Aero-engine gas path fault diagnosis space-time network model of adaptive graph structure advantage

Abstract

The invention belongs to the technical field of aero-engine fault diagnosis, and discloses an aero-engine gas path fault diagnosis space-time network model of a self-adaptive graph structure advantage. Firstly, multi-source sensor data are processed through a time sequence factor extraction module, changes in time dimension are effectively captured, and the stability of a model is improved through a normalization layer. Thirdly, correcting improper spatial relationship representation in the time factor by adopting a random embedding method, constructing an initial graph structure by utilizing cosine similarity calculation, selecting hidden information in the graph structure again in combination with an attention mechanism, and then performing deep fusion on spatial features by using a message passing mechanism; and the node embedding is more accurate. According to the method, graph structure information in sensor data can be adaptively captured, the performance of the model in complex spatio-temporal information processing is remarkably improved, characteristic representation with higher robustness can be generated, and subsequent analysis and prediction are supported.

Description

Technical Field

[0001] The invention belongs to the technical field of aircraft engine fault diagnosis, and relates to an aircraft engine gas path fault diagnosis spatiotemporal network model with an adaptive graph structure advantage, which is used for aircraft engine gas path fault diagnosis and state monitoring. Background Art

[0002] As the core component of aircraft, the operating status of aircraft engines is crucial to flight safety. In the complex system of aircraft engines, the failure of the gas path system is one of the important factors affecting engine performance and safety. Most existing fault diagnosis methods rely on single sensor signal analysis or rule-based fault diagnosis models, which are difficult to effectively deal with the complex nonlinear relationship between multi-sensor data in the gas path system. This leads to insufficient accuracy and real-time performance of fault detection, especially when dealing with sensor noise, data loss and nonlinear dynamic relationships.

[0003] With the advancement of aircraft engine sensor technology, it has become possible to obtain a large amount of multi-dimensional sensor data in real time. These data can reflect the temperature, pressure, flow and other parameters of the engine under different working conditions. However, how to effectively integrate and analyze these complex multi-dimensional data to achieve accurate fault diagnosis of the gas path system is still a technical challenge.

[0004] Yucheng Wang et al. summarized in the article "Local–Global Correlation Fusion-Based GraphNeural Network for Remaining Useful Life Prediction" that traditional methods based on static graphs or without time series features are difficult to capture the dynamic changes and spatial dependencies of sensor data, and cannot fully utilize the complex correlation information between sensors. In addition, existing graph structure learning methods usually lack adaptability and are insufficient to fully capture complex spatial information, thus limiting their performance in prediction. Therefore, there is an urgent need for a new method that can effectively combine spatiotemporal feature extraction, graph structure learning, and adaptive optimization to improve the accuracy and real-time performance of fault diagnosis. Summary of the invention

[0005] In view of the above problems, the present invention proposes a spatiotemporal network model with an adaptive graph structure advantage for use in aero-engine gas path fault diagnosis.

[0006] The technical solution of the present invention:

[0007] An aero-engine gas path fault diagnosis spatiotemporal network model with an adaptive graph structure advantage comprises the following steps:

[0008] 101. Data preparation

[0009] Collect data from N sensors, clean the data to remove noise, missing values ​​or outliers, and regularize the data to ensure that the dimensions of data from different sensors are consistent; then, divide the processed data into training set and test set, where the training data sequence is S train , the corresponding acquisition time length is T train ; The test data sequence is S test , the corresponding acquisition time length is T test ;

[0010] 102. Sliding Window Construction

[0011] By constructing a sliding window of length w, the data finally used for model training and testing is generated; the format of the training data is:

[0012] S1={S train (1:w),…,S train (T train-w+1 :T train )}

[0013] The data segment S1(t) used at the tth training time step is S train (t:t+w), this data segment is a w×N matrix, which is the measurement value of N sensors in a window of length w;

[0014] The test data is processed in the same way as the training data, and the format of the obtained test data is:

[0015] S2={S test (1:w),…,S test (T test-w+1 :T test )}

[0016] The data segment S2(t) used at the tth test time step is S test (t:t+w), this data segment is a w×N matrix, which is the measurement value of N sensors in a window of length w;

[0017] 103. Time Series Feature Extraction

[0018] The input data S1(t) of the t-th training time step contains information spanning w time steps, and the output data of the t-th time step is Represents the final fault classification result; first, the LSTM model is used to process the input data S1(t), capture the features within the time range t to t+w, and generate a hidden state matrix H containing time series features t , provide follow-up support for the output of the model;

