An aircraft system fault diagnosis method in a high-noise environment based on a node sensitivity graph network

Abstract

The application discloses a kind of based on node sensitivity graph network under high-noise environment aircraft system fault diagnosis method, comprising: the aircraft system fault data of acquisition is mapped to frequency domain by fast fourier transform, and it is divided into training set and test set;Based on the sensitivity between each node adaptively constructs initialization adjacency matrix;Node information is updated by learnable polynomial, and a complete graph convolution layer is formed;Node information is input again through graph convolution layer, and data characteristics are obtained;Iterative training is carried out, and loss weight is dynamically adjusted;The interpretability of the method is explained by visualizing each parameter weight;Test set data with different noise levels are input into the trained graph network, and fault diagnosis results are output.The application can effectively process early fault signals of aircraft system under high-noise environment, realize robustness and high-accuracy fault diagnosis, and further provide protection for safe and stable operation of aircraft system.

Description

TECHNICAL FIELD

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

[0002] As a core component in the field of modern industry, the early fault diagnosis of an aircraft system is crucial for preventing major faults and ensuring stable operation of the system.

[0003] Deep learning is widely used in aircraft system fault diagnosis due to its strong feature representation capability, but such methods highly depend on the quality of training data. In actual applications, data is contaminated by noise, and its potential structural information is disturbed, which leads the model to learn non-target feature patterns, increases the risk of model overfitting, and weakens its generalization capability.

[0004] Based on its node information propagation mechanism, a graph neural network (GNN) can effectively capture the complex spatio-temporal relationship between various sensor data, thereby improving the fault diagnosis precision and reliability. However, existing GNN methods do not fully consider the sensitivity and importance between various sensors, lack quantitative evaluation of the model performance in a high-noise environment, and have poor model interpretability. SUMMARY

[0005] The application aims to overcome the deficiencies in the prior art and provide an aircraft system fault diagnosis method in a high-noise environment based on a node sensitivity graph network, which takes the graph network structure as a starting point to solve the problems that existing aircraft system fault diagnosis technologies cannot improve the model generalization capability, interpretability transparency, and other issues under different noise levels, can effectively process early fault signals of an aircraft system in a high-noise environment, realize robust and high-accuracy fault diagnosis, and thus provide protection for the safe and stable operation of an aircraft system.

[0006] The technical scheme is as follows: to achieve the above-mentioned purpose, the application provides an aircraft system fault diagnosis method in a high-noise environment based on a node sensitivity graph network, which comprises the following steps:

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

[0008] Step B: based on the sensitivity between nodes, adaptively constructing an initialization adjacency matrix under the training set data divided in step A;

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

[0010] Step D: the node information obtained in step C is input again through the graph convolution layer, and data features are obtained through the cycle of the encoding layer and the decoding layer;

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

[0012] Step F: based on the trained graph network parameters obtained in step E, the interpretability of the method is explained through visualization of the parameter weights;

[0013] Step G: under the interpretability explanation in step F, test set data with different noise levels are input into the trained graph network, and 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, take the first half data of each channel;

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

[0018] Further, the construction method of the adjacency matrix in step B comprises:

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

[0020] Step B2: if the two nodes in the combination are the same, the corresponding position in the adjacency matrix is 0, and the two nodes have no connection;

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

[0022] Step B4: the weighted Euclidean distance and the cosine distance are used as the measurement of the distance between nodes:

[0023] Step B5: To facilitate the unified measure, the distance between all nodes is standardized by deviation: So that it falls entirely into the interval [0, 1], where D min And D max Are the minimum and maximum values of the distance between nodes, respectively;

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

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

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

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

[0028] Further, the step C specifically comprises:

[0029] Step C1: Obtain the normalized Laplacian 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 realize K-order ARMA filtering combined with L: Where I is the identity matrix, k = 1, 2,..., K, p k , q k Are learnable parameters;

[0031] Step C3: Replace the inverse operation by ARMA iteration weighting T times, and the output of each layer is: Where R and S are learnable matrix parameters, t = 0, 1,..., T, and X , i.e. the initial node feature;

[0032] Step C4: The output of each layer is passed through a nonlinear activation: Where σ(·) is a nonlinear activation function;

[0033] Step C5: At this time, the output feature of ARMA is:

[0034] Step C6: To prevent over-smoothing of message passing, message passing and aggregation are called at the kth order and the tth layer. During the message passing process, for a node v, its neighbor node u ∈ E(v), the message passing process is represented as: Where m is the message of each node, M is the message passing function, which is usually an affine transformation, A uvis the connection between node v and its neighbor node u; the message of each node can be obtained by averaging aggregation: where W is the weight corresponding to the message of each node;

