Avionics system fault diagnosis method and system based on graph anomaly detection

Abstract

The invention provides an avionics system fault diagnosis method and system based on graph anomaly detection, and the method comprises the steps: obtaining sensor time sequence data of an avionics system, and constructing a sensor network topological graph based on the sensor time sequence data; updating the embedded representation of the sensor nodes according to the sensor network topological graph; according to the method and the corresponding system, a global attention mechanism and a graph structure are used for fault diagnosis, and the fault type of the avionics system is judged by creating a topological graph, dynamically updating the embedded representation and combining anomaly detection. The whole set of data-driven method closely combines possible fault inference and positioning, so that the accuracy and the response speed of fault diagnosis are greatly improved, and a powerful guarantee is provided for stable operation of an avionics system.

Description

Technical Field

[0001] The present invention proposes an avionics system fault diagnosis method and system based on graph anomaly detection, and relates to the technical field of civil aviation airborne route maintenance. Background Art

[0002] As we all know, avionics systems play a vital role in modern aircraft. The reliability and accuracy of avionics systems directly affect the operation of aircraft and the safety of passengers. Therefore, sensor fault detection has become a hot research area. In the past few decades, a lot of research has been done on the detection of sensor fault states. Traditional fault detection methods are generally based on fault trees, models, and data-driven. However, these methods require accurate process mathematical models and cannot utilize the hidden positional relationships between sensors in non-Euclidean space, so they are only suitable for small systems with fewer inputs and clear mathematical models. However, these performances are greatly affected in the presence of unmodeled interference and uncertainty. In recent years, graph neural networks have emerged. In the real world, faults can change the operating state of the process and cause changes in the dependencies between measurements. Based on this understanding, researchers began to explore GNNs for fault diagnosis tasks. A large number of researchers use structure graph-based methods to diagnose whether a sensor is faulty from the perspective of system topology graphs.

[0003] Although deep neural network models have shown excellent performance in the field of fault detection, we still need to face four key issues. First, most current methods focus on anomaly prediction or classification of tabular data and calculating anomaly scores for fault diagnosis without fully considering the correlations hidden in non-Euclidean space between sensor data. Second, GNN methods usually require accurate access to the connection topology of the sensor network, but in practical applications, it is often difficult to obtain this information. Third, current GNN-based fault detection methods mainly focus on the regression task of local anomalies and detect sensor anomalies from the perspective of root causes, but cannot achieve direct sensor anomaly detection in one step. Finally, in the field of graph anomaly detection, the anomaly scores of nodes in the graph are usually detected by autoencoders or matrix decomposition, but they do not take into account the different dimensions between sensor network topology and attributes. Summary of the invention

[0004] The present invention provides an avionics system fault diagnosis method and system based on graph anomaly detection to solve the above-mentioned problems: The present invention proposes an avionics system fault diagnosis method based on graph anomaly detection, the method comprising: Acquire sensor time series data of the avionics system, and construct a sensor network topology map based on the sensor time series data; updating an embedded representation of a sensor node according to the sensor network topology graph; The reconstruction error is obtained based on the updated embedded representation, and the fault type of the avionics system is determined according to the reconstruction error.

[0005] Furthermore, obtaining sensor time series data of the avionics system and constructing a sensor network topology map based on the sensor time series data include: Acquire time series data of all sensors in the avionics system, the time series data including: voltage, temperature and current; Cleaning and preprocessing the time series data, wherein the cleaning includes using low-pass filtering to remove noise in the time series data and process missing values ​​in the data, and the preprocessing includes unifying the time series data format and timestamp so that all sensor data are synchronously aligned; Select a time window and calculate statistical features for the time series data in each time window, wherein the statistical features include mean, standard deviation, maximum value and minimum value; Calculating a correlation value of the statistical feature of the sensor by using the Pearson correlation coefficient, and when the correlation value does not exceed a first preset threshold, constructing an implicit connection tag of the sensor as no, otherwise constructing an implicit connection tag of the sensor as yes; Constructing feature data and corresponding label data into a data set, and dividing the data set into a training set and a validation set; Creating a recurrent neural network model, inputting the training set into the recurrent neural network model, calculating the loss and updating the parameters of the recurrent neural network model; Run the updated recurrent neural network model on the validation set, obtain the logits output, and generate the Gumbel Softmax output based on the logits output: in, represents the connection probability of sensor nodes i and j in the avionics system, Represents the output probability of the recurrent neural network model, that is, logits output, represents the noise generated from Gumbel distribution, τ represents the temperature parameter; Construct sensor network topology graph based on Gumbel Softmax output.

[0006] Further, updating the embedded representation of the sensor node according to the sensor network topology graph includes: The embedding information of all layer nodes is learned through the learning model. The specific learning model is: in, shows the correlation measure between the i-th sensor node and the j-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the i-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the jth sensor node of the avionics system in the kth layer, f represents the non-negative decreasing function of z, and N represents the number of layers of the recurrent neural network; Where T represents the transpose operation of the vector; Based on the learned embedding information of all layer nodes, the embedding vector of each layer node i is updated through the state update model. Specifically, the state update model is: in, represents the updated embedding representation of the i-th sensor node of the avionics system in the k-th layer, and τ represents the temperature parameter.

[0007] Furthermore, a reconstruction error is obtained based on the updated embedded representation, and the fault type of the avionics system is determined according to the reconstruction error, including: Obtaining a reconstruction error according to the embedded representation of the sensor updated by the recurrent neural network model and the embedded representation of the sensor actually measured; If the reconstruction error is greater than a second preset threshold, marking the sensor node as abnormal; The state of fault propagation is updated based on the marked anomalies using the neighborhood information of the sensor network topology graph, and then the classifier is used to determine the fault type.

[0008] The present invention proposes an avionics system fault diagnosis system based on image anomaly detection, the system comprising: A sensor network topology map construction module is used to obtain sensor time series data of the avionics system and to construct a sensor network topology map based on the sensor time series data; An updating embedded representation module is applied to update the embedded representation of the sensor node according to the sensor network topology diagram; The fault type determination module is used to obtain a reconstruction error based on the updated embedded representation and determine the fault type of the avionics system according to the reconstruction error.

