Remaining service life prediction method based on double-attention multi-feature fusion network

Abstract

The invention provides a residual service life prediction method based on a double-attention multi-feature fusion network, and the method comprises the steps: S1, constructing a data set for collected equipment multi-sensor data, carrying out the screening, and carrying out the normalization processing of the screened data; s2, on the basis of the training set sample, constructing a multi-feature fusion network model based on a double-attention mechanism; s3, inputting a test sample into the trained double-attention-based multi-feature fusion network model to realize the prediction of the residual service life of the equipment; according to the method, manual features are fused, a multi-feature fusion framework is constructed, final RUL prediction is realized, and the accuracy of predicting the residual service life of the equipment is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment failure prediction and health management, and in particular relates to a remaining service life prediction method based on a dual-attention multi-feature fusion network. Background Art

[0002] With the rapid development of technologies such as the Industrial Internet, the Internet of Things, and 5G, equipment is becoming larger, faster, more integrated, and more intelligent. These devices quickly accumulate a large amount of data during operation, providing new development opportunities for the advancement of equipment status prediction and health management technology. As a core research content in the PHM system, remaining service life prediction plays a key role in complex equipment (such as aircraft engines and high-speed trains) and their components.

[0003] Existing remaining life prediction methods are mainly divided into physical model-based methods and data-driven methods. The physical model-based prediction method predicts the remaining life of the equipment by modeling the physical behavior, wear mechanism and failure mode of the equipment. This type of method relies on the physical principles and operating conditions of the equipment, and usually simulates the operating state and failure process of the equipment through mathematical models or simulation models. Commonly used physical model-based remaining service life prediction methods include Gamma process-based methods and Wiener process-based methods. However, with the rapid development of sensor technology and big data, the complexity and interconnectivity of mechanical systems are increasing, making the physical model-based remaining life prediction methods gradually show their limitations. These methods face challenges in processing complex and multidimensional data, mainly because they rely too much on accurate modeling of the physical behavior and failure mode of the equipment. Therefore, data-driven methods have gradually become a hot topic of research. Data-driven methods extract features from equipment operation data and use machine learning algorithms to establish a mapping relationship between equipment degradation process and remaining life. In contrast, data-driven methods mainly rely on extracting features from equipment operation data and using machine learning and other methods to establish prediction models. Data-driven remaining useful life prediction methods include artificial neural networks, extreme learning machines (ELM), support vector machines (SVM), support vector regression (SVR) models, radial basis function neural networks (RBF) and random forests (RF). These methods have achieved good results in the application of RUL prediction. At present, deep learning, as an end-to-end learning method, has received widespread attention in the field of RUL prediction. Unlike traditional machine learning methods, it no longer requires manual feature selection, but can automatically learn representative features from raw sensory data, which is more convenient and effective in application.

[0004] At present, deep learning, as an end-to-end learning method, has received widespread attention in the field of remaining useful life (RUL) prediction. Different from traditional machine learning methods, it does not require manual feature selection, but can automatically learn representative features from raw sensory data, which is more convenient and effective in application. Among these deep learning-based methods, the long short-term memory network LSTM (Long Short-Term Memory) and the convolutional neural network CNN (Convolutional Neural Network) are two core deep learning models that have been widely used in RUL prediction.

[0005] Traditional RUL prediction methods based on LSTM and CNN still have some limitations: the CNN network has excellent feature extraction capabilities, but it does not weight different features, which leads to no difference in the importance of the extracted features. The LSTM-based RUL prediction method is a neural network that analyzes previous data in a sequential manner, which often leads to long-term dependency problems. That is, only the learning features of the last time step are regressed or classified, while the contribution of features extracted at other times is ignored. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a remaining useful life prediction method based on a dual-attention multi-feature fusion network for the above-mentioned problems, combine the multi-feature fusion model of CNN-CAM and LSTM-SAM, extract features from the time and space dimensions, and comprehensively capture feature categories to achieve the purpose of accurately predicting the remaining useful life.

[0007] The technical solution adopted by the present invention to solve the above technical problems is: a remaining useful life prediction method based on a dual-attention multi-feature fusion network, characterized in that it includes the following steps:

[0008] Step 1: Construct a data set from the collected multi-sensor data of the equipment, merge and filter it, and normalize the filtered data;

[0009] Step 2: Based on the training set samples, a multi-feature fusion network model based on the dual attention mechanism is constructed. First, CNN is combined with CAM to preliminarily extract features in the spatial dimension; then, LSTM is combined with SAM to extract features in the temporal dimension, completing the integration of feature information in the spatial and temporal dimensions.

