Aircraft system fault diagnosis method in high-noise environment based on node sensitivity graph network

Abstract

The invention discloses an aircraft system fault diagnosis method in a high-noise environment based on a node sensitivity graph network, and the method comprises the steps: enabling collected aircraft system fault data to be mapped to a frequency domain through fast Fourier transform, and dividing the data into a training set and a test set; an initialized adjacency matrix is adaptively constructed based on the sensitivity between the nodes; performing node information updating through a learnable polynomial to form a complete graph convolution layer; the node information serves as input to pass through the graph convolution layer again, and data features are obtained; iterative training is carried out, and the loss weight is dynamically adjusted; explaining the interpretability of the proposed method by visualizing the weight of each parameter; and inputting the test set data added with different noise levels into the trained graph network, and outputting a fault diagnosis result. According to the method, early fault signals of the aircraft system in a high-noise environment can be effectively processed, fault diagnosis with robustness and high accuracy is realized, and safe and stable operation of the aircraft system is guaranteed.

Description

Technical Field

[0001] The invention belongs to the field of aviation, relates to aircraft system fault diagnosis technology, and specifically to an aircraft system fault diagnosis method in a high noise environment based on a node sensitivity graph network. Background Art

[0002] As a core component of the modern industrial field, early fault diagnosis of aircraft systems is crucial to preventing major failures and ensuring stable operation of the system.

[0003] Deep learning has been widely used in aircraft system fault diagnosis due to its powerful feature representation ability, but this type of method is highly dependent on the quality of training data. In practical applications, data is contaminated by noise and its potential structural information is disturbed, causing the model to learn non-target feature patterns, increasing the risk of model overfitting and weakening its generalization ability.

[0004] Graph Neural Network (GNN) can effectively capture the complex spatiotemporal relationship between sensor data based on its information propagation mechanism between nodes, thereby improving the accuracy and reliability of fault diagnosis. However, existing GNN methods do not fully consider the sensitivity and importance between sensors, lack quantitative evaluation of model performance in high-noise environments, and lack model interpretability. Summary of the invention

[0005] Purpose of the invention: In order to overcome the deficiencies in the prior art, a method for diagnosing aircraft system faults in a high-noise environment based on a node sensitivity graph network is provided. Starting from the graph network structure, the problem that the existing aircraft system fault diagnosis technology is unable to improve the model generalization ability and explainability transparency under different noise levels is solved. The method can effectively process the early fault signals of the aircraft system in a high-noise environment, realize robust and high-accuracy fault diagnosis, and thus provide guarantee for the safe and stable operation of the aircraft system.

[0006] Technical solution: To achieve the above-mentioned purpose, the present invention provides an aircraft system fault diagnosis method in a high noise environment based on a node sensitivity graph network, comprising the following steps:

[0007] Step A: Map the collected aircraft system fault data to the frequency domain through fast Fourier transform and divide it into a training set and a test set;

[0008] Step B: Based on the training set data divided in step A, adaptively construct an initial adjacency matrix based on the sensitivity between each node;

[0009] Step C: Based on the adjacency matrix and original data established in step B, the node information is updated through a learnable polynomial to form a complete graph convolution layer;

[0010] Step D: The node information obtained in step C is used as input and passed through the graph convolution layer again, and then through the cycle of the encoding layer and the decoding layer to obtain the data features;

[0011] Step E: Based on the data features obtained in step D, iterative training is performed through the cosine similarity loss function, and the loss weight is dynamically adjusted based on its convergence speed;

[0012] Step F: Based on step E, the trained graph network parameters are obtained, and the interpretability of the proposed method is explained by visualizing the weights of each parameter;

[0013] Step G: Under the interpretability description of step F, the test set data with different noise levels added is input into the trained graph network, and the fault diagnosis results are output.

[0014] Further, the step A comprises:

[0015] Step A1: traverse the N sensor channel data corresponding to each fault, and perform fast Fourier transform on each channel data;

[0016] Step A2: According to the symmetry of fast Fourier transform, each channel takes the first half of the data;

[0017] Step A3: Divide the data. Based on the problem of data imbalance in actual working conditions, the training set only contains normal samples, and the test set contains normal and fault samples.

