the first fully connected layer respectively; W 2d and b 2d are the weight matrix and bias parameters of the second fully connected layer respectively; ReLU(·) is the ReLU activation function; Sigmod(·) is the Sigmoid activation function.

6. The method for predicting the remaining useful life of equipment based on a dual-path architecture according to claim 5, characterized in that: The specific data processing process in the dynamic adaptive weight module is expressed as follows: f combine =αf output1 +βf output2 f final =W0(Flatten(f combine )+b0) Among them, α and β are dynamic weight variables; W0 and b0 are the weight matrix and bias parameters of the linear layer; Flatten(·) is the Flatten layer.

7. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images