[0010] Step three, input the test samples into the trained dual-attention multi-feature fusion network model to predict the remaining service life of the equipment.

[0011] According to the above scheme, the data eliminated in the sensor data screening described in step 1 are: the data in the degradation experiment are stable, do not show degradation characteristics, and cannot provide obvious degradation information.

[0012] According to the above scheme, the data normalization process described in step 1 includes the following contents:

[0013] The Min-Max method is used to normalize the data. The specific formula is as follows:

[0014]

[0015] Where: x i Indicates the value at the current time point; and They refer to the maximum and minimum values ​​of the current sensor at all time points; x i ' represents the data obtained after normalization calculation.

[0016] According to the above scheme, the construction of the CNN-CAM module combining CNN and CAM described in step 2 includes the following contents:

[0017] The original time series data is formed into input data after data preprocessing. The input data features can be expressed as

[0018] Shown as:

[0019] X=[x1,x2,x3,…,x k ],where X∈R T×K

[0020] Where: x k represents the Kth channel in the feature map X; T and K represent the time step and the amount of sensor data respectively;

[0021] The convolution calculation is expressed as follows:

[0022]

[0023] Where: X i-1 represents the input of the i-th convolutional layer; W i Represents the i-th convolution kernel; and b i Represent the activation function and the bias term of the function respectively;

[0024] CAM is used to process the data output by the convolution layer. After squeezing the feature map X output by the convolution layer, the spatial information contained in the data is integrated through the global average pooling layer and the global maximum pooling layer. Given that the spatial dimension of the feature map is h×w×c, it becomes 1×1×c after squeezing. After squeezing, the feature map X of each channel will generate the corresponding channel descriptor P K , the formula is as follows:

[0025]

[0026] Where: x k (i,j) represents the element in the i-th row and j-th column in the feature map X;

[0027] P K Perform the excitation operation to generate the extracted descriptor L, which is expressed as:

[0028] L=E(P,FC)=tanh(FC2tanh(FC1P))

[0029] Where: FC1 represents a dimensionality reduction layer; FC2 represents a dimensionality increase layer; tanh is the activation function;

[0030] To avoid the problem of central symmetry, the tanh function sets the output weight between [-1, 1], and then adopts a weighted processing method to complete the recalibration of the original features in the channel dimension. Finally, in order to integrate the data on the two branches of the global average pooling layer and the global maximum pooling layer, the element summation method is used to generate the output feature map, and the softmax function is used to normalize the weights. The normalized weights are assigned to the features of each channel through scaling operations, and the final output of the module is directly input to the next network layer.

[0031] According to the above scheme, the construction of the LSTM-SAM module combining LSTM and SAM described in step 2 includes the following contents:

[0032] Before data flows into the LSTM network, it first passes through the forget gate, whose main function is to selectively forget the data transmitted at the previous moment. Its update calculation formula is:

[0033] f t =σ(W f [h t-1 ,x t ]+b f )

[0034] Where: x t is the input at time step t; h t-1 is the hidden state at time step t-1; W f represents the weight matrix; bf represents the bias term; σ represents the sigmoid function;

[0035] After being filtered by the forget gate, the data is passed to the input gate. The LSTM network input signal is determined by the input gate to determine the information and content to be updated, as shown below:

[0036] i t =σ(W i [h t-1 ,x t ]+b i )

[0037]

[0038] Where: W i , W C represents the weight matrix; b i , b C represents the bias term; Indicates that x t and h t-1 The state quantity that integrates the state information of t Indicates the current status;

[0039] The output gate uses the cell state that needs to be output. The input signal determines the useful information for output through the output gate and immediately updates the current LSTM neural network cell state. The update calculation formula is as follows:

[0040] o t =σ(W o [h t-1 ,x t ]+b o )

[0041] h t =o t *tanh(C t )

[0042] Where: W o represents the weight matrix; b o represents the bias term;