[0018] Furthermore, the construction method of initializing the adjacency matrix in step B includes:

[0019] Step B1: traverse each node and combine each node in pairs to form N×N pairs of combinations;

[0020] Step B2: If two nodes in the combination are the same, the corresponding position in the adjacency matrix is ​​0, and the two nodes are not connected;

[0021] Step B3: If the two nodes in the combination are different, calculate the Euclidean distance between the two nodes i and j and cosine distance X i and X j are the feature vectors corresponding to nodes i and j respectively, F is the dimension of the vector, X i,f and X j,f They are vector X i and X j The component in the fth dimension;

[0022] Step B4: Weighted Euclidean distance and cosine distance as the distance measure between nodes:

[0023] Step B5: To facilitate uniform measurement, the distances between all nodes are normalized: So that they all fall into the interval [0,1], where D min and D max are the minimum and maximum distances between nodes respectively;

[0024] Step B6: Set a distance threshold e, 0<e<1, and determine the sensitivity between two nodes according to the threshold;

[0025] Step B7: If the distance between the two nodes is greater than the threshold e, the sensitivity between the two nodes is considered low, the corresponding position in the adjacency matrix is ​​0, and the two nodes are not connected;

[0026] Step B8: If the distance between the two nodes is less than or equal to the threshold value e, the sensitivity between the two nodes is considered high, the corresponding position in the adjacency matrix is ​​1, and the two nodes are connected;

[0027] Step B9: Get the initialized adjacency matrix A.

[0028] Furthermore, the step C specifically includes:

[0029] Step C1: Obtain the normalized Laplace matrix L of the graph through the adjacency matrix A and the degree matrix D. -1 / 2 AD -1 / 2 ;

[0030] Step C2: Input the original data and implement K-order ARMA filtering in combination with L: Where I is the identity matrix, k = 1, 2..., K, p k ,q k is a learnable parameter;

[0031] Step C3: Perform the weighted replacement inversion operation T times through ARMA iteration, and the output of each layer is: Where R and S are learnable matrix parameters, t=0,1,..., That is, the initial node characteristics;

[0032] Step C4: Each layer output is activated through nonlinearity: Where σ(·) is a nonlinear activation function;

[0033] Step C5: At this time, the output characteristics of ARMA are:

[0034] Step C6: To prevent over-smoothing of message passing, message passing and aggregation are called at the kth order and tth layer; in the message passing process, for node v, its neighbor node u∈E(v), the message passing process is expressed as: Where m is the message of each node, M is the message transfer function, usually an affine transformation, A uvis the connection relationship between node v and its neighbor node u; by averaging and aggregating the messages of each node, we can get: Where W is the corresponding weight of each node message;

[0035] Step C7: Through dynamic weighting, the output of the k-th order t-th layer is: Where α is a learnable parameter;

[0036] Step C8: Therefore, the complete graph convolution output is:

[0037] Furthermore, the step D specifically includes:

[0038] Step D1: The output of the previous layer of graph convolution is passed through BatchNorm and Relu activation functions to obtain the output of the first layer of graph convolution encoder, which is used as the input of the second layer of graph convolution;

[0039] Step D2: By adjusting the input and output dimensions, the output of the second-layer graph convolution encoder is obtained, including the linear mapping of the second-layer graph convolution output and the first-layer graph convolution encoder output, and it is used as the third-layer graph convolution input;

[0040] Step D3: Repeat step D2 to obtain the output of the third-layer graph convolution encoder, the output of the first-layer graph convolution decoder, and the output of the second-layer graph convolution decoder in sequence;

[0041] Step D4: The output of the second-layer graph convolution decoder passes through BatchNorm and Relu activation functions as the input of the third-layer graph convolution;

[0042] Step D5: Get the output of the third-layer graph convolutional decoder, that is, the graph network output.