[0019] H t =LSTM(S1(t))

[0020] Among them, H t It is a w×M matrix, indicating that within a time window of length w, the LSTM layer generates a feature embedding of length M for each time step;

[0021] 104. Temporal Embedding Generation

[0022] For the hidden state matrix H containing temporal features t Crop and only keep the hidden layer information of the most recent K time steps to obtain the processed hidden layer information H' t ;

[0023] H' t =H t [-K:,∶]

[0024] Among them, H' t is a K×M matrix, representing H' t Contains the hidden layer feature information of the last K time steps, and the hidden information dimension of each time step is M;

[0025] Subsequently, the hidden layer feature information is mapped from the high-dimensional feature space with dimension M to the low-dimensional feature space with dimension N' through the fully connected layer; then, the mapping result is transposed to generate the initial embedding E of the graph neural network node in the subsequent generation. t Be prepared;

[0026] First, the processed hidden layer information is further integrated through the fully connected layer:

[0027] X=(WH' t +b) T

[0028] Where X is a matrix of N'×K, indicating that there are N' nodes in total, the dimension of the feature vector of each node is K, W is the weight matrix of the fully connected layer, and b is the bias vector of the fully connected layer;

[0029] Finally, the training process is further stabilized by layer normalization to generate the initial embedding E of the graph neural network nodes with N' nodes. t ;

[0030] E t =LayerNorm(X)

[0031] Among them, E t is an N'×K matrix, which represents the initial embedding matrix of the node that integrates the time features. N' is the number of nodes in the subsequent graph neural network input graph, where the dimension of the feature vector of each node is K;

[0032] 105.Spatial relationship correction

[0033] Generate a random embedding matrix E and use it as a random embedding vector to supplement node features;

[0034] E=randn(N',D)

[0035] Among them, E is the generated random embedding matrix, N' is the number of nodes of the subsequent graph neural network, and D is the set random embedding dimension;

[0036] Initially embed the nodes that incorporate time features into E t Concatenate with the generated random embedding matrix E to correct the initial node embedding E t The incorrect spatial relationships that may exist in the node initial embedding E t There are potentially more hidden patterns in ;

[0037] V=E t +E

[0038] Among them, V is the node embedding generated after spatial relationship correction, which is an N'×(K+D) matrix;

[0039] 106. Dynamic graph construction

[0040] The above steps generate a feature embedding of length K+D for each of the N' nodes; the feature embedding v of the i-th node i and the feature embedding v of the jth node j The similarity between them is used to indicate the similarity between the behaviors of node i and node j. The preliminary graph structure is constructed by calculating the cosine similarity between the feature embeddings of the nodes and combining the TopK method.

[0041] Use a directed graph to build the relationship between nodes, and use the adjacency matrix A to represent the directed graph, where A ij Indicates the existence of a directed edge from node i to node j. In the absence of prior information, the candidate relationships of node i are all nodes except itself. Each node i has a set of candidate relationships, that is, the nodes that the node i may depend on:

[0042]

[0043] In order to select the dependency relationship of node i among these candidate relationships, we calculate the embedding vector v of node i. i With its candidate nodes The embedding vector v j The similarity between them is calculated, and the TopK nodes with the highest similarity are selected as the neighbor nodes of the node. The calculation formula is as follows:

[0044]

[0045] First calculate e ji , that is, the embedding vector v of node i i Relationship with Candidate The normalized dot product between the embedding vectors of all nodes in ; then, the top k normalized dot products are selected; TopK represents the index of the top k values ​​in its input, which is the normalized dot product; the value of k is selected by the user according to the desired sparsity level;

[0046] 107.Graph Information Fusion

[0047] The embedding vector of the i-th node is v i , the embedding vector of the jth node is v j , these vectors represent different behaviors of different nodes; according to the obtained adjacency matrix A, based on the attention mechanism, the information between nodes is further screened and fused, and the aggregate representation z of node i is calculated as follows i :

[0048]

[0049] Among them, z i represents the feature representation of the i-th node after aggregation, α ii is the attention coefficient of node i to itself, α ij is the attention coefficient of node i to neighbor node j, W att is a trainable weight matrix used to linearly transform feature vectors, N(i) is the set of neighbor nodes of node i; ReLU represents the linear rectification function;

[0050] α ij is the normalized attention coefficient, which is solved as follows:

[0051]

