Multi-scale spatial-temporal feature fusion aero-engine residual life prediction model based on sensor data

Abstract

The invention belongs to the technical field of aero-engine life prediction, and discloses a multi-scale spatial-temporal feature fusion aero-engine residual life prediction model based on sensor data. According to the method, aiming at the problems that an existing model depends on a prior graph structure and is insufficient in adaptability, an adaptive graph construction mechanism is provided, a topological relation is dynamically learned by calculating a Pearson's correlation coefficient between sensors, and a threshold filtering and communication guarantee strategy is introduced. The model fuses multi-scale features, extracts statistical features to describe data distribution, captures sensor spatial dependence by using a graph attention network, and models spatio-temporal evolution information in combination with a bidirectional long-short term memory network, the graph attention network and a time sequence attention mechanism. And finally, statistical, space and time-space features are spliced, and residual life prediction is realized through a full connection layer. The method has strong adaptability, multi-feature fusion capability and relatively high interpretability, the prediction precision and robustness are remarkably improved, and the method is suitable for real-time life evaluation of multi-sensor systems such as aero-engines and the like.

Description

Technical Field

[0001] This invention belongs to the field of aero-engine life prediction technology, and relates to a multi-scale spatiotemporal feature fusion model for predicting the remaining life of aero-engines based on sensor data. Background Technology

[0002] Due to their inherent complexity and strong coupling, aero-engines face the challenge of establishing accurate single-physics prediction models in practice. Meanwhile, advancements in sensor and IoT technologies have made acquiring massive amounts of operational data feasible, thus driving the rise of data-driven methods. However, traditional data-driven methods have limitations when processing the high-dimensional and complex monitoring data of modern aero-engines, including insufficient nonlinear representation capabilities and reliance on extensive manual feature engineering. In recent years, with the continuous development of deep learning technology, researchers have begun to explore advanced models capable of simultaneously modeling spatiotemporal characteristics, providing new solutions for aero-engine life prediction.

[0003] In their review article, "A survey on graph neural networks for maintaining useful life prediction: Methodologies, evaluation, and future trends," Yucheng Wang et al. conducted a unified benchmark test on various cutting-edge graph neural network methods, analyzed their performance and complexity, and pointed out that a general graph model adaptable to different tasks, datasets, and sensor configurations is the future direction. However, most models in the field of aero-engine life prediction, represented by the STFA method proposed by Ziqian Kong in his article "Spatio-Temporal Fusion Attention: A Novel Approach for Maintaining Useful Life Prediction Based on Graph Neural Network," require prior knowledge for graph structure construction, resulting in insufficient adaptability. Therefore, a novel method that can effectively integrate spatiotemporal feature extraction, graph structure learning, and adaptive optimization has become an urgent challenge and key direction for current research. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a multi-scale spatiotemporal feature fusion model based on sensor data for predicting the remaining lifespan of aero-engines.

[0005] The technical solution of the present invention:

[0006] A multi-scale spatiotemporal feature fusion model for predicting the remaining life of aero-engines based on sensor data includes the following steps:

[0007] Step 101: Data Preparation;

[0008] Data is collected from N sensors. The data is cleaned to remove noise, missing values, or outliers, and regularized to ensure consistent dimensions across different sensors. The processed data is then divided into training and testing sets, where the original training data sequence is... The corresponding collection time length is The original sequence of the test data is The corresponding collection time length is The original sequences of both training and test data are uniformly adopted using S... raw The corresponding data collection time length is uniformly represented by T. raw This indicates that by constructing a sliding window of length w and combining it with a sliding step size of length p, the final dataset S for training and testing the multi-scale spatiotemporal feature fusion model for predicting the remaining life of aero-engines is generated.

[0009]

[0010] In the formula: the total number of windows M is given by the formula Decision, among which Indicates rounding down;

[0011] From the above formula, we can see that the dataset S has a total of Given a sample set with batch size b, the dataset used in the k-th batch is... To obtain from the dataset The sample consists of b randomly selected samples:

[0012]

[0013] In the formula: k 1 ， k 2 ， … ， k b =Sample [ 1 ， M] ,Right now From the interval [ 1 ， M] The indices of b randomly sampled integer values; This indicates taking the first element from dataset S. Number of samples; k represents the batch index, used to identify the sequential position of the currently processed data batch in the overall data sequence; For the number of sensors; Represents the real number field. express A three-dimensional real tensor of shape (b, N, W);

[0014] Step 102: Adaptive graph structure construction;

[0015] A data-driven adaptive graph construction method is proposed; for the dataset used in the k-th batch... The graph construction process is as follows:

[0016] Representing sensor networks as graph structures ,in, It is a set of nodes, with each node corresponding to a sensor; Let be the set of edges, representing the connection relationships between sensors; This is an adjacency matrix that encodes the connection strength between nodes.