[0009] Furthermore, the module for constructing a sensor network topology diagram includes: A module for acquiring time series data is used to acquire time series data of all sensors in the avionics system, and the time series data includes: voltage, temperature and current; A preprocessing module, used for cleaning and preprocessing the time series data, wherein the cleaning includes using low-pass filtering to remove noise in the time series data and process missing values ​​in the data, and the preprocessing includes unifying the time series data format and timestamp so that all sensor data are synchronously aligned; A feature extraction module is used to select a time window and calculate statistical features for the time series data in each time window, wherein the statistical features include mean, standard deviation, maximum value and minimum value; A module for constructing an implicit label is used to calculate a correlation value of a statistical feature of a sensor by using a Pearson correlation coefficient. When the correlation value does not exceed a first preset threshold, an implicit connection label of the sensor is constructed as no, otherwise an implicit connection label of the sensor is constructed as yes; A data set partitioning module is used to construct a data set from feature data and corresponding label data, and to partition the data set into a training set and a validation set; A training model module, used to create a recurrent neural network model, input a training set into the recurrent neural network model, calculate the loss and update the parameters of the recurrent neural network model; The output module is used to run the updated recurrent neural network model on the validation set, obtain the logits output, and generate the Gumbel Softmax output based on the logits output: in, represents the connection probability of sensor nodes i and j in the avionics system, Represents the output probability of the recurrent neural network model, that is, logits output, represents the noise generated from Gumbel distribution, τ represents the temperature parameter; Building module for constructing sensor network topology graph based on Gumbel Softmax output.

[0010] Further, the updating of the embedded representation module includes: The learning embedding information module is used to learn the embedding information of all layer nodes through the learning model. The specific learning model is: in, represents the correlation measure between the i-th sensor node and the j-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the i-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the jth sensor node of the avionics system in the kth layer, f represents the non-negative decreasing function of z, and N represents the number of layers of the recurrent neural network; Where T represents the transpose operation of the vector; The state update module is used to update the embedding vector of each layer node i through the state update model based on the learned embedding information of all layer nodes. Specifically, the state update model is: in, represents the updated embedding representation of the i-th sensor node of the avionics system in the k-th layer, and τ represents the temperature parameter.

[0011] Furthermore, the fault type determination module includes: A reconstruction error obtaining module is used to obtain a reconstruction error according to the embedded representation of the sensor updated by the recurrent neural network model and the embedded representation of the sensor actually measured; an abnormal marking module, used for marking the sensor node as abnormal if the reconstruction error is greater than a second preset threshold; The judgment module is used to update the state of fault propagation using the neighborhood information of the sensor network topology map according to the marked anomaly, and then use the classifier to judge the fault type.

[0012] Beneficial effects of the present invention: (1) Data diversity: Currently, no open source avionics system-related graph topology data set has been found. Therefore, the first module of this patent can be separated out separately to construct graph topology data from tabular time series data for downstream task research. In addition, when calculating attribute anomalies and structural anomalies, two modules are used to calculate and summarize them separately. Therefore, the framework proposed by this patent can simultaneously meet the anomaly detection tasks of three types of data: time series data, topology graph data, and attribute graph data. It only needs to call different processing modules according to the data type. (2) Diagnostic efficiency: Due to the characteristics of the sensors themselves in the avionics system, in the physical world, there is message flow between each sensor. Therefore, the traditional GNN and the most basic MLP aggregation neighbor method are discarded, and an attention mechanism for calculating all-pair full attention is adopted. This method can avoid the method of GNN and MLP traversing each layer of neighbors, greatly improving the diagnostic efficiency. (3) Granularity independence: In the framework proposed by this patent, each part can be separated out as a submodule for the research of other tasks. For example, if only the attribute anomaly detection module is retained, the framework will fall back to the category of time series data anomaly detection; if only the structure anomaly detection module is retained, the framework will fall back to the category of graph structure anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A schematic diagram of an avionics system fault diagnosis method based on graph anomaly detection according to the present invention. DETAILED DESCRIPTION

[0014] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0015] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. The embodiments described are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0017] One embodiment of the present invention provides an avionics system fault diagnosis method based on graph anomaly detection, the method comprising: Acquire sensor time series data of the avionics system, and construct a sensor network topology map based on the sensor time series data; updating an embedded representation of a sensor node according to the sensor network topology graph; The reconstruction error is obtained based on the updated embedded representation, and the fault type of the avionics system is determined according to the reconstruction error.

[0018] The working principle and effect of the above technical solution are as follows: in the data collection and graph construction stage, the time series data of each sensor of the avionics system is collected, and a network topology graph is constructed based on the correlation between the sensors. The nodes in the graph represent the sensors, and the edges represent the correlation between the sensors. In the embedded representation update stage, the graph neural network is used to learn the feature representation of each sensor node. The embedded representation of nodes is optimized through iterative updating, taking into account the topological relationship and timing characteristics between nodes; the reconstruction error is calculated in the fault diagnosis stage, and the difference between the original data and the reconstructed data is compared; different types of faults are identified through error patterns; faults are classified based on the size and distribution characteristics of the reconstruction error. The diagnostic accuracy is improved, and the correlation information between sensors is fully utilized to reduce the false alarm rate and missed alarm rate, and improve the accuracy of fault identification; the diagnostic ability is enhanced, and complex fault modes can be identified, and multi-sensor collaborative faults can be handled, and the dynamic changes of the system can be adapted; there is no need for a lot of expert experience, and the system characteristics are automatically learned, with strong scalability; the efficiency is improved, real-time monitoring and diagnosis, and the fault source is quickly located, reducing maintenance time and cost; the cause of the fault is analyzed through reconstruction error, and the basis for fault diagnosis is provided, which is convenient for maintenance personnel to understand and handle. It combines the advantages of graph theory, deep learning and fault diagnosis, and provides an intelligent and automated solution for the fault diagnosis of avionics systems.