[0052] Among them, π(i,j) is the attention coefficient between node i and node j; v i is the embedding vector of the i-th node; v j is the embedding vector of the jth node; represents concatenation; a is the learning coefficient vector of the attention mechanism; LeakyReLU is used as the nonlinear activation function to calculate the attention coefficient; α ij is the attention coefficient normalized by the softmax function, N(i) is the set of neighbor nodes of node i;

[0053] 108.Fault diagnosis and prediction

[0054] The feature vector after the information of each node is aggregated is z i , concatenate the features of N' nodes and input them into the multi-layer perceptron for classification;

[0055]

[0056] Among them, f represents the fully connected layer; The output classification result; the model outputs a C-dimensional vector Represents the predicted probability of C-1 types of faults; the model is optimized using cross entropy loss for training.

[0057] Beneficial effects of the invention: The model of the invention combines the spatiotemporal feature extraction, graph structure learning and adaptive embedding enhancement of sensor data, can efficiently and accurately diagnose the fault conditions of the air path system of an aircraft engine, 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 aircraft engine fault diagnosis algorithms and has broad application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a spatiotemporal network model with the advantages of adaptive graph structure and is used in the aero-engine fault diagnosis flowchart.

[0059] Figure 2 It is a flow chart for modeling spatiotemporal network models with the advantages of adaptive graph structure.

[0060] Figure 3 It is a flow chart of the algorithm of the spatiotemporal network model with the advantages of adaptive graph structure.

[0061] Figure 4 It is a schematic diagram of the adaptive graph structure module.

[0062] Figure 5 It is a schematic diagram of the performance of different random embedding dimensions.

[0063] Figure 6 It is a 2D dimensionality reduction visualization diagram of the feature extraction effect of the present invention and other commonly used methods, where (a) is LSTM, (b) is 1D-CNN, (c) is 2D-CNN, (d) is LSTM_GCN, (e) is the present invention (embedding dimension is 0), and (f) is the present invention (embedding dimension is 15). DETAILED DESCRIPTION

[0064] The present invention is further described below in conjunction with the accompanying drawings and implementation modes.

[0065] Reference Figure 1 , a spatiotemporal network model for aircraft engine gas path fault diagnosis with the advantages of adaptive graph structure, including the following contents:

[0066] The aircraft engine fault diagnosis system of the present invention includes a sensor acquisition module, a data processing module, a model building module and a fault detection module. The system relies on a precision sensor network in the engine to collect key parameters such as fuel pressure, oil volume, flow, temperature and lubricating oil pressure in real time. These data are transmitted to the monitoring workstation and the accuracy and consistency of the data are ensured through preprocessing steps such as correlation analysis, filtering, regularization and window division. Window division enables the system to better capture the time-varying characteristics in the data. The processed data is combined with time embedding technology and graph model technology to extract spatiotemporal features and construct an adaptive dynamic graph structure to accurately capture the spatiotemporal associations between sensors. Based on the neural network algorithm, the dynamic graph structure is continuously updated to adapt to real-time sensor data to ensure high accuracy and stability of fault detection. Finally, the system can monitor the health status of the aircraft engine in real time, accurately classify and predict potential faults, issue warnings in time, and provide data support for maintenance and decision-making, thereby ensuring flight safety and optimizing engine performance.

[0067] In a multi-sensor environment, each sensor may have different characteristics and there are complex nonlinear correlations between them. For example, increasing the fuel flow rate may cause changes in combustion temperature and pressure, which will affect the response of the automatic control system. These effects may not be immediately reflected in the data. Therefore, ideally, each sensor should be modeled flexibly to capture its potential "factor" changes in multiple dimensions. The present invention aims to better express the spatiotemporal correlation and dynamic changes between sensor data by constructing an adaptive graph structure spatiotemporal neural network.

[0068] Reference Figure 2 to Figure 4 , the specific steps of constructing an adaptive graph structured spatiotemporal network are described in detail:

[0069] 1. Data Collection and Preparation

[0070] This paper injects predefined faults into the JT9D aircraft engine model multiple times to generate experimental data containing 8 virtual sensor data to verify the effectiveness of the present invention. The collected sensor data is cleaned by interpolation method to remove noise, missing values ​​and outliers to ensure the quality of the data. The data of each sensor is standardized or normalized to ensure that the dimensions of different sensors are consistent, which is convenient for subsequent model training. The cleaned and regularized data are divided into a training set and a test set. The training set is used for model training, and the test set is used for model verification.

