An aeronautical sensor fault detection and classification method based on neural process network

Abstract

The application discloses an aviation sensor fault detection and classification method based on a neural process network, and belongs to the field of aviation technology, and comprises the following steps: S1, sampling and preprocessing the measurement values of a plurality of aviation sensors under different fault categories to obtain training images suitable for input into a neural process network; S2, constructing a neural process network, training the neural process network by using the training images and corresponding fault categories, and then obtaining an aviation sensor fault detection and classification model based on the neural process network; and S3, sampling and preprocessing the measurement values of a to-be-detected aviation sensor in the same manner to obtain a to-be-detected image, processing the to-be-detected image by using the model, and obtaining the fault category of the to-be-detected aviation sensor. The application has a faster convergence speed, a more stable loss value decline curve, a higher accuracy, and higher applicability and effectiveness.

Description

Technical Field

[0001] This invention relates to the field of aviation technology, and in particular to a method for fault detection and classification of aviation sensors based on neural process networks. Background Technology

[0002] In today's intelligent equipment condition monitoring tasks, deploying sensors allows for the acquisition of real-time signals generated by these devices. After preprocessing the collected signals, such as noise reduction, it becomes possible to diagnose whether the intelligent equipment has experienced a malfunction. In the aviation field, both civil airliners and military aircraft are equipped with numerous sensors. These sensors measure information such as speed and angle of attack, which are crucial for ensuring the safe operation of aircraft. However, these sensors are generally mounted on the outer surface of the aircraft and are prone to malfunction after exposure to rain, icing, and humidity. These malfunctions can severely impact the safe operation of the aircraft. Therefore, fault detection and classification technology for aviation sensors is extremely important for ensuring the safe operation of aircraft.

[0003] Existing airborne sensor fault detection and classification technologies can be divided into two categories: model-based detection methods and data-driven detection methods.

[0004] In model-based knowledge-based methods, some scholars have proposed using state-space models, Gaussian processes, and other models to analyze signals measured by airborne sensors, extracting and defining the residual between the model output and the sensor output, and then using classification algorithms to perform fault detection and classification tasks. This method requires analysis of sensor signals based on model knowledge, and the classification algorithm is generally based on decision trees or rules, requiring a threshold function to determine whether a sensor fault has occurred. This makes the classification method less reliant on large amounts of data, thus possessing the advantage of high feasibility; however, due to the limitations of knowledge / rules, the models or classification rules constructed using this method suffer from poor robustness and poor generalization.

[0005] In recent years, the rapid development of computer technology has provided new ideas for fault detection and classification research. Scholars have proposed a series of data-driven methods, including clustering, regression, support vector machines, association rules, and deep neural networks. Among them, methods represented by deep neural networks have been widely used in the field of fault detection and classification due to their advantages such as self-association, self-adaptation, and the fact that they do not require the establishment of precise mathematical models. They have achieved good scientific research progress and preliminary engineering applications.

[0006] Currently, representative deep neural networks include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Autoencoders, and Graph Neural Networks (GNNs). While fault detection and classification algorithms using deep neural networks have certain advantages, their drawbacks are also quite significant. Specifically, the input-output mapping calculation of deep neural networks is highly dependent on the network training process, resulting in poor generalization ability for diverse / heterogeneous scenarios and tasks.

[0007] Neural Processes (NPs) are a novel data-driven approach proposed and vigorously developed by DeepMind in recent years. Unlike the rigid input-output mapping of deep neural networks, NP networks employ a meta-learning architecture. During training, they use stochastic processes based on model knowledge to model multi-task scenarios and construct a deep neural network encoder based on diverse observation data to perform probabilistic inference. Finally, a deep neural network decoder outputs the distribution function of the stochastic process, achieving an input-output mapping based on stochastic processes and deep neural networks. It is evident that NPs combine the advantages of both model-knowledge-based and deep neural network-based methods, offering flexibility in training processes and data dimensionality configuration, as well as excellent model generalization performance. Currently, they have been widely applied in fields such as image classification, data regression, and natural language processing. Summary of the Invention

[0008] To address the problem that existing methods for fault detection and classification of aviation sensors have poor generalization capabilities for diverse / heterogeneous scenarios and tasks, the present invention aims to provide an aviation sensor fault detection and classification method based on neural process networks, so as to at least partially solve the above-mentioned problems.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows:

[0010] In a first aspect, the present invention provides a method for fault detection and classification of airborne sensors based on neural process networks, comprising the following steps:

[0011] S1. Sample and preprocess the measurement values ​​of several aviation sensors under different fault categories to obtain training images suitable for input into the neural network.