[0019] In one embodiment of the present invention, obtaining sensor time series data of an avionics system and constructing a sensor network topology map based on the sensor time series data includes: Acquire time series data of all sensors in the avionics system, the time series data including: voltage, temperature and current; Cleaning and preprocessing the time series data, wherein the cleaning includes using low-pass filtering to remove noise in the time series data and process missing values ​​in the data, and the preprocessing includes unifying the time series data format and timestamp so that all sensor data are synchronously aligned; Select a time window and calculate statistical features for the time series data in each time window, wherein the statistical features include mean, standard deviation, maximum value and minimum value; Calculating a correlation value of the statistical feature of the sensor by using the Pearson correlation coefficient, and when the correlation value does not exceed a first preset threshold, constructing an implicit connection tag of the sensor as no, otherwise constructing an implicit connection tag of the sensor as yes; Constructing feature data and corresponding label data into a data set, and dividing the data set into a training set and a validation set; Creating a recurrent neural network model, inputting the training set into the recurrent neural network model, calculating the loss and updating the parameters of the recurrent neural network model; Run the updated recurrent neural network model on the validation set, obtain the logits output, and generate the Gumbel Softmax output based on the logits output: in, represents the connection probability of sensor nodes i and j in the avionics system, Represents the output probability of the recurrent neural network model, that is, logits output, represents the noise generated from Gumbel distribution, τ represents the temperature parameter; Construct sensor network topology graph based on Gumbel Softmax output.

[0020] The working principle and effect of the above technical solution are as follows: Since the structure of the avionics system involves national security and military secrets, its confidentiality requirements are extremely high. Therefore, there is currently no public avionics system network topology map data set available for use. This confidentiality also limits the research and application of avionics systems in academia and industry. However, although it is impossible to directly obtain the network topology map, valuable information can still be mined from the sensor time series data. Sensor time series data is an important data type generated during the operation of the avionics system. It records the real-time status of each sensor and the connection relationship between them. These data contain rich information, including the working status of the sensor, the signal transmission path, the interaction between devices, etc. By analyzing these data, the network topology of the avionics system can be inferred. In order to extract the position connection relationship of each sensor from the sensor time series data, the Gumbel-softmax algorithm is used. This algorithm is a probabilistic model for discrete sequence processing, which can effectively process the uncertainty and noise in the sensor time series data, thereby accurately extracting the connection relationship between each sensor. In this way, the original data set of the avionics system sensor network topology map can be constructed as the basis for downstream tasks. The working principle of the above technical solution is as follows: in the data collection stage, time series data is collected from the sensors (including voltage, temperature and current) of the avionics system. The time series data is the dynamic operating state of the sensor and is used for scientific calculation of features and analysis of network topology; high-frequency noise in the time series data is removed by low-pass filtering, missing values ​​are processed, data quality is improved, and the timestamp and format of the data are unified to ensure that multiple sensor data are synchronized in time, providing a reliable data basis for subsequent analysis and modeling; the data is divided into multiple time windows (fixed length) to facilitate capturing the dynamic changes of the sensor, and statistical features (such as mean, standard deviation, maximum value, minimum value) are calculated for the time series data in each time window to form quantifiable feature data; the correlation of statistical features between sensors is calculated by the Pearson correlation coefficient to measure the degree of linear correlation between two sensors, and the connection of the sensor is marked as "yes" or "no" based on the correlation value and the preset threshold. Implicit connection labels can describe the preliminary relationship diagram between sensor nodes; integrate feature data and implicit connection labels into a data set for training machine learning models, and divide the data set into a training set and a validation set to serve model training and effect verification; Construct a neural network model (such as a recurrent neural network, RNN) and input the training set. Continuously optimize the model parameters by calculating the loss function (such as cross entropy loss). Run the optimized model on the validation set to obtain the logits output. Use Gumbel Softmax to convert logits into node connection probabilities: Correct the model output by using the logits value (normalized probability) and Gumbel random noise. Use the temperature parameter τ to control the transition of the output from continuous probability to discrete connection state. Based on the output of Gumbel Softmax, the connection probability of each pair of sensor nodes is Transformed into specific connections (connected / unconnected). Intelligently construct sensor network topology maps, and realize automated topology map construction by combining time series data, correlation analysis, and deep learning technology, without relying on artificial expert knowledge or predefined rules; the model can dynamically extract potential relationships between sensors, and improve the model's fault diagnosis capabilities for avionics systems. The constructed sensor network topology map can capture the correlation between sensors and provide data support for further analysis (such as fault detection and location); improve the robustness and accuracy of the system, data cleaning (denoising, missing value compensation) and preprocessing (synchronous alignment) improve the quality of input data, and use recurrent neural networks to process time series data to improve the ability to handle dynamic characteristics; use GumbelSoftmax The randomness and continuity of the model enhance the generalization ability of the model, avoid overfitting, and improve the ability to quantify the association between nodes. The output of the connection between nodes is a probability value. Through flexible processing, it can not only describe whether the connection exists, but also quantify the strength of the connection. This detailed quantitative analysis helps to more accurately infer the interaction intensity or abnormal state that deviates from the normal mode between the key sensors in the avionics system. It realizes the fully automated process of data collection, cleaning, feature extraction, model training and topology construction, and integrates the processing of each stage in an integrated manner. From inputting sensor time series data to building a complete sensor network topology map, the process does not require human intervention, which simplifies the complexity of the system. The scheme has the advantage of strong adaptability by introducing Pearson correlation coefficient and neural network modeling: it is suitable for different types of sensor data (such as temperature, voltage, current, etc.), supports avionics systems or other similar networks of different sizes; dynamically adjusts temperature parameters , which can flexibly balance between exploring multiple topological structures and converging to a specific structure; by integrating data analysis, deep learning and probabilistic modeling, it successfully constructs a network topology map based on the time series data of sensors. Its working principle is to start with data collection and preprocessing under the coordination of multiple modules, and gradually generate a topological map through feature extraction, correlation analysis and modeling; Quantify the connectivity between nodes in the form of probability. The sensor connection label (connected / unconnected, i.e., {0, 1}) is a discrete value, but in order to train and optimize the model, it needs to be differentiable in gradient descent; an important function of Gumbel Softmax is to provide a continuous distribution that approximates discrete variables, i.e. The value of can be regarded as the probability representation of whether nodes i and j are connected or not (0 or 1); Gumbel noise is added to optimize the randomness Gumbel distribution Perturb the logits and add noise; noise can help the model explore different possibilities, prevent it from falling into the local optimal solution, and enhance the generalization ability of the model; temperature parameter The introduction of Controls the smoothness of the Softmax output; when When , the output approaches hard classification (close to discrete value 0 / 1); when When , all categories of output tend to be evenly distributed, and connectivity cannot be distinguished; by dynamically adjusting , the model can gradually transition from the initial relatively smooth probability distribution to a nearly discrete output. One of the advantages of using Gumbel Softmax is that it can transform discrete variable problems into continuous variable approximations, thereby supporting end-to-end gradient optimization in deep learning, which is very important in neural network training; Keep the smoothness of Softmax output and try to explore multiple possible topological structures. Later, gradually reduce , and obtain outputs close to the real discrete labels (connected / unconnected), which are more in line with the connectivity of the actual sensor network; automatically construct the sensor network topology map, calculate the node connection probability through logits and GumbelSoftmax output, and dynamically generate the sensor network topology map without pre-setting a fixed topology structure.