[0071] 2. Sliding Window Construction

[0072] A sliding window with a length of 500 is constructed to extract local features of time series data. The data in each sliding window forms a 500×8 matrix, where 8 is the number of virtual sensors.

[0073] 3. Time series feature extraction

[0074] The LSTM (Long Short-Term Memory) model is used to extract time features from time series data. The input data is a matrix consisting of 500 time steps of sensor data, which is processed by the LSTM layer into a 500×100 matrix, where 100 is the output dimension of the LSTM. The LSTM network generates a matrix containing time series features through hidden states, which provides support for subsequent fault diagnosis.

[0075] 4. Temporal Embedding Generation

[0076] From the hidden layer feature matrix generated by LSTM, select the hidden layer information of the most recent 10 time steps. Process it into a 10×100 matrix, which means that the features within 10 time steps are retained, and the dimension of the features of each time step is 100. The 100-dimensional hidden layer information is mapped to a lower-dimensional feature space through the fully connected layer to obtain a low-dimensional feature vector. In this implementation, the compressed dimension is set to 10, which also means that the number of nodes in the subsequent graph neural network is 10. Then, through the transposition operation, a matrix with a dimension of 10×10 is obtained, indicating that the initial time embedding with a dimension of 10 is generated for each of the 10 nodes, in preparation for the initial embedding of nodes in the subsequent graph neural network.

[0077] 5. Spatial relationship correction

[0078] Generate a random embedding matrix for each node, and set the dimension of each feature vector in the matrix to 15. As we know from the above, the number of nodes in the graph network is 10, so the matrix dimension is 10×15, which is used to supplement the node feature space and correct the spatial relationship between nodes.

[0079] The node embedding of the time feature is concatenated with the random embedding matrix to obtain a node feature matrix V containing time and space information with a dimension of 10×(10+15).

[0080] 6. Dynamic graph construction

[0081] By calculating the cosine similarity between node embedding vectors, we select the top K nodes with similarity, and consider that they have similar performance in the gas path system, so they can be connected. Based on the similarity between nodes, we construct an adjacency matrix A to represent the directed relationship between nodes, where the element A of the adjacency matrix is ij 1 indicates that there is a connection relationship between node i and node j, and the element A of the adjacency matrix ij A value of 0 indicates that there is no connection between node i and node j.

[0082] 7. Graph Information Fusion

[0083] According to the adjacency matrix A, the graph convolutional network (GCN) is used to aggregate and update the node information. The aggregate representation Z of node i i It is obtained by calculating the weighted average of the information of its neighboring nodes. In the graph neural network, the attention mechanism is introduced to further adjust the contribution of each neighboring node to the feature aggregation of node i. By normalizing the attention coefficient α ij To control the weight of neighbor node information.

[0084] 8. Fault diagnosis and prediction

[0085] After concatenating and aggregating the information of all nodes, a global feature vector is obtained and input into a multi-layer perceptron (MLP) for classification. The model outputs a vector containing the probability of each fault type, which is trained and optimized through the cross entropy loss function to finally complete fault diagnosis and prediction.

[0086] To verify the performance of the proposed method, it is compared with the following methods: LSTM model that only extracts temporal features, 1D CNN model, 2D CNN model that only extracts two-dimensional spatial features, and LSTM_2D CNN model that combines spatiotemporal features. The distribution of features extracted by different methods after PCA dimensionality reduction is shown in Figure 2. Figure 6 As shown in the figure, it can be seen that the distribution of the features extracted by 1DCNN in space is relatively clustered, but the distribution of some categories still overlaps. Compared with 1DCNN, the distribution of LSTM feature points is more dispersed, and the category distribution is also relatively separate. 2DCNN is similar to LSTM, with a relatively dispersed distribution, but some categories are more clearly distinguished. The distribution of LSTM_GCN feature points combined with spatiotemporal features is obviously more regular, and the category distinction is better. The model proposed in the present invention has a good separation of each category when the random embedding dimension is 0. However, the model proposed in the present invention performs best when the stacked embedding dimension is 15, and the category separation effect is clearer and there is almost no overlap. The fault diagnosis prediction accuracy of different models is shown in Table 1. It can be seen that the proposed model achieves the highest accuracy (99.351%) when the embedding dimension is 15, which is far higher than other models, especially significantly higher than 1DCNN (97.616%).