[0043] Furthermore, the step E specifically includes:

[0044] Step E1: Calculate the graph network reconstruction loss through the cosine similarity loss function: (c=1,2,…,C), where L c is the reconstruction loss corresponding to the c-th working condition data, X c is the input data corresponding to the cth working condition, X′ c The output result corresponding to the cth working condition;

[0045] Step E2: Dynamically adjust the loss weight based on the model convergence speed. First, calculate the training speed of the it-th iteration under the c-th working condition. The smaller the value, the faster the speed. When it is 0, is 1;

[0046] Step E3: Calculate the weight of the reconstruction loss corresponding to the it-th iteration under the c-th condition according to the training speed , based on Dynamic optimization;

[0047] Step E4: Calculate the total reconstruction loss for the it-th iteration

[0048] Step E5: The iteration ends and the network training is completed.

[0049] Further, the step F comprises:

[0050] Step F1: call graph network related parameter weights;

[0051] Step F2: Visualize the weights and explain the interpretability of the proposed method through attention distribution.

[0052] Furthermore, the step G specifically includes:

[0053] Step G1: Import the training set data into the trained graph network, calculate all loss values, and set the maximum loss value as the threshold for fault data detection;

[0054] Step G2: Add two types of noise to the test set data: Random Gaussian:X m =X m +ε and RandomScale:X m =X m ×δ, where X m is the mth sampling point of the sample; ε is the randomly added Gaussian noise, ε~N(0,μ); δ is the proportional factor, δ~N(1,μ), μ is the variance; in each sample, the probability of the two types of noise being randomly added is 50%;

[0055] Step G3: Control the noise level in the graph network by adjusting μ, μ ranges from 0 to 1, and the larger μ is, the higher the corresponding noise level is;

[0056] Step G4: Verify the performance of the graph network under different noise levels, obtain the calculation loss, and determine whether the data is normal or faulty based on the detection threshold obtained previously;

[0057] Step G5: Output the overall evaluation indicators, including accuracy, precision, F1-score, and the area under the ROC curve (Area Under Curve, AUC).

[0058] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0059] (1) The method of the present invention uses the distance between nodes to quantify the sensitivity between nodes, adaptively constructs and initializes the adjacency matrix, reduces unnecessary node connections, reduces redundancy and noise interference, and ensures the accuracy and robustness of diagnosis;

[0060] (2) The method of the present invention iteratively trains the graph network through the cosine similarity loss function, and dynamically adjusts the loss weight based on the convergence speed, balances the training stages, speeds up the convergence speed, and ensures the stability of the network;

[0061] (3) The method of the present invention enhances the network interpretability by visualizing the network parameter weights, making the internal working principle of the network transparent and improving the network trust;

[0062] (4) The method of the present invention uses evaluation indicators such as accuracy, precision, F1-score and AUC to measure the generalization performance of the proposed method under different noise levels. Compared with the results without using this method, the method of the present invention can improve the noise resistance and generalization ability of the model and achieve fault diagnosis of aircraft systems with higher accuracy and precision. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is an overall flow chart of the method of the present invention;

[0064] Figure 2 A schematic diagram of the initialization adjacency matrix constructed by the method of the present invention;

[0065] Figure 3 A schematic diagram of implementing node information update using a learnable polynomial in the present invention;

[0066] Figure 4 This is the interpretability analysis result diagram in the present invention;

[0067] Figure 5 This is a comparison chart of reconstruction loss results under different noise levels in the present invention. DETAILED DESCRIPTION

[0068] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0069] like Figure 1 As shown, this embodiment provides an aircraft system fault diagnosis method in a high noise environment based on a node sensitivity graph network, comprising the following steps:

[0070] Step A: Map the collected aircraft system fault data to the frequency domain through fast Fourier transform and divide it into a training set and a test set, including the following steps:

[0071] (1) Traverse the 6 sensor channel data corresponding to each fault and perform fast Fourier transform on each channel data;

[0072] (2) According to the symmetry of fast Fourier transform, each channel takes the first half of the data;

[0073] (3) The data was divided and collected under three different working conditions: 200 Hz, 0 N; 300 Hz, 1000 N; 400 Hz, 1400 N. Based on the problem of data imbalance under actual working conditions, the training set only contains normal samples. The corresponding imbalance ratio of the three working conditions is 200:100:10, that is, the total number of training set samples is 310. The test set contains normal and fault samples. Under each working condition, there are 200 normal samples and 300 fault samples, that is, the total number of test set samples is 1500.