[0021] In one embodiment of the present invention, updating the embedded representation of the sensor node according to the sensor network topology graph includes: The embedding information of all layer nodes is learned through the learning model. The specific learning model is: in, represents the correlation measure between the i-th sensor node and the j-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the i-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the jth sensor node of the avionics system in the kth layer, f represents the non-negative decreasing function of z, and N represents the number of layers of the recurrent neural network; Where T represents the transpose operation of the vector; Based on the learned embedding information of all layer nodes, the embedding vector of each layer node i is updated through the state update model. Specifically, the state update model is: in, represents the updated embedding representation of the i-th sensor node of the avionics system in the k-th layer, Represents the temperature parameter.

[0022] The working principle and effect of the above technical solution are as follows: construct an embedded representation of the sensor nodes in the avionics system, analyze the correlation between sensors, and thus describe the overall operating status or fault distribution mode of the avionics system network. By recursively updating the embedded representation, the digital expression of the interaction relationship between sensors is gradually optimized; learn the correlation between nodes. In the kth layer, the correlation (similarity) between sensors i and j is calculated through the learning model. First, the embedding of each pair of sensors and Calculate the square of the difference and use function f to calculate the similarity. This function combines the distance information and the cosine similarity of the feature direction. Normalize the similarities of all nodes to get the correlation between the two sensors. , used to characterize the strength of their interaction; update the embedding representation of each layer of nodes, according to the adjacency relationship of each node (through Encoding), recursively update the embedding vector; The first term represents the self-sustaining property of sensor i, combined with The weaker connection is attenuated; the second term: represents the information propagation (or dynamic interaction) of the adjacent sensor node (j) to the current sensor i; recursive multi-layer update captures features, and through recursive multi-layer update embedding representation, each layer captures the deeper correlation characteristics between avionics sensors. The embedding representation obtained in the final layer can be used to construct the avionics sensor network topology; the correlation clarifies which sensors have strong correlations and their correlation strengths; The updated embedding captures the high-level features of the sensor and is used to infer the fault propagation path or network status. The state perception capability of the avionics system is improved. The updating process of the embedding vector integrates the self-characteristics of the sensor and the output of the adjacent sensors, and can capture the complex spatiotemporal dependencies between sensors in the avionics system. The obtained embedding features can be further used for system performance monitoring, state prediction, and anomaly detection; the correlation between sensors can be accurately quantified, and the correlation between sensors i and j can be quantified, as well as their interaction strength can be captured. Strongly associated sensor groups can be used to detect possible fault propagation paths or operating modes; automatic inference of complex topologies is supported, and the input sensor network can be of any structure (such as fully connected or sparsely connected). The recursive update replaces the manually set topology assumptions, making the model more adaptable to the complexity of the avionics system; recursive multi-layer updates transmit information about the entire network, allowing the system to maintain the stability of the entire network even when local data is abnormal. To assist in avionics system optimization and fault location, the final embedding and association vectors can be used to detect implicit abnormal correlations between sensors and find system weaknesses. The topology construction scheme provides a reliable basis for the monitoring and troubleshooting of complex avionics systems. Through the design of recursive updates and non-negative decreasing functions, the efficiency and reliability of avionics system fault diagnosis, association analysis and overall status monitoring are improved.

[0023] In one embodiment of the present invention, a reconstruction error is obtained based on an updated embedded representation, and a fault type of an avionics system is determined according to the reconstruction error, including: Obtaining a reconstruction error according to the embedded representation of the sensor updated by the recurrent neural network model and the embedded representation of the sensor actually measured; If the reconstruction error is greater than a second preset threshold, marking the sensor node as abnormal; The state of fault propagation is updated based on the marked anomalies using the neighborhood information of the sensor network topology graph, and then the classifier is used to determine the fault type.

[0024] The process of updating the fault propagation state: construct the sensor network topology model of the avionics system to capture the correlation between sensors; initialize the abnormal state and mark the initial abnormal node according to the reconstruction error; update the state based on the neighborhood propagation, and Diffusion of abnormal impact, update the status layer by layer; analyze the propagation results, obtain the global abnormal status diagram, and locate possible system-level fault sources.