[0035] Step C7: Through dynamic weighting, the output of the kth order and the tth layer is: where a is a learnable parameter;

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

[0037] Further, the step D specifically comprises:

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

[0039] Step D2: The input and output dimensions are adjusted to obtain the output of the second layer graph convolution encoder, including the linear mapping of the second layer graph convolution output and the first layer graph convolution encoder output, and the output of the second layer graph convolution encoder is taken as the input of the third layer graph convolution;

[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 turn;

[0041] Step D4: The output of the second layer graph convolution decoder is passed through the BatchNorm and Relu activation function to be taken as the input of the third layer graph convolution;

[0042] Step D5: The output of the third layer graph convolution decoder is obtained, that is, the output of the graph network.

[0043] Further, the step E specifically comprises:

[0044] Step E1: The reconstruction loss of the graph network is calculated by the cosine similarity loss function: (c = 1, 2, …, C), where L c is the reconstruction loss corresponding to the cth 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;

[0045] Step E2: The loss weight is dynamically adjusted based on the model convergence speed. First, the training speed of the ith iteration under the cth working condition is calculated The smaller the value, the faster the speed, when it is 0, is 1;

[0046] Step E3: The weight of the reconstruction loss corresponding to the ith iteration under the cth working condition is calculated according to the training speed , according to dynamic optimization;

[0047] Step E4: Calculate the total reconstruction loss of the it iteration

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

[0049] Further, the step F comprises:

[0050] Step F1: Call the graph network related parameter weight;

[0051] Step F2: Visualize the weight, and explain the method interpretability through the attention distribution.

[0052] Further, the step G specifically comprises:

[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 value of fault data detection;

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

[0055] Step G3: Control the noise level in the graph network by adjusting μ, μ is 0-1, the larger μ is, the higher the noise level is;

[0056] Step G4: Verify the performance of the graph network under different noise levels, obtain the calculation loss, and judge whether the data is normal or fault according to the detection threshold value obtained in the foregoing;

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

[0058] Beneficial effects: compared with the prior art, the present application has the following advantages:

[0059] (1) The method of the present application quantifies the sensitivity between nodes by the distance between nodes, adaptively constructs the initial adjacency matrix, reduces unnecessary node connection, reduces redundancy and noise interference, and ensures the accuracy and robustness of diagnosis;

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

[0061] (3) The method of the present application visualizes the network parameter weight, enhances the network interpretability, realizes the transparency of the network internal working principle, and improves the network trustworthiness;

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

[0063] Figure 1 is the overall flowchart of the method of the present application;

[0064] Figure 2 is the schematic diagram of the initial adjacency matrix constructed by the method of the present application;

[0065] Figure 3 is the schematic diagram of the node information update realized by the learnable polynomial in the present application;

[0066] Figure 4 is the explainability analysis result graph in the present application;

[0067] Figure 5 is the reconstruction loss result comparison graph under different noise levels in the present application. DETAILED DESCRIPTION

[0068] The present application will be further illustrated below in conjunction with the drawings and specific embodiments, and it should be understood that these embodiments are only used to illustrate the present application and not to limit the scope of the present application. After reading the present application, those skilled in the art can modify various equivalent forms of the present application, which all fall within the scope defined by the appended claims.

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

[0070] Step A: the aircraft system fault data collected is mapped to frequency domain by fast Fourier transform, and it is divided into training set and 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, take the first half data of each channel;

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

[0074] Step B: Under the training set data divided in step A, the initial adjacency matrix is adaptively constructed based on the sensitivity between nodes, as shown in the following formula: Figure 2 The initial adjacency matrix is shown in the following figure, including the following steps:

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

[0076] (2) If the two nodes in the 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 and the cosine distance between the two nodes i and j i and X j are the feature vectors corresponding to nodes i and j, F is the dimension of the vector, X i,f and X j,f are the components of the vector X i and X j in the f-th dimension;

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

[0079] (5) In order to unify the measurement, the deviation standardization is performed on the distance between all nodes: so that it falls into the interval [0, 1], where D min and D max are the minimum and maximum values of the distance between nodes, respectively;

[0080] (6) Set a distance threshold e = 0.5, and determine the sensitivity between two nodes according to the threshold;

[0081] (7) If the distance between two nodes is greater than 0.5, the sensitivity between the two nodes is 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 high, the corresponding position in the adjacency matrix is 1, and the two nodes are connected;