[0074] Step B: Under the training set data divided in step A, adaptively construct the initial adjacency matrix based on the sensitivity between each node, such as Figure 2 The figure shows a schematic diagram of initializing the adjacency matrix, which includes the following steps:

[0075] (1) Traverse each node and combine each node in pairs to form 36 pairs of combinations;

[0076] (2) If two nodes in a combination are the same, the corresponding position in the adjacency matrix is ​​0, and the two nodes are not connected;

[0077] (3) If the two nodes in the combination are different, calculate the Euclidean distance between the two nodes i and j and cosine distance X i and X j are the feature vectors corresponding to nodes i and j respectively, F is the dimension of the vector, X i,f and X j,f They are vector X i and X j The component in the fth dimension;

[0078] (4) Weighted Euclidean distance and cosine distance as the measure of distance between nodes:

[0079] (5) To facilitate unified measurement, the distances between all nodes are standardized: So that they all fall into the interval [0,1], where D min and D max are the minimum and maximum distances between nodes respectively;

[0080] (6) Set a distance threshold e = 0.5 and determine the sensitivity between two nodes based on this threshold;

[0081] (7) If the distance between two nodes is greater than 0.5, the sensitivity between the two nodes is considered low, the corresponding position in the adjacency matrix is ​​0, and the two nodes are not connected;

[0082] (8) If the distance between two nodes is less than or equal to 0.5, the sensitivity between the two nodes is considered high, the corresponding position in the adjacency matrix is ​​1, and the two nodes are connected;

[0083] (9) Get the initialized adjacency matrix A.

[0084] Step C: Based on the adjacency matrix A established in step B and the original data, the node information is updated through a learnable polynomial to form a complete graph convolution layer, such as Figure 3 The figure shows a schematic diagram of node information update using a learnable polynomial, which includes the following steps:

[0085] (1) The normalized Laplace matrix L of the graph is obtained by using the adjacency matrix A and the degree matrix D. -1 / 2 AD -1 / 2 ;

[0086] (2) Input the original data and combine L to implement K-order ARMA filtering: Where I is the identity matrix, k = 1, 2..., K, p k ,q k is a learnable parameter;

[0087] (3) Through ARMA iterative weighted substitution inversion operation T times, the output of each layer is: Where R and S are learnable matrix parameters, t=0,1,..., That is, the initial node characteristics;

[0088] (4) Each layer output is activated through nonlinearity: Where σ(·) is a nonlinear activation function;

[0089] (5) At this time, the output characteristics of ARMA are:

[0090] (6) To prevent over-smoothing of message passing, message passing and aggregation are called at the kth order and tth layer. In the message passing process, for node v, its neighbor node u∈E(v), the message passing process can be expressed as: Where m is the message of each node, M is the message transfer function, usually an affine transformation, A uv is the connection relationship between node v and its neighbor node u; by averaging and aggregating the messages of each node, we can get: Where W is the corresponding weight of each node message;

[0091] (7) Through dynamic weighting, the output of the k-th order and t-th layer is: Where α is a learnable parameter;

[0092] (8) Then the complete graph convolution output is: Taking K=T=1, the graph convolution output formula is: in

[0093] Step D: The node information obtained in step C is used as input and passed through the graph convolution layer again for six cycles, the first three times corresponding to the encoding layer and the last three times corresponding to the decoding layer to obtain data features; including the following steps:

[0094] (1) The output of the previous layer of graph convolution passes through the BatchNorm and Relu activation functions to obtain the output of the first layer of graph convolution encoder, which is used as the input of the second layer of graph convolution;

[0095] (2) By adjusting the input and output dimensions, the output of the second-layer graph convolution encoder is obtained, including the linear mapping of the second-layer graph convolution output and the first-layer graph convolution encoder output, and it is used as the third-layer graph convolution input;

[0096] (3) Repeat the above step (2) to obtain the output of the third-layer graph convolution encoder, the output of the first-layer graph convolution decoder, and the output of the second-layer graph convolution decoder in sequence;

[0097] (4) The output of the second-layer graph convolution decoder passes through BatchNorm and Relu activation functions and serves as the input of the third-layer graph convolution;

[0098] (5) Obtain the output of the third-layer graph convolutional decoder, that is, the graph network output.