[0017] Calculate the i-th sample Pearson correlation matrix among all sensors:

[0018]

[0019] In the formula: For the i-th sample in the k-th batch The Pearson correlation coefficients between all pairs of N sensors; Representative calculation sample The Pearson correlation coefficients between any two of the N sensors in the dataset; express It is a two-dimensional real matrix of shape (N, N);

[0020] The correlation coefficient is defined as:

[0021]

[0022] In the formula: , representing the u-th and v-th sensors; , represents the sampled value at time t in the feature tensor of length w; , This represents the sampling at time t from the u-th and v-th sensors; , This represents the average of the sampled values ​​from sensors u and v within the length of the sliding window.

[0023] Aggregate the correlation tensors of all b samples within the k-th batch to obtain the average correlation tensor of the k-th batch. , It is a two-dimensional real matrix of dimension (N, N):

[0024]

[0025] The correlation matrix is ​​converted into an adjacency matrix, and then absolute value transformation and self-join removal are performed:

[0026]

[0027] In the formula: This represents element-wise absolute value operation; Extracting the diagonal elements of a matrix when the input is a vector. Construct a diagonal matrix;

[0028] use The operation involves selecting the node with the strongest correlation for each node. One neighbor:

[0029] for

[0030] In the formula: The K values ​​representing the largest Pearson correlation coefficients between the u-th sensor and the other N-1 sensors are a one-dimensional tensor of size (1, K). For the corresponding neighbor index, there is a one-dimensional tensor of size (1, K);

[0031] Weakly correlated connections are filtered out by applying a set threshold τ, and the edge is only retained if the Pearson correlation coefficient between the two sensors is greater than the set threshold:

[0032] for ,have

[0033]

[0034] In the formula: This represents the j-th value among the K largest Pearson correlation coefficients between the u-th sensor and every N-1 other sensors. When this value exceeds a set threshold... At that time, at the edge set Add edge , This is the index corresponding to the j-th sensor;

[0035] To ensure the connectivity of the graph structure, a guarantee mechanism is introduced: when no edges exist after threshold filtering, a minimum connected circular graph is constructed as the default structure.

[0036]

[0037] In the formula: This is the final set of edges;

[0038] Step 103: Spatial feature extraction;

[0039] For the dataset used in the kth batch The measurement sequence of each node within the time window is used as its feature tensor. To better perform batch operations, the input node feature matrix is ​​set to... Sliding window edge index is The forward propagation process is as follows:

[0040]

[0041] In the formula: It is a dimension A matrix, where each row is a... A tensor of dimension 1 represents the node features after processing by a graph attention network. For graph attention networks, For batch normalization operations, For activation functions;

[0042] Step 104: Statistical feature extraction;

[0043] For the dataset used in the kth batch Calculate the mean along the third dimension, that is, for :

[0044]

[0045] In the formula: The value at time t represents the sampled value in the feature tensor of sliding window length w; Representative batch data

[0046] The value of the u-th sensor at time t;

[0047] Similarly, the standard deviation is calculated along the third dimension, i.e., for :

[0048]

[0049] In the formula: It is a dimension The matrix, This represents the data of the kth batch. Standard value; Represents the kth batch of data The sample, the mean value collected within the u-th sensor window;

[0050] Concatenate the mean and standard deviation along the feature dimension:

[0051]

[0052] In the formula, It is the final statistical characteristic; This indicates splicing along the second dimension; It is a fully connected layer; The statistical feature dimensions are set;

[0053] Step 105: Generation of graph-level feature vectors;

[0054] The node features obtained in step 103 Reshape:

[0055]

[0056] In the formula: reshape is the matrix dimension transformation operation, which transforms the two-dimensional tensor. Convert to 3D tensor The reshaping operation restores the data to its original state. Three-dimensional real tensor;

[0057] Perform global pooling operation, for The second dimension is averaged, that is, the features of all nodes are averaged to obtain the graph-level feature vector. :