[0012] S2. Construct a neural process network, train the neural process network with the training images and corresponding fault categories, and then obtain an aviation sensor fault detection and classification model based on the neural process network.

[0013] S3. Sample and preprocess the measured values ​​of the airborne sensor under test in the same manner as in step S1 to obtain the image under test. Process the image under test using the model described in step S2 to obtain the fault category of the airborne sensor under test.

[0014] Preferably, in step S1, the step of sampling and preprocessing the measurement values ​​of several airborne sensors to obtain training images suitable for input into the neural network includes:

[0015] In the time dimension, the measurements of m airborne sensors are sampled and stacked in a historical window of duration n to obtain an n×m dimensional matrix;

[0016] The n×m dimensional matrix is ​​normalized and visualized to obtain a grayscale image;

[0017] The grayscale image is subjected to data augmentation processing to obtain the training image.

[0018] Preferably, the grayscale image is subjected to data enhancement processing using a mosaic method.

[0019] Preferably, in step S2, the step of constructing the neural process network specifically includes:

[0020] Set reference set:

[0021] Select N training images and their corresponding fault categories to form a set D, D = {(x1, y1)..., (x... N y N )}, where x represents the training image and y represents the fault category;

[0022] Select several reference points from the set D to form a reference set R, R = {x} r 1, ..., x r K}, and make set M = D\R and set B = R∪M;

[0023] Establish feature space mapping:

[0024] The Lenet-5 network is built using the network module provided by PyTorch. The Lenet-5 network acts as a function distribution mapper, mapping the input training images from a low-dimensional space to a high-dimensional space, so that the subsequent graph network can construct the required neural process network in the high-dimensional space.

[0025] Building a graph network:

[0026] By constructing a bipartite graph representing the dependencies between data points in R and data points in M, and a directed acyclic graph representing the dependencies between data points in R, the neural network described above has been constructed.

[0027] Preferably, the process of setting the reference set further includes: dividing the training images into a training set and a test set according to a preset ratio, wherein the set D is selected from the training set.

[0028] Preferably, after constructing the graph network, the following steps are also included:

[0029] Test the model using the test set:

[0030] Input the data from the test set into the trained model, activate the model's predictions using the Softmax function, and then call PyTorch's built-in function argmax to obtain the fault classification results.

[0031] Preferably, the Lenet-5 network comprises two two-dimensional convolutional layers and two pooling layers, using the ReLU function as the activation function and setting the pooling function to max pooling.

[0032] Preferably, in step S1, there are 15 aviation sensors that respectively collect the aircraft's speed, pitch angle, sideslip angle, roll rate, pitch rate, yaw rate, roll angle, pitch angle, yaw angle, lateral displacement, side displacement, vertical displacement, x-axis overload, y-axis overload, and z-axis overload; the fault categories include no fault, speed measurement drift, pitch angle measurement drift, sideslip angle measurement drift, pitch angle measurement noise, sideslip angle measurement noise, three-axis angular velocity measurement drift, three-axis overload measurement drift, three-axis angular velocity measurement noise, and three-axis overload measurement noise.

[0033] In a second aspect, the present invention also provides an electronic device, including a memory storing executable program code and a processor coupled to the memory; wherein the processor calls the executable program code stored in the memory to execute the method described above.

[0034] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the method described above.