[0099] Step E: Based on the data features obtained in step D, iterative training is performed through the cosine similarity loss function, and the loss weight is dynamically adjusted based on its convergence speed, including the following steps:

[0100] (1) Calculate the graph network reconstruction loss through the cosine similarity loss function: Where L c is the reconstruction loss corresponding to the c-th working condition data, X c is the input data corresponding to the cth working condition, X c ′ is the output result corresponding to the cth working condition;

[0101] (2) Dynamically adjust the loss weight based on the model convergence speed. First, calculate the training speed of the it-th iteration under the c-th working condition. The smaller the value, the faster the speed. When it is 0, r c 0 is 1;

[0102] (3) Calculate the weight of the reconstruction loss corresponding to the it-th iteration under the c-th working condition according to the training speed Its basis Dynamic optimization;

[0103] (4) Calculate the total reconstruction loss of the it-th iteration

[0104] (5) The iteration ends and the network training is completed.

[0105] Step F: Based on the trained graph network parameters obtained in step E, the interpretability of the method of the present invention is explained by visualizing the parameter weights to show the dynamic changes of its attention weights, deepen the understanding of the network working mechanism, and at the same time show the compactness advantage of the method of the present invention in attention distribution through comparison.

[0106] "Explainability" mainly lies in its ability to enhance understanding of the decision-making process within the network, thereby improving the transparency and credibility of the model. By revealing the working principles of the network, potential biases in the data can be more effectively identified and prevented, ensuring the fairness and reliability of the network's prediction results.

[0107] like Figure 4 The following figure shows the result of interpretability analysis, which includes the following steps:

[0108] (1) Call the weights of the first and last layers of graph convolution in the graph network;

[0109] (2) In order to conduct subsequent analysis of various performance indicators of the model, two comparison methods were selected: FGDAE (published in Reliability Engineering and System Safety in 2023) and MAE (published in International Conference on Computer Vision in 2019);

[0110] (3) Since MAE is a codec structure, FGDAE and the proposed method are both multi-layer graph convolution structures, only FGDAE and the proposed method are compared in the weight visualization part. The attention distribution of the first and last layers of graph convolution of the proposed method and the attention distribution of the corresponding layers of FGDAE are visualized (e.g. Figure 4 As shown), and analyze it to reveal the interpretability of the method of the present invention;

[0111] (4) From Figure 4It can be seen that in the first layer, the weight distribution of the proposed method is between -0.6 and 1.0, with a range of 1.6, showing a more concentrated attention distribution; while the weight distribution of FGDAE is between -1.5 and 1.0, with a range of 2.5, which is relatively dispersed. This comparison shows that in the first layer, the proposed method can better focus on key features and reduce unnecessary node connections and noise interference.

[0112] As the number of network layers increases, the weight distribution of the method of the present invention is further concentrated between -0.4 and 0.8 in the last layer, and the range of variation is reduced to 1.2, indicating that the model gradually focuses on some significant features in the process of transmitting information layer by layer, making the weight distribution more compact and concentrated. In contrast, the weight distribution of FGDAE in the last layer is between -1.0 and 1.5, and the range of variation is still 2.5, which has not decreased.

[0113] From this changing trend, it can be seen that the method of the present invention can better maintain the stability of the weight distribution and reduce unnecessary node connections and data redundancy in the process of transmitting information layer by layer, thereby improving the robustness and accuracy of the model. In comparison with FGDAE, this advantage of the method of the present invention is further highlighted. Specifically, the method of the present invention can adaptively initialize the adjacency matrix based on the differences in sensitivity between sensors, thereby more effectively reducing unnecessary node connections and data redundancy and improving the fault diagnosis capability of the model. Therefore, in a high-noise environment, the method of the present invention can demonstrate superior fault diagnosis capabilities, providing a strong guarantee for the reliable operation of aircraft systems;

[0114] Step G: Based on the visualization and interpretability analysis of the graph network parameter weights in step F, the test set data with different noise levels is input into the trained graph network to fully verify the robustness and accuracy of the method. This step aims to further confirm the stability and accuracy of the method of the present invention in different noise environments through actual data verification and performance evaluation, and includes the following steps:

[0115] (1) Import the training set data into the trained graph network, calculate all loss values, and set the maximum loss value as the threshold for fault data detection;

[0116] (2) Add two types of noise to the test set data: RandomGaussian:X m =X m +ε and RandomScale:X m =X m ×δ, where X mis the mth sampling point of the sample; ε is the randomly added Gaussian noise, ε~N(0,μ); δ is the proportional factor, δ~N(1,μ), μ is the variance. In each sample, the probability of the two types of noise being randomly added is 50%;

[0117] (3) The noise level in the control graph network is controlled by adjusting μ, which can be between 0 and 1. The larger the μ, the higher the corresponding noise level.

[0118] (4) The performance of the graph network was verified under the conditions of noise levels of 0, 0.5 and 1, and the reconstruction loss values ​​of the proposed method, FGDAE and MAE under all test set samples were obtained, as shown in the figure. Figure 5 shown. Figure 5 The first 200 samples in the figure represent normal samples, and the last 300 samples are fault samples. The three colored curves correspond to three different working conditions. The black dotted line is the threshold, which is adaptively determined based on step E. Figure 5 It can be seen that the three comparison methods can distinguish normal samples from faulty samples, but the method of the present invention shows higher effectiveness and stability. For the method of the present invention, by comparing (a), (b), and (c), it can be seen that the noise interference causes its fluctuation to intensify, but the overall loss is kept below 7, and the threshold is stable at about 2. At the same time, the faulty samples are basically above the threshold, which can achieve a higher fault detection accuracy. The maximum loss value of FGDAE is nearly 18, and the maximum loss value of MAE has exceeded 20. When the noise level is 0.5 and 1, (b), (e), (h) and (c), (f), (i) are compared respectively. At this time, there are a large number of samples below the threshold in the faulty sample loss values ​​of FGDAE and MAE, and the misjudgment ratio is high. In addition, both are significantly affected by noise and fluctuate violently. In comparison, the proposed method is more robust. This result reflects the accuracy and robustness of the proposed method.

[0119] (5) Output the overall evaluation indicators, including accuracy, precision, F1-score and AUC, and compare the evaluation indicators of the proposed method with those of FGDAE and MAE. The results of each indicator are shown in Table 1, Table 2, Table 3 and Table 4;

[0120] Table 1 Comparison of the accuracy of each method under different noise levels

[0121] Proposed method FGDAE MAE μ=0 0.9735 0.9653 0.8675 μ=0.2 0.9427 0.8933 0.8144 μ=0.4 0.9160 0.8521 0.7764 μ=0.6 0.8953 0.8309 0.7513 μ=0.8 0.8600 0.8067 0.7327 μ=1 0.8440 0.7801 0.7167

[0122] Table 2 Comparison of the accuracy of each method under different noise levels

[0123]

[0124]

[0125] Table 3 Comparison of F1-score of various methods at different noise levels

[0126]

[0127] Table 4 AUC comparison of various methods under different noise levels

[0128]

[0129] (6) From the results in the above tables, it can be seen that the various indicators (accuracy, precision, F1-score, AUC) of the proposed method show stable and superior performance under different noise levels. As the noise level gradually increases, the performance indicators of the three methods all show a downward trend, which is caused by the intensification of noise interference. In contrast, the performance degradation of the proposed method at each noise level is relatively small, showing strong robustness.

[0130] Specifically, when μ = 0, that is, when there is no noise interference, all methods achieve the best performance, but the accuracy, precision, F1-score and AUC of the proposed method are 0.9735, 0.9835, 0.9782 and 0.9981, respectively, which are better than FGDAE and MAE. This is because under ideal conditions without noise, the proposed method strengthens the connection of sensors with higher sensitivity and weakens the connection of sensors with lower sensitivity based on the sensitivity difference between nodes, optimizes the message transmission path, and thus improves the accuracy of the proposed method.