[0087] In general, this paper proposes an adaptive graph-structured spatiotemporal neural network that solves the nonlinear association problem in sensor data by flexibly representing each sensor and capturing the complex temporal and spatial relationships behind it. The main advantages of this method include:

[0088] 1) Comprehensive extraction of spatiotemporal features: LSTM is used to capture the time series dependency and dynamic changes of sensor data, and the graph attention mechanism is used to extract the spatial correlation between sensors, ensuring the full utilization of spatiotemporal features.

[0089] 2) Adaptive graph structure learning: The graph structure between sensors is dynamically constructed through embedding and similarity measurement, which can flexibly represent the dependency relationship of sensors and allow different types of sensors to play different roles in the graph structure, thereby improving the flexibility and robustness of the model.

[0090] 3) Random embedding and spatial relationship correction: By introducing random embedding vectors to supplement the time series features, the complex relationship between sensors can be better captured and the spatial relationship correction can be achieved, especially when the embedding dimension is 15, showing the best classification effect and stability.

[0091] 4) High accuracy and stability: Experimental results show that this method performs better than traditional models in sensor fault diagnosis tasks. When the embedding dimension is 15, the accuracy of the model reaches 99.351%, significantly surpassing traditional models such as 1D CNN and LSTM, showing extremely high accuracy and stability of model training.

[0092] These advantages make this method have wide application potential in complex multi-sensor systems, especially for tasks that need to deal with both temporal and spatial dependencies, such as aircraft engine fault diagnosis.

[0093]

[0094] Table 1.

Claims

1. A spatiotemporal network model for aircraft engine gas path fault diagnosis with an adaptive graph structure advantage, characterized in that: The following steps are involved:

101. Data preparation; 102. Sliding window construction; 103. Temporal feature extraction; 104.Time embedding generation; 105.Spatial relationship correction; 106. Dynamic graph construction; 107.Graph information fusion; 108.Fault diagnosis and prediction.

2. The aero-engine gas path fault diagnosis spatiotemporal network model with the adaptive graph structure advantage according to claim 1 is characterized in that: Step 101 data preparation is as follows: Collect data from N sensors, clean the data to remove noise, missing values ​​or outliers, and regularize the data to ensure that the dimensions of data from different sensors are consistent; Then, the processed data is divided into a training set and a test set, where the training data sequence is S train , the corresponding acquisition time length is T train ; The test data sequence is S test , the corresponding acquisition time length is T test .

3. The aero-engine gas path fault diagnosis spatiotemporal network model with adaptive graph structure advantage according to claim 1 is characterized in that: Step 102 sliding window construction is specifically as follows: By constructing a sliding window of length w, the data finally used for model training and testing is generated; the format of the training data is: S1={S train (1:w),…,S train (T train-w+1 :T train )} The data segment S1(t) used at the tth training time step is S train (t:t+w), this data segment is a w×N matrix, which is the measurement value of N sensors in a window of length w; The test data is processed in the same way as the training data, and the format of the obtained test data is: S2={S test (1:w),…,S test (T test-w+1 :T test )} The data segment S2(t) used at the tth test time step is S test (t:t+w), the data segment is a w×N matrix, which is the measurement value of N sensors in a window of length w.

4. The aero-engine gas path fault diagnosis spatiotemporal network model with adaptive graph structure advantage according to claim 1 is characterized in that: Step 103 of extracting time series features is as follows: The input data S1(t) of the t-th training time step contains information spanning w time steps, and the output data of the t-th time step is Represents the final fault classification result; first, the LSTM model is used to process the input data S1(t), capture the features within the time range t to t+w, and generate a hidden state matrix H containing time series features t , provide follow-up support for the output of the model; H t =LSTM(S1(t)) Among them, H t It is a w×M matrix, indicating that within a time window of length w, the LSTM layer generates a feature embedding of length M for each time step.

5. The aero-engine gas path fault diagnosis spatiotemporal network model with adaptive graph structure advantage according to claim 1 is characterized in that: Step 104, time embedding generation is specifically as follows: For the hidden state matrix H containing temporal features t Crop and only keep the hidden layer information of the most recent K time steps to obtain the processed hidden layer information H' t ; H’ t =H t [-K:,∶] Among them, H' t is a K×M matrix, representing H' t Contains the hidden layer feature information of the last K time steps, and the hidden information dimension of each time step is M; Subsequently, the hidden layer feature information is mapped from the high-dimensional feature space with dimension M to the low-dimensional feature space with dimension N' through the fully connected layer; then, the mapping result is transposed to generate the initial embedding E of the graph neural network node in the subsequent generation. t Be prepared; First, the processed hidden layer information is further integrated through the fully connected layer: X=(WH’ t +b) T Where X is a matrix of N'×K, indicating that there are N' nodes in total, the dimension of the feature vector of each node is K, W is the weight matrix of the fully connected layer, and b is the bias vector of the fully connected layer; Finally, the training process is further stabilized by layer normalization to generate the initial embedding E of the graph neural network nodes with N' nodes. t ; E t =LayerNorm(X) Among them, E t is an N'×K matrix, which represents the initial node embedding matrix that integrates time features. N' is the number of nodes in the subsequent graph neural network input graph, where the dimension of the feature vector of each node is K.