[0035] The beneficial effects of this invention using the above technical solution are as follows: This invention employs a neural process network to model the airborne sensor fault detection and classification task as an inference problem based on stochastic processes. By establishing and preprocessing a diverse dataset of airborne sensor faults, the dataset is used as input to the neural process network. The output of the neural process network is set as the fault category of the airborne sensor, and the neural process network model is obtained through training. Compared to the current best airborne sensor fault detection and classification method based on VGG16 networks, the technical solution of this invention has a faster convergence speed and a more stable loss value descent curve for diverse task scenarios, and also achieves higher accuracy, demonstrating greater applicability and effectiveness in heterogeneous tasks. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method according to Embodiment 1 of the present invention;

[0037] Figure 2a This is a schematic diagram of data acquisition for speed, elevation angle, and sideslip angle in Embodiment 1 of the present invention;

[0038] Figure 2b This is a schematic diagram of data acquisition for the x-axis angular velocity, y-axis angular velocity, and z-axis angular velocity in Embodiment 1 of the present invention;

[0039] Figure 2c This is a schematic diagram of data acquisition for roll angle, pitch angle, and yaw angle in Embodiment 1 of the present invention;

[0040] Figure 2d This is a schematic diagram of data acquisition for x-coordinate, y-coordinate, and height coordinate in Embodiment 1 of the present invention;

[0041] Figure 2e This is a schematic diagram of data acquisition for x-axis overload, y-axis overload, and z-axis overload in Embodiment 1 of the present invention;

[0042] Figure 3 This is a schematic diagram showing the data stacking of the three dimensions of speed, elevation angle, and sideslip angle in Embodiment 1 of the present invention;

[0043] Figure 4 This is a schematic diagram of the data augmentation process in Embodiment 1 of the present invention;

[0044] Figure 5 This is a schematic diagram of the model structure of the neural process network constructed in Embodiment 1 of the present invention;

[0045] Figure 6 This is a schematic diagram illustrating the relationship between the various sets in Embodiment 1 of the present invention;

[0046] Figure 7 This is a schematic diagram of the model structure of the VGG16 network introduced in Embodiment 1 of the present invention;

[0047] Figure 8 This is a schematic diagram illustrating the accuracy of the VGG16 network model in Embodiment 1 of the present invention for detecting three types of aircraft datasets.

[0048] Figure 9 This is a schematic diagram illustrating the accuracy of the neural process network model in Embodiment 1 of the present invention in detecting three types of aircraft datasets.

[0049] Figure 10 This is a schematic diagram of the model structure of the AlexNet network introduced in Embodiment 1 of the present invention;

[0050] Figure 11 This is a schematic diagram showing the changes in loss values ​​of the three network models during the training process in Embodiment 1 of the present invention;

[0051] Figure 12 This is a schematic diagram of the electronic device in Embodiment 3 of the present invention. Detailed Implementation

[0052] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0053] Example 1

[0054] A method for fault detection and classification of airborne sensors based on neural process networks, such as Figure 1 As shown, it includes the following steps:

[0055] S1. Sample and preprocess the measurements of several airborne sensors under different fault categories to obtain training images suitable for input into the neural network.

[0056] The aviation sensor configuration is 15, such as Figures 2a-2e As shown, these parameters are used to measure the aircraft's speed v, pitch angle α, sideslip angle β, and roll angular velocity (x-axis angular velocity) ω, respectively. x Pitch angular velocity (y-axis angular velocity) ω y Yaw angular velocity (z-axis angular velocity) ω z Roll angle, pitch angle, yaw angle, lateral displacement (represented by the x-axis), side displacement (represented by the y-axis), vertical displacement (represented by the height coordinate), and x-axis overload G. x y-direction overload G y Z-direction overload G z .

[0057] The aircraft targeted by the airborne sensors are the Y-8, B747, and DH-6. The overview of the fault flight databases for these three aircraft is shown in Table 1. Various flight states were employed, including pilot control, autopilot control, low-altitude takeoff and landing, and high-altitude cruise. Data for the B747 was measured and recorded by onboard airborne sensors during actual flight; data for the Y-8 and DH-6 were based on flight simulation calculations. In this embodiment, the total flight times for the Y-8, B747, and DH-6 are 520 minutes, 516 minutes, and 867 minutes, respectively.