[0131] As the noise level increases, the various indicators of FGDAE and MAE drop sharply. Although the method of the present invention is also affected by noise to a certain extent, its various performance indicators drop significantly slower and still remain at a relatively high level. For example, when μ = 1, the accuracy, precision, F1-score and AUC of the method of the present invention are maintained at 0.8440, 0.8219, 0.8812 and 0.8071, respectively, which are 8.1913%, 10.2038%, 5.9644%, 14.2230% higher than FGDAE, and 17.7620%, 21.3674%, 14.0860%, 24.8453% higher than MAE. This is because the method of the present invention uses the distance between nodes to quantify the sensitivity between nodes, reduces unnecessary node connections, thereby reducing the interference of noise on the graph network structure and enhancing the robustness of the proposed method.

[0132] In summary, the method of the present invention can resist noise interference, enhance the fault diagnosis capability of the aircraft system in a high noise environment, and thus achieve more accurate and reliable fault detection and early warning, providing a strong guarantee for the safe and stable operation of the aircraft system.

Claims

1. A method for diagnosing aircraft system faults in a high noise environment based on a node sensitivity graph network, characterized in that: The steps include: Step A: Map the collected aircraft system fault data to the frequency domain through fast Fourier transform and divide it into a training set and a test set; Step B: Based on the training set data divided in step A, adaptively construct an initial adjacency matrix based on the sensitivity between each node; Step C: Based on the adjacency matrix and original data established in step B, the node information is updated through a learnable polynomial to form a complete graph convolution layer; Step D: The node information obtained in step C is used as input and passed through the graph convolution layer again, and then through the cycle of the encoding layer and the decoding layer to obtain the data features; Step E: Based on the data features obtained in step D, iterative training is performed through the cosine similarity loss function, and the loss weight is dynamically adjusted based on its convergence speed; Step F: Based on step E, the trained graph network parameters are obtained, and the interpretability of the proposed method is explained by visualizing the weights of each parameter; Step G: Under the interpretability description of step F, the test set data with different noise levels added is input into the trained graph network, and the fault diagnosis results are output.

2. The method for diagnosing aircraft system faults in a high noise environment based on a node sensitivity graph network according to claim 1, characterized in that: The step A comprises: Step A1: traverse the N sensor channel data corresponding to each fault, and perform fast Fourier transform on each channel data; Step A2: According to the symmetry of fast Fourier transform, each channel takes the first half of the data; Step A3: Divide the data. Based on the problem of data imbalance in actual working conditions, the training set only contains normal samples, and the test set contains normal and fault samples.

3. The method for diagnosing aircraft system faults in a high noise environment based on a node sensitivity graph network according to claim 1, characterized in that: The construction method of initializing the adjacency matrix in step B includes: Step B1: traverse each node and combine each node in pairs to form N×N pairs of combinations; Step B2: If two nodes in the combination are the same, the corresponding position in the adjacency matrix is ​​0, and the two nodes are not connected; Step B3: If the two nodes in the combination are different, calculate the Euclidean distance between the two nodes i and j and cosine distance X i and X j are the feature vectors corresponding to nodes i and j respectively, F is the dimension of the vector, X i,f and X j,f They are vector X i and X j The component in the fth dimension; Step B4: Weighted Euclidean distance and cosine distance as the distance measure between nodes: Step B5: To facilitate uniform measurement, the distances between all nodes are normalized: So that they all fall into the interval [0,1], where D min and D max are the minimum and maximum distances between nodes respectively; Step B6: Set a distance threshold e, 0<e<1, and determine the sensitivity between two nodes according to the threshold; Step B7: If the distance between the two nodes is greater than the threshold e, the sensitivity between the two nodes is considered low, the corresponding position in the adjacency matrix is ​​0, and the two nodes are not connected; Step B8: If the distance between the two nodes is less than or equal to the threshold value e, the sensitivity between the two nodes is considered high, the corresponding position in the adjacency matrix is ​​1, and the two nodes are connected; Step B9: Get the initialized adjacency matrix A.