[0025] The working principle and effect of the above technical solution are: through the recurrent neural network (RNN) model, the historical time series data of the sensor node is learned, and the sensor embedded representation is updated and generated. , capturing the dynamic operation behavior of sensor nodes and their characteristics in the time dimension; the actual measurement data also generates an embedded representation of the sensor , represents the state actually observed by the sensor; compared with the predicted embedding and the actual embedding The reconstruction error is calculated by: is a scalar that indicates whether the current state of sensor i in the avionics system deviates from the normal prediction of the historical model. If it exceeds the second preset threshold, it indicates that the observation value of sensor i deviates greatly from the normal state, and the sensor is marked as an abnormal node; the features of the input classifier are extracted for the abnormally marked nodes and their global propagation state graph, such as the embedded representation of each sensor, the number and distribution of abnormal sensors, the topological structure of the fault propagation path and the impact range and abnormality information of the neighborhood propagation; the classifier model is trained. The classifier can be a machine learning model (such as logistic regression, random forest) or a deep learning model (such as multi-layer perceptron, graph neural network). The model goal is to determine the fault type of the avionics system based on the global abnormal state graph and topological features: power system failure: such as voltage / current anomalies directly affecting a group of sensors, or indirectly affecting subsystem functions; temperature control system failure, such as temperature anomalies spreading to multiple nodes, resulting in failure of associations between different subsystems; communication system failure, such as multiple sensors reporting anomalies at the same time, presenting a network communication problem. Accurately capture system anomalies, sensor anomaly detection, recurrent neural network model combined with reconstruction error, can accurately capture the abnormal state of the sensor, threshold control ensures sensitive response to real abnormal conditions (such as temperature / voltage fluctuations), and does not hinder the small-scale fluctuations that hinder the normal operation of the sensor; global system anomaly perception, using the neighborhood association in the sensor network, can start from local anomalies, analyze the abnormal distribution and impact range of the entire avionics system, and capture the global state of system anomalies more comprehensively; from anomaly to fault inference, it is not limited to capturing abnormal nodes, but through topological propagation analysis and global state feature extraction, accurately classify the system type and cause of the fault; support complex working condition associations of avionics systems, such as the mutual influence between multiple subsystems; support multiple fault type classifications, and the classifier can be trained to identify various typical fault types of avionics systems, covering common subsystem faults such as power supply, temperature, communication, and navigation; real-time update Function, through continuous real-time data stream input to the recurrent neural network, the embedded representation and topological propagation state of the corresponding sensor will be updated at each time step, which can realize real-time detection and response of faults; from sensor anomaly detection to avionics system fault classification, all are end-to-end automatic processing, assisting engineers to quickly locate problems; reduce false alarm and missed alarm rates, apply embedded representation and reconstruction errors to filter out random noise or occasional single-point anomalies, and exclude irrelevant influencing sensors through topological propagation updates, focusing on actual fault modes; through anomaly detection, anomaly propagation state updates and centralized classification, not only can abnormal behavior of sensors in avionics systems be detected, but also the propagation mode of anomalies within the system can be analyzed, so as to accurately determine the fault type of avionics systems and improve the accuracy of fault location and classification; optimize system safety and stability; realize automated fault diagnosis process, support real-time monitoring and intelligent operation and maintenance of avionics systems in complex environments.

[0026] One embodiment of the present invention is an avionics system fault diagnosis system based on image anomaly detection, the system comprising: A sensor network topology map construction module is used to obtain sensor time series data of the avionics system and to construct a sensor network topology map based on the sensor time series data; An updating embedded representation module is applied to update the embedded representation of the sensor node according to the sensor network topology diagram; The fault type determination module is used to obtain a reconstruction error based on the updated embedded representation and determine the fault type of the avionics system according to the reconstruction error.

[0027] The working principle and effect of the above technical solution are as follows: in the data collection and graph construction stage, the time series data of each sensor of the avionics system is collected, and a network topology graph is constructed based on the correlation between the sensors. The nodes in the graph represent the sensors, and the edges represent the correlation between the sensors. In the embedded representation update stage, the graph neural network is used to learn the feature representation of each sensor node. The embedded representation of nodes is optimized through iterative updating, taking into account the topological relationship and timing characteristics between nodes; the reconstruction error is calculated in the fault diagnosis stage, and the difference between the original data and the reconstructed data is compared; different types of faults are identified through error patterns; faults are classified based on the size and distribution characteristics of the reconstruction error. The diagnostic accuracy is improved, and the correlation information between sensors is fully utilized to reduce the false alarm rate and missed alarm rate, and improve the accuracy of fault identification; the diagnostic ability is enhanced, and complex fault modes can be identified, and multi-sensor collaborative faults can be handled, and the dynamic changes of the system can be adapted; there is no need for a lot of expert experience, and the system characteristics are automatically learned, with strong scalability; the efficiency is improved, real-time monitoring and diagnosis, and the fault source is quickly located, reducing maintenance time and cost; the cause of the fault is analyzed through reconstruction error, and the basis for fault diagnosis is provided, which is convenient for maintenance personnel to understand and handle. It combines the advantages of graph theory, deep learning and fault diagnosis, and provides an intelligent and automated solution for the fault diagnosis of avionics systems.

[0028] In one embodiment of the present invention, the module for constructing a sensor network topology diagram includes: A module for acquiring time series data is used to acquire time series data of all sensors in the avionics system, and the time series data includes: pressure, temperature and current; A preprocessing module, used for cleaning and preprocessing the time series data, wherein the cleaning includes using low-pass filtering to remove noise in the time series data and process missing values ​​in the data, and the preprocessing includes unifying the time series data format and timestamp so that all sensor data are synchronously aligned; A feature extraction module is used to select a time window and calculate statistical features for the time series data in each time window, wherein the statistical features include mean, standard deviation, maximum value and minimum value; A module for constructing an implicit label is used to calculate a correlation value of a statistical feature of a sensor by using a Pearson correlation coefficient. When the correlation value does not exceed a first preset threshold, an implicit connection label of the sensor is constructed as no, otherwise an implicit connection label of the sensor is constructed as yes; A data set partitioning module is used to construct a data set from feature data and corresponding label data, and to partition the data set into a training set and a validation set; A training model module, used to create a recurrent neural network model, input a training set into the recurrent neural network model, calculate the loss and update the parameters of the recurrent neural network model; The output module is used to run the updated recurrent neural network model on the validation set, obtain the logits output, and generate the Gumbel Softmax output based on the logits output: in, represents the connection probability of sensor nodes i and j in the avionics system, Represents the output probability of the recurrent neural network model, that is, logits output, represents the noise generated from a Gumbel distribution, represents the temperature parameter; Building module for constructing sensor network topology graph based on Gumbel Softmax output.