[0043] The learning features output by the LSTM network can be expressed as X = {X1, X2, ..., X n} T , where T represents the transpose operation, and as the input x of SAM i ={x1,x2,…,x m} T , m represents the time step of the data, and the specific formula is as follows:

[0044] u i=softmax(W T x i +b)

[0045]

[0046] Where: u i represents the attention weight; W T and b are the weight matrices and bias terms in the attention network, a i is the influence of the ith eigenvalue on the target sequence data. All a i Add up to 1;

[0047] The feature sequence is weighted to obtain the output sequence, which is expressed as:

[0048]

[0049] The beneficial effects of the present invention are as follows: a remaining useful life prediction method based on a dual-attention multi-feature fusion network is provided, and a dual-attention mechanism framework is adopted. CNN-CAM is responsible for assigning different weights to channels in the spatial dimension, thereby extracting important features; and LSTM-SAM further determines the importance of features in the time dimension, and applies higher weights to more important features and time steps; at the same time, the present invention integrates manual features, constructs a multi-feature fusion framework, realizes the final RUL prediction, and improves the accuracy of the remaining useful life prediction of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The figure is a flowchart of an embodiment of the present invention.

[0051] Figure 2 This is a multi-feature fusion network diagram of a dual attention mechanism according to an embodiment of the present invention.

[0052] Figure 3a , Figure 3b , Figure 3c , Figure 3d It is a fitting diagram of the prediction results and the actual RUL of four sub-data sets FD001, FD002, FD003, and FD004 according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to better understand the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0054] In order to make up for the shortcomings of CNN in feature weighting, the present invention introduces the channel attention mechanism CAM (Channel Attention Mechanism). The channel attention mechanism, first proposed by Hu et al., can adaptively allocate channel weights so that the network can focus on more important features. Combining CNN and CAM not only enhances the feature extraction capability of the model, but also overcomes the shortcomings of traditional CNN in feature attention, thereby improving the accuracy of remaining useful life prediction.

[0055] At the same time, the spatial attention mechanism SAM (Spatial Attemtion Mechanism) is considered to be an effective way to solve long-term dependencies. First proposed by Bahdanau, the spatial attention mechanism provides direct correlation between different states of a sequence in a prediction model. As the name suggests, the spatial attention mechanism assigns different attention weights to different parts of a sequence, allowing the model to pay more attention to important information during training. SAM can effectively overcome the long-term dependency problem of LSTM networks, and combined with LSTM (LSTM-SAM) can better extract features in the time dimension.

[0056] like Figure 1 , Figure 2 As shown, the present invention proposes a remaining useful life prediction method based on a dual-attention multi-feature fusion network, comprising the following steps:

[0057] Step 1: Construct a data set for the collected multi-sensor data of the equipment, merge and filter it, and normalize the filtered data.

[0058] 1. Among the sensors, the data in the degradation experiment are relatively stable, which means that these data do not show degradation characteristics and cannot provide obvious degradation information. Therefore, the data of these sensors are eliminated and the data from the remaining sensors are used to train the model proposed by the present invention.

[0059] 2. The data has high noise, so it is necessary to normalize the data to eliminate this effect.

[0060] The Min-Max method is used to normalize the data. The specific formula is as follows:

[0061]

[0062] Where: x i Indicates the value at the current time point; and They refer to the maximum and minimum values ​​of the current sensor at all time points; x i ' represents the data obtained after normalization calculation.

[0063] Step 2: Based on the training set samples, a multi-feature fusion network model based on the dual attention mechanism is constructed. First, CNN is combined with CAM to preliminarily extract features in the spatial dimension; then, LSTM is combined with SAM to extract features in the temporal dimension, completing the integration of feature information in the spatial and temporal dimensions and enhancing the ability of feature extraction. Finally, the trained neural network model is used for RUL prediction.

[0064] In view of the excellent feature extraction capability of CNN, the present invention proposes a CNN-CAM module which integrates CNN and CAM.