4. The method for diagnosing aircraft system faults in a high noise environment based on a node sensitivity graph network according to claim 3 is characterized in that: The step C specifically comprises: Step C1: Obtain the normalized Laplace matrix L of the graph through the adjacency matrix A and the degree matrix D. -1 / 2 AD -1 / 2 ; Step C2: Input the original data and implement K-order ARMA filtering in combination with L: Where I is the identity matrix, k×1,2...,K,p k ,q k is a learnable parameter; Step C3: Perform the weighted replacement inversion operation T times through ARMA iteration, and the output of each layer is: Where R and S are learnable matrix parameters, That is, the initial node characteristics; Step C4: Each layer output is activated through nonlinearity: Where σ(·) is a nonlinear activation function; Step C5: At this time, the output characteristics of ARMA are: Step C6: To prevent over-smoothing of message passing, message passing and aggregation are called at the kth order and tth layer; in the message passing process, for node v, its neighbor node u∈E(v), the message passing process is expressed as: Where m is the message of each node, M is the message transfer function, A uv is the connection relationship between node v and its neighbor node u; by averaging and aggregating the messages of each node, we can get: Where W is the corresponding weight of each node message; Step C7: Through dynamic weighting, the output of the k-th order t-th layer is: Where α is a learnable parameter; Step C8: Therefore, the complete graph convolution output is:

5. The method for diagnosing aircraft system faults in a high noise environment based on a node sensitivity graph network according to claim 1, characterized in that: The step D is cycled six times, the first three times correspond to the encoding layer, and the last three times correspond to the decoding layer, specifically including: Step D1: The output of the previous layer of graph convolution is passed through BatchNorm and Relu activation functions to obtain the output of the first layer of graph convolution encoder, which is used as the input of the second layer of graph convolution; Step D2: By adjusting the input and output dimensions, the output of the second-layer graph convolution encoder is obtained, including the linear mapping of the second-layer graph convolution output and the first-layer graph convolution encoder output, and it is used as the third-layer graph convolution input; Step D3: Repeat step D2 to obtain the output of the third-layer graph convolution encoder, the output of the first-layer graph convolution decoder, and the output of the second-layer graph convolution decoder in sequence; Step D4: The output of the second-layer graph convolution decoder passes through BatchNorm and Relu activation functions as the input of the third-layer graph convolution; Step D5: Get the output of the third-layer graph convolutional decoder, that is, the graph network output.

6. The method for diagnosing aircraft system faults in a high noise environment based on a node sensitivity graph network according to claim 1, characterized in that: The step E specifically comprises: Step E1: Calculate the graph network reconstruction loss through the cosine similarity loss function: Where L c is the reconstruction loss corresponding to the c-th working condition data, X c is the input data corresponding to the cth working condition, X c ′ is the output result corresponding to the cth working condition; Step E2: Dynamically adjust the loss weight based on the model convergence speed. First, calculate the training speed of the it-th iteration under the c-th working condition. The smaller the value, the faster the speed. When it is 0, is 1; Step E3: Calculate the weight of the reconstruction loss corresponding to the it-th iteration under the c-th condition according to the training speed Its basis Dynamic optimization; Step E4: Calculate the total reconstruction loss for the it-th iteration Step E5: The iteration ends and the network training is completed.

7. The method for diagnosing aircraft system faults in a high noise environment based on a node sensitivity graph network according to claim 1, characterized in that: The step F comprises: Step F1: call graph network related parameter weights; Step F2: Visualize the weights and explain the interpretability of the proposed method through attention distribution.

8. The method for diagnosing aircraft system faults in a high noise environment based on a node sensitivity graph network according to claim 1, characterized in that: The step G specifically comprises: Step G1: Import the training set data into the trained graph network, calculate all loss values, and set the maximum loss value as the threshold for fault data detection; Step G2: Add two types of noise to the test set data: RandomGaussian:X m ×X m +ε and RandomScale:X m ×X m ×δ, where X m is the mth sampling point of the sample; ε is the randomly added Gaussian noise, ε~N(0,μ); δ is the proportional factor, δ~N(1,μ), μ is the variance; in each sample, the probability of the two types of noise being randomly added is 50%; Step G3: Control the noise level in the graph network by adjusting μ, where μ ranges from 0 to 1; Step G4: Verify the performance of the graph network under different noise levels, obtain the calculation loss, and determine whether the data is normal or faulty based on the detection threshold obtained previously; Step G5: Output the overall evaluation indicators, including accuracy, precision, F1-score, and the area under the ROC curve and the coordinate axis.

Images