[0083] (9) Obtain 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, as shown in Figure 3 The schematic diagram of implementing node information update by a learnable polynomial is shown in the following figure:

[0085] (1) Obtain the normalized Laplacian matrix L of the graph through the adjacency matrix A and the degree matrix D: -1 / 2 AD -1 / 2 ;

[0086] (2) Input the original data, and realize K-order ARMA filtering combined with L: Where I is an identity matrix, k = 1, 2,..., K, p k , q k are learnable parameters;

[0087] (3) Replace the inverse operation T times through ARMA iteration weighting, and the output of each layer is: Where R and S are learnable matrix parameters, t = 0, 1,..., T , that is, the initial node feature;

[0088] (4) The output of each layer is passed through a nonlinear activation: Where σ(·) is a nonlinear activation function;

[0089] (5) The output feature of ARMA at this time is:

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

[0091] (7) Through dynamic weighting, the output of the kth order and the tth layer at this time is: wherein a is a learnable parameter;

[0092] (8) The complete graph convolution output is: Take K=T=1, and the graph convolution output formula is: wherein

[0093] Step D: The node information obtained in step C is input again through the graph convolution layer, and the cycle is six times, the first three times correspond to the encoding layer, and the last three times correspond to the decoding layer, to obtain the data feature; including the following steps:

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

[0095] (2) The input and output dimensions are adjusted to obtain the output of the second layer of graph convolution encoder, including the linear mapping of the output of the second layer of graph convolution and the output of the first layer of graph convolution encoder, and the output is taken as the input of the third layer of graph convolution;

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

[0097] (4) The output of the second layer of graph convolution decoder is passed through BatchNorm and Relu activation function to be taken as the input of the third layer of graph convolution;

[0098] (5) The output of the third layer of graph convolution decoder is obtained, which is the output of the graph network.

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

[0100] (1) The graph network reconstruction loss is calculated through the cosine similarity loss function: wherein L c is the reconstruction loss corresponding to the cth 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) The loss weight is dynamically adjusted based on the model convergence speed, and first the training speed of the ith iteration under the cth working condition is calculated The smaller the value is, the faster the speed is, and when it is 0, r c 0 is 1;

[0102] (3) According to the training speed, the weight of the reconstruction loss corresponding to the it iteration under the c working condition is calculated According to the dynamic optimization

[0103] (4) The total reconstruction loss of the it iteration is calculated

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

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

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

[0107] As Figure 4 The explainability analysis result graph is shown, including the following steps:

[0108] (1) The weights of the first layer and the last layer graph convolution of the graph network are called;

[0109] (2) For subsequent analysis of various performance indicators of the model, two comparison methods are 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 coding and decoding structure, FGDAE and the method proposed in the present application are both multi-layer graph convolution structures, so only FGDAE and the method proposed in the present application are compared in the weight visualization part. By visualizing the attention distribution of the first layer and the last layer graph convolution of the method proposed in the present application, and the corresponding layer attention distribution of FGDAE (as shown in Figure 4 ), and analyzing them, the explainability of the method proposed in the present application is revealed;

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

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

[0113] From this change trend, it can be seen that the method of the present application can better maintain the stability of the weight distribution in the process of layer-by-layer information transmission, reduce unnecessary node connections and data redundancy, thereby improving the robustness and accuracy of the model. In comparison with FGDAE, the advantage of the method of the present application is further highlighted. Specifically, the method of the present application can adaptively initialize the adjacency matrix based on the difference 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 application can exhibit superior fault diagnosis capability, providing strong support for the reliable operation of the aircraft system;

[0114] Step G: On the basis of the visualization processing and interpretability analysis of the graph network parameter weights in step F, test set data with different noise levels are input into the trained graph network to comprehensively verify the robustness and accuracy of the method. This step aims to further confirm the stability and accuracy of the method of the present application in different noise environments through actual data verification and performance evaluation, including 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 kinds 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 a random added Gaussian noise, ε ~ N(0, μ); δ is a proportional factor, δ ~ N(1, μ), and μ is a variance. In each sample, the random addition probability of the two types of noise is 50%;

[0117] (3) The noise level in the graph network is controlled by adjusting μ, μ can be 0-1, the larger μ is, the higher the corresponding noise level is;