[0058] Table 1 - Overview of the Aircraft Fault Flight Database

[0059]

[0060]

[0061] For the same aircraft, considering all 15 aviation sensors it carries, the fault categories of the aviation sensors are divided into 10 categories, denoted as 0-9, as shown in Table 2. They are: all 15 aviation sensors are fault-free [0], velocity measurement drift [1], pitch angle measurement drift [2], sideslip angle measurement drift [3], pitch angle measurement noise [4], sideslip angle measurement noise [5], three-axis angular velocity measurement drift [6], three-axis overload measurement drift [7], three-axis angular velocity measurement noise [8], and three-axis overload measurement noise [9]. It is assumed that the three aviation sensors used to measure overload (or the three aviation sensors used to measure angular velocity) simultaneously experience the same type of fault, i.e., fault category [6]-[9].

[0062] Table 2 - Fault Categories and Amplitude Definitions for Aircraft Sensors

[0063]

[0064] In this embodiment, the steps of sampling and preprocessing the measurement values ​​of 15 airborne sensors to obtain training images suitable for input into the neural network specifically include:

[0065] S11. Sample and stack the measurements of 15 airborne sensors in a historical window with a duration of 30 seconds in the time dimension (sampling frequency 1Hz) to obtain a 15×31 dimensional matrix.

[0066] like Figure 3 As shown, this diagram illustrates how a 3×31 matrix is ​​obtained by sampling and stacking the three dimensions: velocity v, elevation angle α, and sideslip angle β. Similarly, a 15×31 matrix can be obtained by sampling and stacking 15 dimensions.

[0067] S12. Normalize and visualize the 15×31 dimensional matrix to obtain a grayscale image.

[0068] S13. Perform data augmentation processing on the grayscale image to obtain the training image.

[0069] In this embodiment, the image format suitable for input to the neural process network is 224×224 dimensions. Therefore, the grayscale image is data augmented using the mosaic method to transform it from 15×31 dimensions to 224×224 dimensions, so that the neural process network can fully extract the channel features of the measurement values ​​of each aviation sensor.

[0070] Specifically, such as Figure 4 As shown in Equation 1, M represents a grayscale image, and its subscripts m and n represent the row and column numbers corresponding to the elements, respectively. In Equation 1, m = 15 and n = 31. In the mosaic image method, each element in M ​​is repeated a times along the column and row directions, respectively (a = 14, b = 7), and then zeros are added to the matrix edges to expand it to a 224×224 dimension.

[0071]

[0072] S2. Construct a neural process network. Train the neural process network using training images and their corresponding fault categories to obtain an aviation sensor fault detection and classification model based on the neural process network.

[0073] The hardware configuration required for the neural process network constructed in this embodiment to run is as follows: CPU model i9-11900K, GPU model Nvidia RTX3090, and 64GB of memory; the software environment is Python 3.8 (Windows 10), PyTorch version 1.9.0, and CUDA version 11.4.

[0074] This embodiment constructs a neural process network as follows: Figure 5 As shown, the specific construction steps include: setting a reference set, establishing a feature space mapping, and constructing a graph network.

[0075] Set the reference set, including:

[0076] Select N training images and their corresponding fault categories to form a set D, D = {(x1, y1)..., (x... N y N )}, where x represents the training image (as input to the neural network) and y represents the fault category (as output to the neural network);

[0077] To reduce computational complexity, the sparse Gaussian process uses induced input to partially replace the original input for hypothesis and training. Several reference points are selected from set D to form a reference set R, where R = {x}. r 1, ..., x r K For example, K is 300, and set M = D\R and set B = R∪M.

[0078] like Figure 6 As shown, it specifically illustrates the relationships between the various sets. The large circle on the left is set D, the small circle on the right is set R, the part that belongs to set D but not to set R is set M, and set B is the union of set D and set R.

[0079] Establishing a feature space mapping includes:

[0080] The Lenet-5 network is built using the network module provided by PyTorch. The Lenet-5 network acts as a function distribution mapper, mapping the input training images from a low-dimensional space (1×224×224) to a high-dimensional space (feature vector length is 500), so that the subsequent graph network can construct the required neural process network in the high-dimensional space.