[0029] The working principle and effect of the above technical solution are as follows: Since the structure of the avionics system involves national security and military secrets, its confidentiality requirements are extremely high. Therefore, there is currently no public avionics system network topology map data set available for use. This confidentiality also limits the research and application of avionics systems in academia and industry. However, although it is impossible to directly obtain the network topology map, valuable information can still be mined from the sensor time series data. Sensor time series data is an important data type generated during the operation of the avionics system. It records the real-time status of each sensor and the connection relationship between them. These data contain rich information, including the working status of the sensor, the signal transmission path, the interaction between devices, etc. By analyzing these data, the network topology of the avionics system can be inferred. In order to extract the position connection relationship of each sensor from the sensor time series data, the Gumbel-softmax algorithm is used. This algorithm is a probabilistic model for discrete sequence processing, which can effectively process the uncertainty and noise in the sensor time series data, thereby accurately extracting the connection relationship between each sensor. In this way, the original data set of the avionics system sensor network topology map can be constructed as the basis for downstream tasks. The working principle of the above technical solution is as follows: in the data collection stage, time series data is collected from the sensors (including voltage, temperature and current) of the avionics system. The time series data is the dynamic operating state of the sensor and is used for scientific calculation of features and analysis of network topology; high-frequency noise in the time series data is removed by low-pass filtering, missing values ​​are processed, data quality is improved, and the timestamp and format of the data are unified to ensure that multiple sensor data are synchronized in time, providing a reliable data basis for subsequent analysis and modeling; the data is divided into multiple time windows (fixed length) to facilitate capturing the dynamic changes of the sensor, and statistical features (such as mean, standard deviation, maximum value, minimum value) are calculated for the time series data in each time window to form quantifiable feature data; the correlation of statistical features between sensors is calculated by the Pearson correlation coefficient to measure the degree of linear correlation between two sensors, and the connection of the sensor is marked as "yes" or "no" based on the correlation value and the preset threshold. Implicit connection labels can describe the preliminary relationship diagram between sensor nodes; integrate feature data and implicit connection labels into a data set for training machine learning models, and divide the data set into a training set and a validation set to serve model training and effect verification; Construct a neural network model (such as a recurrent neural network, RNN) and input the training set. Continuously optimize the model parameters by calculating the loss function (such as cross entropy loss). Run the optimized model on the validation set to obtain the logits output. Use Gumbel Softmax to convert logits into node connection probabilities: Correct the model output by using the logits value (normalized probability) and Gumbel random noise. Use the temperature parameter τ to control the transition of the output from continuous probability to discrete connection state. Based on the output of Gumbel Softmax, the connection probability of each pair of sensor nodes is Transformed into specific connections (connected / unconnected). Intelligently construct sensor network topology maps, and realize automated topology map construction by combining time series data, correlation analysis, and deep learning technology, without relying on artificial expert knowledge or predefined rules; the model can dynamically extract potential relationships between sensors, and improve the model's fault diagnosis capabilities for avionics systems. The constructed sensor network topology map can capture the correlation between sensors and provide data support for further analysis (such as fault detection and location); improve the robustness and accuracy of the system, data cleaning (denoising, missing value compensation) and preprocessing (synchronous alignment) improve the quality of input data, and use recurrent neural networks to process time series data to improve the ability to handle dynamic characteristics; use GumbelSoftmax The randomness and continuity of the model enhance the generalization ability of the model, avoid overfitting, and improve the ability to quantify the association between nodes. The output of the connection between nodes is a probability value. Through flexible processing, it can not only describe whether the connection exists, but also quantify the strength of the connection. This detailed quantitative analysis helps to more accurately infer the interaction intensity or abnormal state that deviates from the normal mode between the key sensors in the avionics system. It realizes the fully automated process of data collection, cleaning, feature extraction, model training and topology construction, and integrates the processing of each stage in an integrated manner. From inputting sensor time series data to building a complete sensor network topology map, the process does not require human intervention, which simplifies the complexity of the system. The scheme has the advantage of strong adaptability by introducing Pearson correlation coefficient and neural network modeling: it is suitable for different types of sensor data (such as temperature, voltage, current, etc.), supports avionics systems or other similar networks of different sizes; dynamically adjusts temperature parameters , which can flexibly balance between exploring multiple topological structures and converging to a specific structure; by integrating data analysis, deep learning and probabilistic modeling, it successfully constructs a network topology map based on the time series data of sensors. Its working principle is to start with data collection and preprocessing under the coordination of multiple modules, and gradually generate a topological map through feature extraction, correlation analysis and modeling; Quantify the connectivity between nodes in the form of probability. The sensor connection label (connected / unconnected, i.e., {0, 1}) is a discrete value, but in order to train and optimize the model, it needs to be differentiable in gradient descent; an important function of Gumbel Softmax is to provide a continuous distribution that approximates discrete variables, i.e. The value of can be regarded as the probability representation of whether nodes i and j are connected or not (0 or 1); Gumbel noise is added to optimize the randomness Gumbel distribution Perturb the logits and add noise; noise can help the model explore different possibilities, prevent it from falling into the local optimal solution, and enhance the generalization ability of the model; temperature parameter The introduction of Controls the smoothness of the Softmax output; when When , the output approaches hard classification (close to discrete value 0 / 1); when When , all categories of output tend to be evenly distributed, and connectivity cannot be distinguished; by dynamically adjusting , the model can gradually transition from the initial relatively smooth probability distribution to a nearly discrete output. One of the advantages of using Gumbel Softmax is that it can transform discrete variable problems into continuous variable approximations, thereby supporting end-to-end gradient optimization in deep learning, which is very important in neural network training; Keep the smoothness of Softmax output and try to explore multiple possible topological structures. Later, gradually reduce , and obtain outputs close to the real discrete labels (connected / unconnected), which are more in line with the connectivity of the actual sensor network; automatically construct the sensor network topology map, calculate the node connection probability through logits and GumbelSoftmax output, and dynamically generate the sensor network topology map without pre-setting a fixed topology structure.

[0030] In one embodiment of the present invention, the updating of the embedded representation module comprises: The learning embedding information module is used to learn the embedding information of all layer nodes through the learning model. The specific learning model is: in, represents the correlation measure between the i-th sensor node and the j-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the i-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the jth sensor node of the avionics system in the kth layer, f represents the non-negative decreasing function of z, and N represents the number of layers of the recurrent neural network; Where T represents the transpose operation of the vector; The state update module is used to update the embedding vector of each layer node i through the state update model based on the learned embedding information of all layer nodes. Specifically, the state update model is: in, represents the updated embedding representation of the i-th sensor node of the avionics system in the k-th layer, Represents the temperature parameter.