[0118] (4) The performance of the graph network is verified under the conditions of noise levels of 0, 0.5 and 1 respectively, and the reconstruction loss values of the method, FGDAE and MAE under all test set samples are obtained, as shown in Figure 5 . Figure 5 The first 200 samples in the figure represent normal samples, and the last 300 samples are fault samples. The curves of three colors correspond to three different working conditions respectively, and the black dotted line is the threshold value, which is adaptively determined based on step E. It can be seen from Figure 5 , that the three comparison methods can distinguish normal samples and fault samples, but the method exhibits higher effectiveness and stability. For the method, it can be seen from comparisons (a), (b) and (c) that noise interference causes the fluctuation to intensify, but the overall loss remains below 7, and the threshold value is stable at about 2, while the fault samples are basically above the threshold value, which can achieve a high fault detection accuracy. The highest loss value of FGDAE is nearly 18, and the highest loss value of MAE is more than 20. When the noise level is 0.5 and 1, comparisons (b), (e), (h) and (c), (f), (i) are made respectively. At this time, the loss values of fault samples of FGDAE and MAE both have a large number of samples below the threshold value, the misjudgment proportion is high, and both are significantly affected by noise, with violent fluctuations. In comparison, the proposed method is more robust. This result reflects the accuracy and robustness of the proposed method.

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

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

[0121] The 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 precision of each method under different noise levels

[0123]

[0124]

[0125] Table 3 Comparison of F1-score of each method under different noise levels

[0126]

[0127] Table 4 AUC of each method under different noise levels

[0128]

[0129] (6) From the results of the above tables, it can be seen that the indexes (accuracy, precision, F1-score, AUC) of the method of the present application all show stable and superior performance under different noise levels. With the gradual increase of noise level, the performance indexes of the three methods all show a downward trend, which is due to the aggravation of noise interference. Compared with the above, the performance decline of the proposed method is relatively small under each noise level, showing strong robustness.

[0130] Specifically, when μ = 0, that is, there is no noise interference, each method achieves the best performance, but the accuracy, precision, F1-score and AUC of the method of the present application are 0.9735, 0.9835, 0.9782 and 0.9981 respectively, which are superior to FGDAE and MAE. This is because under the ideal condition of no noise, the method of the present application is based on the difference in sensitivity between nodes, strengthens the connection of sensors with high sensitivity, weakens the connection of sensors with low sensitivity, optimizes the message passing path, and thus improves the accuracy of the proposed method.

[0131] With the increase of noise level, the indexes of FGDAE and MAE decrease sharply, while the indexes of the method of the present application are affected by noise to some extent, but the decline speed is significantly slower, and still maintains at a relatively high level. For example, when μ = 1, the accuracy, precision, F1-score and AUC of the method of the present application are 0.8440, 0.8219, 0.8812 and 0.8071 respectively, which are 8.1913%, 10.2038%, 5.9644%, 14.2230% higher than FGDAE respectively, and 17.7620%, 21.3674%, 14.0860%, 24.8453% higher than MAE respectively. This is because the method of the present application quantifies the sensitivity between nodes by using the distance between nodes, reduces unnecessary node connection, thereby reducing the interference of noise on the structure of graph network, and enhances the robustness of the proposed method.

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

Claims

1. A method for diagnosing faults of an aircraft system in a high-noise environment based on a node-sensitive graph network, characterized in that, The method comprises the following steps: Step A: mapping the collected aircraft system fault data to the frequency domain through fast Fourier transform, and dividing the data into a training set and a test set; Step B: based on the sensitivity between nodes, an initial adjacency matrix is adaptively constructed under the training set data divided in step A; Step C: based on the adjacency matrix and the 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 input into the graph convolution layer again, and the data features are obtained through the cycle of the encoding layer and the decoding layer; Step E: based on the data features obtained in step D, iterative training is performed through a cosine similarity loss function, and the loss weight is dynamically adjusted based on the convergence speed; Step F: based on the trained graph network parameters obtained in step E, the interpretability of the method is explained through visualization of the parameter weights; Step G: under the interpretability explanation in step F, the test set data with different noise levels are input into the trained graph network, and the fault diagnosis results are output.

2. The method of claim 1, wherein, The step A comprises: Step A1: traversing the N sensor channel data corresponding to each fault, and performing fast Fourier transform on each channel data; Step A2: according to the symmetry of fast Fourier transform, the first half data of each channel is taken; Step A3: the data is divided, and based on the problem of data imbalance under actual working conditions, the training set only contains normal samples, and the test set contains normal and fault samples.