[0081] The Lenet-5 network consists of two 2D convolutional layers and two pooling layers, using the ReLU function as the activation function and max pooling as the pooling function.

[0082] Constructing a graph network includes:

[0083] By constructing a bipartite graph representing the dependencies between data points in R and data points in M, and a directed acyclic graph representing the dependencies between data points in R, the neural process network has been constructed.

[0084] like Figure 5 As shown, the relationship between data point 6 and data point 5 can be obtained through the bipartite graph. Similarly, after constructing the directed acyclic graph (DAG) between data points in R, the following dependencies can be obtained through the DAG: ① Data point 3 depends on data point 5, and data point 5 depends on data point 1; ② Data point 6 depends on data point 2, and data point 2 depends on data point 4.

[0085] S3. Sample and preprocess the measured values ​​of the airborne sensor under test in the same manner as in step S1 to obtain the image under test. Process the image under test using the model obtained in step S2 to obtain the fault category of the airborne sensor under test.

[0086] The image under test is also 224×224 in dimensions, and in constructing such... Figure 5Once the neural network model shown is trained, it can be used to process the image under test. The output of the model is the fault detection and classification result of the aviation sensor. By referring to the contents of Table 2, you can know what kind of fault the aviation sensor has experienced.

[0087] Comparative analysis of test results:

[0088] This invention introduces the VGG16 network and the AlexNet network to replace the neural network, respectively, for fault detection and classification of airborne sensors. The model structure of the VGG16 network is as follows: Figure 7 As shown, experiments were conducted using the VGG16 network model on datasets corresponding to the B747, DH6, and Y8 aircraft, and the results are as follows. Figure 8 As shown in the figure, for the DH6 and Y8 aircraft datasets, since both aircraft types are flight simulation data, the classification difficulty is low, and the VGG16 network model achieves the best results. However, for the B747 aircraft dataset, due to the presence of more noise and other interference during actual flight, the classification accuracy of the VGG16 network model decreases, and its accuracy is still less than 95% at the 30th iteration.

[0089] The neural network model constructed in this embodiment of the invention was tested on datasets corresponding to the B747, DH6, and Y8 aircraft, and the results are as follows: Figure 9 As shown in the figure, compared with the VGG16 network model, the neural process network model can achieve high accuracy (greater than 95%) for the DH6 and Y8 aircraft datasets. At the same time, the neural process network model also has excellent classification results for the B747 aircraft dataset, with an accuracy of 95% by the 16th iteration.

[0090] This invention also introduces the AlexNet network to replace the neural process network for fault detection and classification of aviation sensors. The model structure of the AlexNet network is as follows: Figure 10 As shown. The changes in loss values ​​during training for the neural network, VGG16 network, and AlexNet network disclosed in this embodiment are statistically analyzed and plotted, as follows. Figure 11As shown in the figure, the neural process network model has the fastest convergence speed. When the number of iterations is 5, the loss value of the neural process network model (0.601) is much smaller than that of the VGG16 network model (1.409) and the AlexNet network model (2.167). Moreover, when the number of iterations reaches 8, the detection accuracy of the neural process network model for the datasets corresponding to the three types of aircraft is greater than 90%. This verifies that the neural process network model has the characteristics of rapid transfer training and application, thus meeting the requirements of real-world aviation sensor fault detection tasks.

[0091] Example 2

[0092] In step S2, when setting the reference set, all training images are usually divided into training set and test set according to a preset ratio (2:1 or 7:3, etc.), and the aforementioned set D is selected from the training set.

[0093] After constructing the graph network, the following steps are also included:

[0094] Test the model using the test set:

[0095] That is, the data in the test set is input into the trained model, the Softmax function is used to activate the model's predicted values, and then the built-in function argmax in PyTorch is called to obtain the fault detection and classification results.

[0096] After obtaining the fault detection and classification results output by the model, it is convenient to perform accuracy analysis and parameter tuning on the model's output results based on the known fault categories.