[0058]

[0059] In the formula: Representative batch data Node features of the sample;

[0060] Step 106: Spatiotemporal feature modeling;

[0061] For the kth batch of data The time feature matrix is ​​constructed by using N sensor measurement data points at each time point within the time window as its feature vector. The time series data provided by the sensors are as follows:

[0062]

[0063] In the formula: reshape is the matrix dimension transformation operation, which transforms the matrix into its dimension. Transpose the first and second dimensions to obtain a feature tensor suitable for temporal feature processing. ;

[0064] The graph-level feature vectors are expanded so that graph features can be involved at each time point:

[0065]

[0066] In the formula: Indicates the outer product. It is a vector of length w consisting entirely of 1s;

[0067] Concatenate with the time feature matrix:

[0068]

[0069] Input bidirectional long short-term memory network The features obtained by fusing spatiotemporal characteristics are:

[0070]

[0071] In the formula: For the hidden layer dimensions of a bidirectional long short-term memory network, It is a dimension The three-dimensional tensor;

[0072] Step 107: Focusing on spatiotemporal information;

[0073] The output of the Long Short-Term Memory network in step 106 is processed using a temporal attention mechanism. Weighting;

[0074]

[0075] In the formula, Softmax is applied along the last dimension, so that the weights of each feature channel at different time steps form a probability distribution.

[0076] Finally, the attention weights are multiplied element-wise with the original input to obtain the weighted spatiotemporal features. :

[0077]

[0078] In the formula: This represents element-wise multiplication, the Hadamard product;

[0079] Step 108: Multi-scale information stitching;

[0080] After flattening the features obtained in step 107, a fully connected layer is used to perform dimensional mapping to obtain the final spatiotemporal module information:

[0081]

[0082] In the formula,

[0083] Weighted spatiotemporal features reshape The operation of flattening a three-dimensional tensor into a two-dimensional matrix. The spatiotemporal and feature dimensions for each batch;

[0084] Statistical characteristics obtained in step 104 Spatial features obtained in step 105 The spatiotemporal features obtained in step 108 To splice:

[0085]

[0086] Step 109, Remaining Life Prediction:

[0087]

[0088] In the formula, To output the predicted value, The input is a multi-scale fusion feature.

[0089] The beneficial effects of this invention are as follows: The model of this invention, combining multi-scale feature extraction and graph structure learning from sensor data, can efficiently and accurately predict the remaining service life of aero-engines, and is suitable for real-time analysis and processing of complex multi-dimensional sensor data. It improves the diagnostic accuracy and interpretability of data-driven aero-engine life prediction algorithms, and has broad application potential. Attached Figure Description

[0090] Figure 1 It is a feature flow diagram of a multi-scale spatiotemporal feature fusion prediction model for the remaining life of aero-engines.

[0091] Figure 2 This is a framework diagram of the multi-scale spatiotemporal feature fusion prediction model algorithm for the remaining life of aero-engines.

[0092] Figure 3 This is a flowchart of the modeling process for a multi-scale spatiotemporal feature fusion prediction model for the remaining lifespan of aero-engines.

[0093] Figure 4 This is a flowchart for generating a graph structure. Detailed Implementation

[0094] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0095] Reference Figure 1A multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data is proposed, and its multi-scale feature fusion is reflected in:

[0096] Starting with raw sensor data, statistical features are extracted, and the distribution characteristics of the data are characterized by calculating statistical measures such as mean and standard deviation. Secondly, based on an adaptive graph structure and graph attention network, the topological associations and spatial dependencies between sensors are extracted to form spatial features. Finally, the spatial features are fused with local time series extracted based on a sliding window, and long-term time dependencies are modeled through a bidirectional long short-term memory network to construct a spatiotemporal fusion feature that simultaneously contains sensor associations and temporal evolution information. This bottom-up, local-to-global, and single-dimensional-to-multi-dimensional fusion structure effectively enhances the model's ability to understand and model complex sensor data.

[0097] Reference Figures 2-3 This paper details the specific steps of a multi-scale spatiotemporal feature fusion model for predicting the remaining life of aero-engines based on sensor data:

[0098] 1. Data preparation;

[0099] This invention is validated using the NASA C-MAPSS public dataset as an example. The preprocessing process includes selecting appropriate sensor data as features, constructing remaining lifetime labels for the training set data and performing amplitude limiting, and standardizing based on training set statistics. Finally, sequence samples are generated by setting a sliding window of length w=50 to obtain an input tensor with dimensions (number of samples, 17, 50) and the corresponding remaining lifetime labels, providing the model with normalized and structured input data.