[0065] First, the original time series data is preprocessed to form input data. The input data features can be expressed as:

[0066] X=[x1,x2,x3,...,x k ],where X∈R T×K

[0067] Where: x k represents the Kth channel in the feature map X; T and K represent the time step and the amount of sensor data respectively;

[0068] The convolution calculation is expressed as follows:

[0069]

[0070] Where: X i-1 represents the input of the i-th convolutional layer; W i Represents the i-th convolution kernel; and b i Represent the activation function and the bias term of the function respectively;

[0071] CAM is used to process the data output by the convolution layer. After squeezing the feature map X output by the convolution layer, the spatial information contained in the data is integrated through the global average pooling layer and the global maximum pooling layer. Given that the spatial dimension of the feature map is h×w×c, it becomes 1×1×c after squeezing. After squeezing, the feature map X of each channel will generate the corresponding channel descriptor P K , the formula is as follows:

[0072]

[0073] Where: x k (i,j) represents the element in the i-th row and j-th column in the feature map X;

[0074] P K Perform the excitation operation to generate the extracted descriptor L, which is expressed as:

[0075] L=E(P,FC)=tanh(FC2tanh(FC1P))

[0076] Where: FC1 represents a dimensionality reduction layer; FC2 represents a dimensionality increase layer; tanh is the activation function.

[0077] The two fully connected layers in the excitation layer reduce the complexity of the model by reducing the dimension, which can greatly simplify the calculation process and reduce the number of parameters. To avoid the problem of central symmetry, the tanh function sets the output weight between [-1,1], and then adopts the weighted processing method to complete the recalibration of the original features in the channel dimension. Finally, in order to integrate the data on the two branches of the global average pooling layer and the global maximum pooling layer, the element summation method is used to generate the output feature map, and the softmax function is used to normalize the weights. The normalized weights are assigned to the features of each channel through scaling operations, and the final output of this module is directly input to the next network layer.

[0078] The LSTM-SAM module that combines LSTM and SAM includes the following:

[0079] Before data flows into the LSTM network, it first passes through the forget gate, whose main function is to selectively forget the data transmitted at the previous moment. Its update calculation formula is:

[0080] f t =σ(W f [h t-1 ,x t ]+b f )

[0081] Where: x t is the input at time step t; h t-1 is the hidden state at time step t-1; W f represents the weight matrix; b f represents the bias term; σ represents the sigmoid function;

[0082] After being filtered by the forget gate, the data is passed to the input gate. The LSTM network input signal is determined by the input gate to determine the information and content to be updated, as shown below:

[0083] i t =σ(W i [h t-1 ,x t ]+b i )

[0084]

[0085] Where: W i , W Crepresents the weight matrix; b i 、b C represents the bias term; Indicates that x t and h t-1 The state quantity that integrates the state information of t Indicates the current status;

[0086] The output gate uses the cell state that needs to be output. The input signal determines the useful information for output through the output gate and immediately updates the current LSTM neural network cell state. The update calculation formula is as follows:

[0087] o t =σ(W o [h t-1 ,x t ]+b o )

[0088] h t =o t *tanh(C t )

[0089] Where: W o represents the weight matrix; b o represents the bias term;

[0090] The learning features output by the LSTM network can be expressed as X = {X1, X2, ..., X n} T , where T represents the transpose operation, and as the input x of SAM i ={x1,x2,…,x m} T , m represents the time step of the data, and the specific formula is as follows:

[0091] u i =softmax(W T x i +b)

[0092]

[0093] Where: u i represents the attention weight; W T and b are the weight matrices and bias terms in the attention network, a i is the influence of the ith eigenvalue on the target sequence data. All a i Add up to 1;

[0094] The feature sequence is weighted to obtain the output sequence, which is expressed as:

[0095]

[0096] Step three, input the test samples into the trained dual-attention multi-feature fusion network model to predict the remaining service life of the equipment.

[0097] Example 1

[0098] The simulator of the aviation propulsion simulation system simulates 90,000 pounds of thrust. The engine is set to be able to simulate operation at 40,000 feet sea level, with a Mach number of 0 to 0.90 and a temperature of -60 to 103°F at sea level. The components of the power include the fan, the combustion chamber, the low-pressure turbine, the low-pressure compressor, the high-pressure compressor, the high-pressure turbine, and the nozzle. The C-MAPSS data set includes 4 sub-datasets, namely FD001, FD002, FD003, and FD004. Each sub-dataset is further divided into training and test data sets. The training set and test set files are composed of sensor data of a specific engine in each operating cycle. The purpose is to predict the true RUL of each engine in the test file with a given sensor measurement value. The data set is arranged in an Nx26 matrix, that is, a data sample includes 26 variables: each engine unit number, each turbofan engine degradation cycle, 3 turbofan engine operating setting parameters, and the sampling data of 21 sensors on the turbofan engine in each operating cycle, where N corresponds to the number of sensor signals recorded for each engine. The detailed information of the dataset is given in Table 1. The dataset also contains multivariate temporal data of 21 sensors, including 100 time series from operation to failure.