[0097] Furthermore, after obtaining the aforementioned dependency graph, the dependencies between data points can be extracted and the posterior distribution of the data can be approximated after training. That is, after the model training is completed, the approximate posterior distribution of the predicted dependency graph is used as the prior for the testing phase.

[0098] Example 3

[0099] An electronic device, such as Figure 12 As shown, it includes a memory storing executable program code and a processor coupled to the memory; wherein the processor calls the executable program code stored in the memory to execute the method steps disclosed in the above embodiments.

[0100] Example 4

[0101] A computer storage medium storing a computer program, wherein the computer program is executed by a processor to perform the method steps disclosed in the above embodiments.

[0102] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for fault detection and classification of airborne sensors based on neural process networks, characterized in that: Includes the following steps: S1. Sample and preprocess the measurement values ​​of several aviation sensors under different fault categories to obtain training images suitable for input into the neural network. S2. Construct a neural process network, train the neural process network with the training images and corresponding fault categories, and then obtain an aviation sensor fault detection and classification model based on the neural process network. S3. Sample and preprocess the measured values ​​of the airborne sensor under test in the same manner as in step S1 to obtain the image under test. Process the image under test using the model described in step S2 to obtain the fault category of the airborne sensor under test. In step S2, the step of constructing the neural process network specifically includes: Setting a reference set: Select N A set consists of training images and their corresponding fault categories. D , D ={( x 1, y 1)...,( x N , y N )}, x Represents training images, y Indicate the fault category; Select several reference points from the set D to form a reference set. R , R ={x r 1, ..., x r K }, and make the set M = D \ R ,gather B = R ∪ M ; Establish feature space mapping: use PyTorch The provided network module is used to build a Lenet-5 network. The Lenet-5 network acts as a function distribution mapper, mapping the input training images from a low-dimensional space to a high-dimensional space, so that the subsequent graph network can construct the required neural process network in the high-dimensional space. Building a graph network: Construction Representation R Data points and M A bipartite graph showing the dependencies between data points, and the construction of its representation. R The directed acyclic graph of the dependencies between data points in the graph represents the completed construction of the neural process network.

2. The method according to claim 1, characterized in that: In step S1, the step of sampling and preprocessing the measurement values ​​of several airborne sensors to obtain training images suitable for input into the neural network includes: In terms of time dimension, using duration as... n Historical window m The measurements from several airborne sensors are sampled and stacked to obtain... n × m A 3D matrix; Regarding the n × m The matrix is ​​normalized and visualized to obtain a grayscale image; The grayscale image is subjected to data augmentation processing to obtain the training image.

3. The method according to claim 2, characterized in that: The grayscale image is augmented using a mosaic method.

4. The method according to claim 1, characterized in that: The process of setting the reference set also includes: dividing the training images into a training set and a test set according to a preset ratio. D It is then selected from the training set.

5. The method according to claim 4, characterized in that: After constructing the graph network, the following is also included: Test the model using the test set: Input the data from the test set into the trained model, and select... Softmax The function activates the model's predicted values ​​and then calls... PyTorch Built-in functions argmax The fault detection and classification results are obtained.

6. The method according to claim 1, characterized in that: The Lenet-5 network comprises two 2D convolutional layers and two pooling layers, using... RELU The function is used as the activation function, and the pooling function is set to max pooling.

7. The method according to claim 1, characterized in that: In step S1, the aircraft sensors comprise 15 units and collect data on the aircraft's speed, pitch angle, sideslip angle, roll rate, pitch rate, yaw rate, roll angle, pitch angle, yaw angle, lateral displacement, side displacement, and vertical displacement. x To overload, y To overload, z Overload; the fault categories include no fault, speed measurement drift, pitch angle measurement drift, sideslip angle measurement drift, pitch angle measurement noise, sideslip angle measurement noise, triaxial angular velocity measurement drift, triaxial overload measurement drift, triaxial angular velocity measurement noise, and triaxial overload measurement noise.

8. An electronic device, characterized in that: The method includes a memory storing executable program code and a processor coupled to the memory; wherein the processor invokes the executable program code stored in the memory to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program is executed by the processor to perform the method as described in any one of claims 1-7.

Images