[0100] 2. Adaptive graph construction;

[0101] Reference Figure 4 This paper proposes a data-driven approach to dynamically learn the spatial topology of sensor networks. First, Pearson correlation coefficients are calculated for each pair of sensors in each batch of data to construct an initial fully connected weighted graph. Next, the absolute value of the correlation coefficient matrix is ​​taken, and self-connections are removed. Then, for each sensor node, only the top K most correlated neighbors are retained, forming sparse adjacency relationships to highlight key physical associations. Finally, weak connections are filtered by setting a threshold to enhance robustness. A safeguard mechanism is designed: if the filtered graph is empty, a minimum connected circular graph is automatically constructed as the default topology to ensure the graph structure remains valid and provides a reliable foundation for subsequent spatial feature extraction.

[0102] 3. Spatial feature extraction;

[0103] The preprocessed sensor data and the dynamically generated graph structure from the previous step are input into the Graph Attention Network (GAT) module. This module aggregates the neighbor information of each sensor through an attention mechanism, thereby learning the complex nonlinear spatial dependencies between sensors.

[0104] 4. Statistical feature extraction;

[0105] The mean and standard deviation of the raw sensor data are calculated directly within a time window. These statistics serve as descriptive features, characterizing the stability and trends of the equipment's macroscopic operating status.

[0106] 5. Generation of graph-level feature vectors;

[0107] Global pooling is performed on the node-level spatial features output by GAT. This invention uses average pooling to generate a fixed-length graph-level feature vector. This vector summarizes the overall spatial state of the entire sensor network at a certain moment.

[0108] 6. Spatiotemporal feature modeling;

[0109] The aforementioned graph-level feature vectors are concatenated with the original sensor time-series data and input into a bidirectional long short-term memory (LSTM) network. This bidirectional LSTM network can capture complex dynamic temporal patterns during device degradation from both forward and backward perspectives.

[0110] 7. Focus on spatiotemporal information;

[0111] A temporal attention mechanism is introduced on top of the output of the bidirectional long short-term memory network. This mechanism assigns different weights to features at different historical time steps, enabling the model to automatically focus on the most critical historical moments for predicting the current remaining lifespan, thereby enhancing the model's interpretability and prediction accuracy.

[0112] 8. Multi-scale information stitching;

[0113] Features from different levels and perspectives—namely, graph-level feature vectors summarizing the spatial state of the system, statistical feature vectors describing macroscopic statistical laws, and refined spatiotemporal feature vectors filtered by attention—are concatenated to form a comprehensive multi-scale fusion feature representation.

[0114] 9. Remaining life prediction;

[0115] The fused features are input into a fully connected layer for regression, outputting a predicted remaining life of the aero-engine. The entire model is trained end-to-end, optimizing all parameters to minimize prediction error.

[0116] In summary, this invention proposes a spatiotemporal neural network based on an adaptive graph structure. This model can dynamically represent each sensor node and effectively capture the complex spatiotemporal relationships between them, thus solving the challenge of modeling nonlinear dependencies in multi-sensor data from aero-engines. This method has the following main advantages:

[0117] 1) Adaptability: It can dynamically construct the sensor network graph structure based on real-time data, without relying on prior physical knowledge or fixed topology, thus improving the model's adaptability to different operating states and actuator configurations.

[0118] 2) Multi-scale feature fusion: It effectively integrates the local statistical features, spatial dependencies and global temporal evolution patterns of the sensor to form a comprehensive and discriminative degradation representation, thereby enhancing the expressive power of the model.

[0119] 3) Improved interpretability: Through graph attention and temporal attention mechanisms, the model can identify key sensor associations and key historical moments that are crucial to the prediction, providing some mechanistic support for the prediction results.

[0120] 4) High Accuracy and Stability: To verify the effectiveness of this invention, we conducted comparative experiments on the publicly available C-MAPSS dataset. Experimental results show that, compared with traditional time-series models such as LSTM and CNN, as well as graph neural network methods requiring fixed graph structures (such as STFA), the model proposed in this invention achieves significantly better performance in both the root mean square error (shown in Table 1) and the scoring function (shown in Table 2), two key evaluation metrics. This fully demonstrates the enormous potential of the adaptive graph structure and multi-scale spatiotemporal feature fusion strategy in improving the accuracy and robustness of aero-engine remaining life prediction.