[0031] The working principle and effect of the above technical solution are as follows: construct an embedded representation of the sensor nodes in the avionics system, analyze the correlation between sensors, and thus describe the overall operating status or fault distribution mode of the avionics system network. By recursively updating the embedded representation, the digital expression of the interaction relationship between sensors is gradually optimized; learn the correlation between nodes. In the kth layer, the correlation (similarity) between sensors i and j is calculated through the learning model. First, the embedding of each pair of sensors and Calculate the square of the difference and use function f to calculate the similarity. This function combines the distance information and the cosine similarity of the feature direction. Normalize the similarities of all nodes to get the correlation between the two sensors. , used to characterize the strength of their interaction; update the embedding representation of each layer of nodes, according to the adjacency relationship of each node (through Encoding), recursively update the embedding vector; The first term represents the self-sustaining property of sensor i, combined with The weaker connection is attenuated; the second term: represents the information propagation (or dynamic interaction) of the adjacent sensor node (j) to the current sensor i; recursive multi-layer update captures features, and through recursive multi-layer update embedding representation, each layer captures the deeper correlation characteristics between avionics sensors. The embedding representation obtained in the final layer can be used to construct the avionics sensor network topology; the correlation clarifies which sensors have strong correlations and their correlation strengths; The updated embedding captures the high-level features of the sensor and is used to infer the fault propagation path or network status. The state perception capability of the avionics system is improved. The updating process of the embedding vector integrates the self-characteristics of the sensor and the output of the adjacent sensors, and can capture the complex spatiotemporal dependencies between sensors in the avionics system. The obtained embedding features can be further used for system performance monitoring, state prediction, and anomaly detection; the correlation between sensors can be accurately quantified, and the correlation between sensors i and j can be quantified, as well as their interaction strength can be captured. Strongly associated sensor groups can be used to detect possible fault propagation paths or operating modes; automatic inference of complex topologies is supported, and the input sensor network can be of any structure (such as fully connected or sparsely connected). The recursive update replaces the manually set topology assumptions, making the model more adaptable to the complexity of the avionics system; recursive multi-layer updates transmit information about the entire network, allowing the system to maintain the stability of the entire network even when local data is abnormal. To assist in avionics system optimization and fault location, the final embedding and association vectors can be used to detect implicit abnormal correlations between sensors and find system weaknesses. The topology construction scheme provides a reliable basis for the monitoring and troubleshooting of complex avionics systems. Through the design of recursive updates and non-negative decreasing functions, the efficiency and reliability of avionics system fault diagnosis, association analysis and overall status monitoring are improved.

[0032] In one embodiment of the present invention, the fault type determination module includes: A reconstruction error obtaining module is used to obtain a reconstruction error according to the embedded representation of the sensor updated by the recurrent neural network model and the embedded representation of the sensor actually measured; an abnormal marking module, used for marking the sensor node as abnormal if the reconstruction error is greater than a second preset threshold; The judgment module is used to update the state of fault propagation using the neighborhood information of the sensor network topology map according to the marked anomaly, and then use the classifier to judge the fault type.

[0033] The working principle and effect of the above technical solution are: through the recurrent neural network (RNN) model, the historical time series data of the sensor node is learned, and the sensor embedded representation is updated and generated. , capturing the dynamic operation behavior of sensor nodes and their characteristics in the time dimension; the actual measurement data also generates an embedded representation of the sensor , represents the state actually observed by the sensor; compared with the predicted embedding and the actual embedding The difference between , calculate the reconstruction error: is a scalar that indicates whether the current state of sensor i in the avionics system deviates from the normal prediction of the historical model. If it exceeds the second preset threshold, it indicates that the observation value of sensor i deviates greatly from the normal state, and the sensor is marked as an abnormal node; the features of the input classifier are extracted for the abnormally marked nodes and their global propagation state graph, such as the embedded representation of each sensor, the number and distribution of abnormal sensors, the topological structure of the fault propagation path and the impact range and abnormality information of the neighborhood propagation; the classifier model is trained. The classifier can be a machine learning model (such as logistic regression, random forest) or a deep learning model (such as multi-layer perceptron, graph neural network). The model goal is to determine the fault type of the avionics system based on the global abnormal state graph and topological features: power system failure: such as voltage / current anomalies directly affecting a group of sensors, or indirectly affecting subsystem functions; temperature control system failure, such as temperature anomalies spreading to multiple nodes, resulting in failure of associations between different subsystems; communication system failure, such as multiple sensors reporting anomalies at the same time, presenting a network communication problem. Accurately capture system anomalies, sensor anomaly detection, recurrent neural network model combined with reconstruction error, can accurately capture the abnormal state of the sensor, threshold control ensures sensitive response to real abnormal conditions (such as temperature / voltage fluctuations), and does not hinder the small-scale fluctuations that hinder the normal operation of the sensor; global system anomaly perception, using the neighborhood association in the sensor network, can start from local anomalies, analyze the abnormal distribution and impact range of the entire avionics system, and capture the global state of system anomalies more comprehensively; from anomaly to fault inference, it is not limited to capturing abnormal nodes, but through topological propagation analysis and global state feature extraction, accurately classify the system type and cause of the fault; support complex working condition associations of avionics systems, such as the mutual influence between multiple subsystems; support multiple fault type classifications, and the classifier can be trained to identify various typical fault types of avionics systems, covering common subsystem faults such as power supply, temperature, communication, and navigation; real-time update Function, through continuous real-time data stream input to the recurrent neural network, the embedded representation and topological propagation state of the corresponding sensor will be updated at each time step, which can realize real-time detection and response of faults; from sensor anomaly detection to avionics system fault classification, all are end-to-end automatic processing, assisting engineers to quickly locate problems; reduce false alarm and missed alarm rates, apply embedded representation and reconstruction errors to filter out random noise or occasional single-point anomalies, and exclude irrelevant influencing sensors through topological propagation updates, focusing on actual fault modes; through anomaly detection, anomaly propagation state updates and centralized classification, not only can abnormal behavior of sensors in avionics systems be detected, but also the propagation mode of anomalies within the system can be analyzed, so as to accurately determine the fault type of avionics systems and improve the accuracy of fault location and classification; optimize system safety and stability; realize automated fault diagnosis process, support real-time monitoring and intelligent operation and maintenance of avionics systems in complex environments.

[0034] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for avionics system fault diagnosis based on graph anomaly detection, characterized in that: The method comprises: Acquire sensor time series data of the avionics system, and construct a sensor network topology map based on the sensor time series data; updating an embedded representation of a sensor node according to the sensor network topology graph; The reconstruction error is obtained based on the updated embedded representation, and the fault type of the avionics system is determined according to the reconstruction error.