3. The method of claim 1, wherein, The construction method of the initial adjacency matrix in step B comprises: Step B1: traversing each node, combining each node in pairs to form N×N pairs; Step B2: if the 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 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 are the vectors X i and X j are the components in the f-th dimension; Step B4: Weighted Euclidean and Cosine distances as measures of inter-node distance: Step B5: To facilitate a uniform measure, all inter-node distances are standardized by deviation: to the interval [0, 1], where D min and D max are the minimum and maximum values of the inter-node distances, respectively. Step B6: set a distance threshold e, 0 Step B7: if the distance between the two nodes is greater than the threshold e, the sensitivity between the two nodes is 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 e, the sensitivity between the two nodes is high, the corresponding position in the adjacency matrix is 1, and the two nodes are connected; Step B9: obtain the initial adjacency matrix A.

4. The method of claim 3, wherein, The step C specifically comprises: Step Cl: Obtain the normalized Laplacian matrix L of the graph from the adjacency matrix A and the degree matrix D -1 / 2 AD -1 / 2 ; Step C2: input raw data, combine L to implement K-order ARMA filtering: where I is a unit matrix, k = 1, 2,..., K, p k , q k are learnable parameters; Step C3: ARMA iterative weighted T times replacement inverse operation, the output of each layer is: where R and S are learnable matrix parameters, i.e., the initial node features; Step C4: Each layer output passes through a non-linear activation: where σ(·) is a non-linear activation function; Step C5: At this point the output characteristics of the ARMA are: Step C6: To prevent over-smoothing of message passing, message passing and aggregation are called at the k-th layer at the t-th step. During message passing, for a node v, its neighbor node u ∈ E(v), the message passing process is represented as: where m is the message of each node, M is the message passing function, A uv is the connection between node v and its neighbor node u; by averaging the aggregation of each node message, we get: where W is the weight corresponding to each node message; Step C7: By dynamic weighting, the output of the kth order and tth layer at this time is: wherein a is a learnable parameter; Step C8: Thus the complete graph convolution output is:

5. The method of claim 1, wherein, 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, which specifically comprises: Step D1: the output of the previous layer of graph convolution is input into the BatchNorm and Relu activation function to obtain the first layer of graph convolution encoder output, and the first layer of graph convolution encoder output is input into the second layer of graph convolution; Step D2: by adjusting the input and output dimensions, the second layer of graph convolution encoder output is obtained, including the linear mapping of the second layer of graph convolution output and the first layer of graph convolution encoder output, and the second layer of graph convolution encoder output is input into the third layer of graph convolution; Step D3: repeat step D2 to obtain the third layer of graph convolution encoder output, the first layer of graph convolution decoder output, and the second layer of graph convolution decoder output; Step D4: the second layer of graph convolution decoder output is input into the BatchNorm and Relu activation function as the input of the third layer of graph convolution; Step D5: the third layer graph convolution decoder output, i.e. the graph network output, is obtained.

6. The method of claim 1, wherein, The step E specifically comprises: Step E1: Calculate the graph network reconstruction loss by the cosine similarity loss function: where L c is the reconstruction loss corresponding to the cth operating condition data, X c is the input data corresponding to the cth operating condition, X c ' is the output result corresponding to the cth operating condition; Step E2: dynamically adjust the loss weight based on the model convergence speed, first calculate the training speed of the it iteration under the c working condition The smaller the value, the faster the speed, it is 0, 1; Step E3: Calculate the weight of the reconstruction loss corresponding to the it iteration under the c working condition according to the training speed According to Dynamic optimization; Step E4: Calculate the total reconstruction loss for the it iteration Step E5: the iteration is ended, and the network training is completed.

7. The method of claim 1, wherein, The step F comprises: Step F1: the graph network related parameter weight is called; Step F2: the weight is visualized, and the method interpretability is explained through the attention distribution.

8. The method of claim 1, wherein, The step G specifically comprises: Step G1: the training set data is imported into the trained graph network, all loss values are calculated, and the maximum loss value is set as a threshold value for fault data detection; Step G2: Add two kinds of noise in 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 random added Gaussian noise, ε ~ N(0, μ); δ is the scale factor, δ ~ N(1, μ), μ is the variance; in each sample, the probability of randomly adding two kinds of noise is 50%. Step G3: the noise level in the graph network is controlled by adjusting μ, and μ is 0-1; Step G4: the graph network performance is verified under different noise levels, the calculation loss is obtained, and it is judged whether the data is normal or fault according to the detection threshold value obtained in the foregoing; Step G5: the overall evaluation index is output, including the accuracy, precision, F1-score and the area surrounded by the ROC curve and the coordinate axis.

Images