[0121] Table 1

[0122]

[0123]

[0124] Table 2

[0125]

[0126] These advantages make this method a promising candidate for application in complex multi-sensor systems, especially suitable for tasks that require simultaneous handling of time and space dependencies, such as predicting the remaining life of aero-engines.

Claims

1. A multi-scale spatiotemporal feature fusion model for predicting the remaining life of aero-engines based on sensor data, characterized in that, Includes the following steps: Step 101: Data Preparation; Step 102: Adaptive graph structure construction; Step 103: Spatial feature extraction; Step 104: Statistical feature extraction; Step 105: Generation of graph-level feature vectors; Step 106: Spatiotemporal feature modeling; Step 107: Focusing on spatiotemporal information; Step 108: Multi-scale information stitching; Step 109: Remaining life prediction.

2. The multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data according to claim 1, characterized in that, Step 101, Data Preparation: The specific steps are as follows: Data is collected from N sensors, and the data is cleaned to remove noise, missing values ​​or outliers, and the data is regularized to ensure that the dimensions of the data from different sensors are consistent. Subsequently, the processed data was divided into a training set and a test set, where the original sequence of the training data was... The corresponding collection time length is The original sequence of the test data is The corresponding collection time length is The original sequences of both training and test data are uniformly adopted using S... raw The corresponding data collection time length is uniformly represented by T. raw This indicates that by constructing a sliding window of length w and combining it with a sliding step size of length p, the final dataset S for training and testing the multi-scale spatiotemporal feature fusion model for predicting the remaining life of aero-engines is generated. ； In the formula: the total number of windows M is given by the formula Decision, among which Indicates rounding down; From the above formula, we can see that the dataset S has a total of Given a sample set with batch size b, the dataset used in the k-th batch is... To obtain from the dataset Randomly sample b samples composition: ； In the formula: ,Right now From the interval The indices of b randomly sampled integer values; This indicates taking the first element from dataset S. Number of samples; k represents the batch index, used to identify the sequential position of the currently processed data batch in the overall data sequence; Number of sensors; Represents the real number field. express A three-dimensional real tensor of shape (b, N, W).

3. The multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data according to claim 2, characterized in that, Step 102, the adaptive graph structure construction is as follows: A data-driven adaptive graph construction method is proposed; for the dataset used in the k-th batch... The graph construction process is as follows: Representing sensor networks as graph structures ,in, It is a set of nodes, with each node corresponding to a sensor; Let be the set of edges, representing the connection relationships between sensors; This is an adjacency matrix that encodes the connection strength between nodes. Calculate the i-th sample Pearson correlation matrix among all sensors: ； In the formula: For the i-th sample in the k-th batch The Pearson correlation coefficients between all pairs of N sensors; Representative calculation sample The Pearson correlation coefficients between any two of the N sensors in the dataset; express It is a two-dimensional real matrix of shape (N, N); The correlation coefficient is defined as: ； In the formula: , representing the u-th and v-th sensors; , represents the sampled value at time t in the feature tensor of length w; , This represents the sampling at time t from the u-th and v-th sensors; , This represents the average of the sampled values ​​from sensors u and v within the length of the sliding window. Aggregate the correlation tensors of all b samples within the k-th batch to obtain the average correlation tensor of the k-th batch. , It is a two-dimensional real matrix of dimension (N, N): ； The correlation matrix is ​​converted into an adjacency matrix, and then absolute value transformation and self-join removal are performed: ； In the formula: This represents element-wise absolute value operation; Extracting the diagonal elements of a matrix when the input is a vector. Construct a diagonal matrix; use The operation involves selecting the node with the strongest correlation for each node. One neighbor: for ; In the formula: The K values ​​representing the largest Pearson correlation coefficients between the u-th sensor and the other N-1 sensors are a one-dimensional tensor of size (1, K). For the corresponding neighbor index, there is a one-dimensional tensor of size (1, K); Weakly correlated connections are filtered out by applying a set threshold τ, and the edge is only retained if the Pearson correlation coefficient between the two sensors is greater than the set threshold: for ,have ； In the formula: This represents the j-th value among the K largest Pearson correlation coefficients between the u-th sensor and every N-1 other sensors. When this value exceeds a set threshold... At that time, at the edge of the set Add edge , This is the index corresponding to the j-th sensor; To ensure the connectivity of the graph structure, a guarantee mechanism is introduced: when no edges exist after threshold filtering, a minimum connected circular graph is constructed as the default structure. ； In the formula: This is the final set of edges.