[0099] Table 1C-MAPSS dataset information

[0100]

[0101]

[0102] In order to accurately verify the effectiveness and accuracy of the model used, the present invention proposes two evaluation indicators: root mean square error (RMSE) and score function (Score). The smaller the value of both indicators, the better the prediction performance. In this study, these two evaluation indicators will be used to evaluate the prediction performance of the model. The formula and definition of the root mean square error are as follows:

[0103]

[0104] Where: N represents the total number of samples, and i Represent the predicted RUL value and the actual RUL value respectively.

[0105] The formula and definition of the scoring function are as follows:

[0106]

[0107] Where: N represents the total number of samples, and i Represent the predicted RUL value and the actual RUL value respectively.

[0108] In order to verify the contribution of each component of the model, the present invention selects four methods including no CNN-CAM (NLH), no LSTM-SAM (CNH), no manual features (CLN) and the model used in the present invention for comparative experiments. The results are shown in Table 2.

[0109] Table 2 Experimental results of RMSE of ablation experiments on four datasets

[0110]

[0111] From the results, we can see that, in the case of fusion with manual features, the results of the model without channel attention mechanism in each subset are 15.19, 18.11, 14.69 and 18.31 respectively, while the results of the model without LSTM-AM in each subset are 14.55, 15.62, 12.99 and 16.49 respectively, which means that the channel attention mechanism can improve the accuracy of RUL prediction more than the time attention mechanism emphasized by LSTM-AM. The fusion of manual features does optimize the performance of the model to a certain extent, indicating that some manual features with domain knowledge may provide effective help for RUL prediction. Finally, the results of the method proposed in the present invention are reduced by 2.49, 3.64, 2.66 and 2.11 respectively on the four subsets, that is, a relative reduction of 16.39%, 20.1%, 18.1% and 11.52%, which can show that the fusion model used in the present invention is accurate and effective.

[0112] like Figure 3a , Figure 3b , Figure 3c , Figure 3d As shown in the figure, the predicted results and the actual RUL values ​​were fitted. The predicted RUL of the four data sets were well matched with the actual RUL values. The fitting curves on FD001 and FD003 were quite close to the true values, which shows that the model can effectively predict RUL. At the same time, it can be seen that the frequency of early predictions is still higher than that of late predictions, which to some extent still shows the stability of the model.

[0113] Tables 3 and 4 list the comparison of RMSE and Score between some advanced algorithms and the experimental results. In the table, the results in bold font represent the best results obtained among all methods. It can be seen from the table that the results of the architecture proposed in the present invention on the FD002, FD003, and FD004 subsets are better than some advanced algorithms. They are reduced by 0.77, 0.1, and 1.956, respectively, which is equivalent to a reduction of 5.05%, 0.83%, and 10.77%.

[0114] Table 3 Results of the research method of the present invention and other advanced methods on RMSE

[0115]

[0116]

[0117] Table 4 Results of the research method of the present invention and other advanced methods on Score

[0118]

[0119] In summary, compared with some existing advanced algorithms, the model of the present invention has achieved good results in both RMSE and SCORE evaluation indicators, and has been verified on 4 sub-datasets, showing the superiority and reliability of the model of the present invention.

[0120] The above are only preferred embodiments of the present invention, and certainly cannot be used to limit the scope of rights of the present invention. Therefore, equivalent changes made according to the scope of the patent application of the present invention still fall within the protection scope of the present invention.

Claims

1. The remaining useful life prediction method based on dual-attention multi-feature fusion network is characterized by: The steps include: Step 1: Construct a data set from the collected multi-sensor data of the equipment, merge and filter it, and normalize the filtered data; Step 2: Based on the training set samples, a multi-feature fusion network model based on the dual attention mechanism is constructed. First, CNN is combined with CAM to preliminarily extract features in the spatial dimension; then, LSTM is combined with SAM to extract features in the temporal dimension, completing the integration of feature information in the spatial and temporal dimensions. Step three, input the test samples into the trained dual-attention multi-feature fusion network model to predict the remaining service life of the equipment.