6. The aero-engine gas path fault diagnosis spatiotemporal network model with adaptive graph structure advantage according to claim 1 is characterized in that: Step 105 spatial relationship correction is specifically as follows: Generate a random embedding matrix E and use it as a random embedding vector to supplement node features; E=randn(N',D Among them, E is the generated random embedding matrix, N' is the number of nodes of the subsequent graph neural network, and D is the set random embedding dimension; Initially embed the nodes that incorporate time features into E t Concatenate with the generated random embedding matrix E to correct the initial node embedding E t The incorrect spatial relationships that may exist in the node initial embedding E t There are potentially more hidden patterns in ; V=E t +E Among them, V is the node embedding generated after spatial relationship correction, which is an N'×(K+D) matrix.

7. The aero-engine gas path fault diagnosis spatiotemporal network model with adaptive graph structure advantage according to claim 1 is characterized in that: Step 106 dynamic graph construction is as follows: The above steps generate a feature embedding of length K+D for each of the N' nodes; the feature embedding v of the i-th node i and the feature embedding v of the jth node j The similarity between them is used to indicate the similarity between the behaviors of node i and node j. The preliminary graph structure is constructed by calculating the cosine similarity between the feature embeddings of the nodes and combining the TopK method. Use a directed graph to build the relationship between nodes, and use the adjacency matrix A to represent the directed graph, where A ij Indicates the existence of a directed edge from node i to node j. In the absence of prior information, the candidate relationships of node i are all nodes except itself. Each node i has a set of candidate relationships, that is, the nodes that the node i may depend on: In order to select the dependency relationship of node i among these candidate relationships, we calculate the embedding vector v of node i. i With its candidate nodes The embedding vector v j The similarity between them is calculated, and the TopK nodes with the highest similarity are selected as the neighbor nodes of the node. The calculation formula is as follows: First calculate e ji , that is, the embedding vector v of node i i Relationship with Candidate The normalized dot product between the embedding vectors of all nodes in ; then, the top k normalized dot products are selected; TopK represents the index of the top k values ​​in its input, which is the normalized dot product; the value of k is selected by the user according to the desired sparsity level.

8. The aero-engine gas path fault diagnosis spatiotemporal network model with adaptive graph structure advantage according to claim 1 is characterized in that: Step 107: The graph information fusion is specifically as follows: The embedding vector of the i-th node is v i , the embedding vector of the jth node is v j , these vectors represent different behaviors of different nodes; according to the obtained adjacency matrix A, based on the attention mechanism, the information between nodes is further screened and fused, and the aggregate representation z of node i is calculated as follows i : Among them, z i represents the feature representation of the i-th node after aggregation, α ii is the attention coefficient of node i to itself, α ij is the attention coefficient of node i to neighbor node j, W att is a trainable weight matrix used to linearly transform feature vectors, N(i) is the set of neighbor nodes of node i; ReLU represents the linear rectification function; α ij is the normalized attention coefficient, which is solved as follows: Among them, π(i,j) is the attention coefficient between node i and node j; v i is the embedding vector of the i-th node; v j is the embedding vector of the jth node; represents concatenation; a is the learning coefficient vector of the attention mechanism; LeakyReLU is used as the nonlinear activation function to calculate the attention coefficient; α ij is the attention coefficient normalized by the softmax function, and N(i) is the set of neighbor nodes of node i.

9. The aero-engine gas path fault diagnosis spatiotemporal network model with adaptive graph structure advantage according to claim 1 is characterized in that: Step 108 fault diagnosis and prediction are as follows: The feature vector after the information of each node is aggregated is z i , concatenate the features of N' nodes and input them into the multi-layer perceptron for classification; Among them, f represents the fully connected layer; The output classification result; the model outputs a C-dimensional vector Represents the predicted probability of C-1 types of faults; the model is optimized using cross entropy loss for training.

Images