4. The multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data according to claim 3, characterized in that, Step 103, Spatial Feature Extraction, is as follows: For the dataset used in the kth batch The measurement sequence of each node within the time window is used as its feature tensor. To better perform batch operations, the input node feature matrix is ​​set to... Sliding window edge index is The forward propagation process is as follows: ； In the formula: It is a dimension A matrix, where each row is a... A tensor of dimension 1 represents the node features after processing by a graph attention network. For graph attention networks, For batch normalization operations, This is the activation function.

5. The multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data according to claim 1, characterized in that, Step 104, Statistical Feature Extraction, is as follows: For the dataset used in the kth batch Calculate the mean along the third dimension, that is, for : ； In the formula: The value at time t represents the sampled value in the feature tensor of sliding window length w; Representative batch data ; The value of the u-th sensor at time t; Similarly, the standard deviation is calculated along the third dimension, i.e., for : ； In the formula: It is a dimension The matrix, This represents the data of the kth batch. The standard value; Represents the kth batch of data The sample, the mean value collected within the u-th sensor window; Concatenate the mean and standard deviation along the feature dimension: ； In the formula, It is the final statistical characteristic; This indicates splicing along the second dimension; It is a fully connected layer; The statistical feature dimensions are set.

6. The multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data according to claim 5, characterized in that, Step 105, Graph-level feature vector generation, is as follows: The node features obtained in step 103 Reshape: ； In the formula: reshape is the matrix dimension transformation operation, which transforms the two-dimensional tensor. Convert to 3D tensor The reshaping operation restores the data to its original state. Three-dimensional real tensor; Perform global pooling operation, for The second dimension is averaged, that is, the features of all nodes are averaged to obtain the graph-level feature vector. : ； In the formula: Representative batch data Node features of the sample.

7. The multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data according to claim 6, characterized in that, Step 106, Spatiotemporal feature modeling, is as follows: For the kth batch of data The time feature matrix is ​​constructed by using N sensor measurement data points at each time point within the time window as its feature vector. The time series data provided by the sensors are as follows: ； In the formula: reshape is the matrix dimension transformation operation, which transforms the matrix into its dimension. Transpose the first and second dimensions to obtain a feature tensor suitable for temporal feature processing. ; The graph-level feature vectors are expanded so that graph features can be involved at each time point: ； In the formula: Indicates the outer product. It is a vector of length w consisting entirely of 1s; Concatenate with the time feature matrix: ； Input bidirectional long short-term memory network The features obtained by fusing spatiotemporal characteristics are: ； In the formula: For the hidden layer dimensions of a bidirectional long short-term memory network, It is a dimension The three-dimensional tensor.

8. The multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data according to claim 7, characterized in that, Step 107, Spatiotemporal Information Focusing, is as follows: The output of the Long Short-Term Memory network in step 106 is processed using a temporal attention mechanism. Weighting; ； In the formula, Softmax is applied along the last dimension, so that the weights of each feature channel at different time steps form a probability distribution. Finally, the attention weights are multiplied element-wise with the original input to obtain the weighted spatiotemporal features. : ； In the formula: This represents element-wise multiplication, the Hadamard product.

9. The multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data according to claim 8, characterized in that, Step 108, Multi-scale information stitching, is as follows: After flattening the features obtained in step 107, a fully connected layer is used to perform dimensional mapping to obtain the final spatiotemporal module information: ； In the formula, ; Weighted spatiotemporal features reshape The operation of flattening a three-dimensional tensor into a two-dimensional matrix. The spatiotemporal and feature dimensions for each batch; Statistical characteristics obtained in step 104 The spatial features obtained in step 105 The spatiotemporal features obtained in step 108 To splice: 。 10. The multi-scale spatiotemporal feature fusion prediction model for aero-engine remaining life based on sensor data according to claim 1, characterized in that, Step 109, Remaining Life Prediction: The specific prediction is as follows: ； In the formula, To output the predicted value, The input is a multi-scale fusion feature.

Images