2. The remaining useful life prediction method based on dual attention multi-feature fusion network according to claim 1 is characterized in that: The data eliminated in the sensor data screening described in step 1 are: the data in the degradation experiment are stable, do not show degradation characteristics, and cannot provide obvious degradation information.

3. The remaining useful life prediction method based on dual attention multi-feature fusion network according to claim 2 is characterized in that: The data normalization process described in step 1 includes the following: The Min-Max method is used to normalize the data. The specific formula is as follows: Where: x i Indicates the value at the current time point; and They refer to the maximum and minimum values ​​of the current sensor at all time points; x i ' represents the data obtained after normalization calculation.

4. The remaining useful life prediction method based on dual attention multi-feature fusion network according to claim 3 is characterized in that: The construction of the CNN-CAM module combining CNN and CAM described in step 2 includes the following contents: The original time series data is formed into input data after data preprocessing. The input data features can be expressed as: X=[x1,x2,x3,…,x k ],whereX∈R T×K Where: x k represents the Kth channel in the feature map X; T and K represent the time step and the amount of sensor data respectively; The convolution calculation is expressed as follows: Where: X i-1 represents the input of the i-th convolutional layer; W i Represents the i-th convolution kernel; and b i Represent the activation function and the bias term of the function respectively; CAM is used to process the data output by the convolution layer. After squeezing the feature map X output by the convolution layer, the spatial information contained in the data is integrated through the global average pooling layer and the global maximum pooling layer. Given that the spatial dimension of the feature map is h×w×c, it becomes 1×1×c after squeezing. After squeezing, the feature map X of each channel will generate the corresponding channel descriptor P K , the formula is as follows: Where: x k (i,j) represents the element in the i-th row and j-th column in the feature map X; P K Perform the excitation operation to generate the extracted descriptor L, which is expressed as: L=E(P,FC)=tanh(FC2tanh(FC1P)) Where: FC1 represents a dimensionality reduction layer; FC2 represents a dimensionality increase layer; tanh is the activation function; To avoid the problem of central symmetry, the tanh function sets the output weight between [-1, 1], and then adopts a weighted processing method to complete the recalibration of the original features in the channel dimension. Finally, in order to integrate the data on the two branches of the global average pooling layer and the global maximum pooling layer, the element summation method is used to generate the output feature map, and the softmax function is used to normalize the weights. The normalized weights are assigned to the features of each channel through scaling operations, and the final output of the module is directly input to the next network layer.

5. The remaining useful life prediction method based on dual attention multi-feature fusion network according to claim 4 is characterized in that: The LSTM-SAM module construction combining LSTM and SAM described in step 2 includes the following contents: Before data flows into the LSTM network, it first passes through the forget gate, whose main function is to selectively forget the data transmitted at the previous moment. Its update calculation formula is: f t =σ(W f [h t-1 ,x t ]+b f ) Where: x t is the input at time step t; h t-1 is the hidden state at time step t-1; W f represents the weight matrix; b f represents the bias term; σ represents the sigmoid function; After being filtered by the forget gate, the data is passed to the input gate. The LSTM network input signal is determined by the input gate to determine the information and content to be updated, as shown below: i t =σ(W i [h t-1 ,x t ]+b i ) Where: W i , W C represents the weight matrix; b i , b C represents the bias term; Indicates that x t and h t-1 The state quantity that integrates the state information of t Indicates the current status; The output gate uses the cell state that needs to be output. The input signal determines the useful information for output through the output gate and immediately updates the current LSTM neural network cell state. The update calculation formula is as follows: the t =σ(W o [h t-1 ,x t ]+b o ) h t =o t *tanh(C t ) Where: W o represents the weight matrix; b o represents the bias term; The learning features output by the LSTM network can be expressed as X = {X1,X2,·.·,X n } T , where T represents the transpose operation, and as the input x of SAM i ={x1,x2,·.·,x m } T , m represents the time step of the data, and the specific formula is as follows: u i =softmax(W T x i +b) Where: u i represents the attention weight; W T and b are the weight matrices and bias terms in the attention network, a i is the influence of the ith eigenvalue on the target sequence data. All a i Add up to 1; The feature sequence is weighted to obtain the output sequence, which is expressed as:

Images