2. The avionics system fault diagnosis method based on graph anomaly detection according to claim 1 is characterized in that: Acquiring sensor time series data of the avionics system and constructing a sensor network topology map based on the sensor time series data, including: Acquire time series data of all sensors in the avionics system, the time series data including: voltage, temperature and current; Cleaning and preprocessing the time series data, wherein the cleaning includes using low-pass filtering to remove noise in the time series data and process missing values ​​in the data, and the preprocessing includes unifying the time series data format and timestamp so that all sensor data are synchronously aligned; Select a time window and calculate statistical features for the time series data in each time window, wherein the statistical features include mean, standard deviation, maximum value and minimum value; Calculating a correlation value of the statistical feature of the sensor by using the Pearson correlation coefficient, and when the correlation value does not exceed a first preset threshold, constructing an implicit connection tag of the sensor as no, otherwise constructing an implicit connection tag of the sensor as yes; Constructing feature data and corresponding label data into a data set, and dividing the data set into a training set and a validation set; Creating a recurrent neural network model, inputting the training set into the recurrent neural network model, calculating the loss and updating the parameters of the recurrent neural network model; Run the updated recurrent neural network model on the validation set, obtain the logits output, and generate the Gumbel Softmax output based on the logits output: in, represents the connection probability of sensor nodes i and j in the avionics system, Represents the output probability of the recurrent neural network model, that is, logits output, represents the noise generated from Gumbel distribution, τ represents the temperature parameter; Construct sensor network topology graph based on Gumbel Softmax output.

3. The avionics system fault diagnosis method based on graph anomaly detection according to claim 1 is characterized in that: Updating the embedded representation of the sensor node according to the sensor network topology graph includes: The embedding information of all layer nodes is learned through the learning model. The specific learning model is: in, represents the correlation measure between the i-th sensor node and the j-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the i-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the jth sensor node of the avionics system in the kth layer, f represents the non-negative decreasing function of z, and N represents the number of layers of the recurrent neural network; Wherein, T represents the transpose operation of the vector; Based on the learned embedding information of all layer nodes, the embedding vector of each layer node i is updated through the state update model. Specifically, the state update model is: in, represents the updated embedding representation of the i-th sensor node of the avionics system in the k-th layer, and τ represents the temperature parameter.

4. The avionics system fault diagnosis method based on graph anomaly detection according to claim 1 is characterized in that: The reconstruction error is obtained based on the updated embedded representation, and the fault type of the avionics system is determined according to the reconstruction error, including: Obtaining a reconstruction error according to the embedded representation of the sensor updated by the recurrent neural network model and the embedded representation of the sensor actually measured; If the reconstruction error is greater than a second preset threshold, marking the sensor node as abnormal; The state of fault propagation is updated based on the marked anomalies using the neighborhood information of the sensor network topology graph, and then the classifier is used to determine the fault type.

5. An avionics system fault diagnosis system based on graph anomaly detection, characterized in that: The system comprises: A sensor network topology map construction module is used to obtain sensor time series data of the avionics system and to construct a sensor network topology map based on the sensor time series data; An updating embedded representation module is applied to update the embedded representation of the sensor node according to the sensor network topology diagram; The fault type determination module is used to obtain a reconstruction error based on the updated embedded representation and determine the fault type of the avionics system according to the reconstruction error.

6. The avionics system fault diagnosis system based on graph anomaly detection according to claim 5, characterized in that: The module for constructing a sensor network topology diagram includes: A module for acquiring time series data is used to acquire time series data of all sensors in the avionics system, and the time series data includes: voltage, temperature and current; A preprocessing module, used for cleaning and preprocessing the time series data, wherein the cleaning includes using low-pass filtering to remove noise in the time series data and process missing values ​​in the data, and the preprocessing includes unifying the time series data format and timestamp so that all sensor data are synchronously aligned; A feature extraction module is used to select a time window and calculate statistical features for the time series data in each time window, wherein the statistical features include mean, standard deviation, maximum value and minimum value; A module for constructing an implicit label is used to calculate a correlation value of a statistical feature of a sensor by using a Pearson correlation coefficient. When the correlation value does not exceed a first preset threshold, an implicit connection label of the sensor is constructed as no, otherwise an implicit connection label of the sensor is constructed as yes; A data set partitioning module is used to construct a data set from feature data and corresponding label data, and to partition the data set into a training set and a validation set; A training model module, used to create a recurrent neural network model, input a training set into the recurrent neural network model, calculate the loss and update the parameters of the recurrent neural network model; The output module is used to run the updated recurrent neural network model on the validation set, obtain the logits output, and generate the Gumbel Softmax output based on the logits output: in, represents the connection probability of sensor nodes i and j in the avionics system, Represents the output probability of the recurrent neural network model, that is, logits output, represents the noise generated from Gumbel distribution, τ represents the temperature parameter; Building module for constructing sensor network topology graph based on Gumbel Softmax output.

7. The avionics system fault diagnosis system based on graph anomaly detection according to claim 5, characterized in that: The updating embedded representation module comprises: The learning embedding information module is used to learn the embedding information of all layer nodes through the learning model. The specific learning model is: in, represents the correlation measure between the i-th sensor node and the j-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the i-th sensor node of the avionics system in the k-th layer, represents the embedded representation of the jth sensor node of the avionics system in the kth layer, f represents the non-negative decreasing function of z, and N represents the number of layers of the recurrent neural network; Wherein, T represents the transpose operation of the vector; The state update module is used to update the embedding vector of each layer node i through the state update model based on the learned embedding information of all layer nodes. Specifically, the state update model is: in, represents the updated embedding representation of the i-th sensor node of the avionics system in the k-th layer, and τ represents the temperature parameter.

8. The avionics system fault diagnosis system based on graph anomaly detection according to claim 5, characterized in that: The fault type determination module comprises: A reconstruction error obtaining module is used to obtain a reconstruction error according to the embedded representation of the sensor updated by the recurrent neural network model and the embedded representation of the sensor actually measured; an abnormal marking module, used for marking the sensor node as abnormal if the reconstruction error is greater than a second preset threshold; The judgment module is used to update the state of fault propagation using the neighborhood information of the sensor network topology map according to the marked anomaly, and then use the classifier to judge the fault type.

Images