An abnormal detection method for eVTOL flight control based on dynamic graph cross-modal reconstruction and prototype comparison learning

Abstract

This invention provides an anomaly detection method for eVTOL flight control systems based on dynamic graph cross-modal reconstruction and prototype comparative learning, relating to the field of aircraft monitoring. It includes acquiring multimodal sensor data during eVTOL flight and performing feature fusion; constructing a dynamic graph structure to represent the relationships between various flight control systems and extracting spatiotemporal state features using a dynamic graph neural network; reconstructing masked modes using a cross-modal reconstruction network to obtain reconstructed state representations and multi-scale reconstruction errors; constructing a normal flight state prototype library and optimizing the state representation space using a prototype comparative learning method; calculating anomaly scores based on the distribution deviation between the actual and reconstructed states, and completing anomaly detection and localization based on the anomaly scores. This invention achieves accurate identification and fault localization of eVTOL flight control anomalies by constructing a collaborative framework of dynamic graph modeling, cross-modal reconstruction, and prototype comparative learning, utilizing the correlation and state evolution laws between multi-source sensing data.

Description

Technical Field

[0001] This invention relates to the field of aircraft monitoring technology, specifically to an eVTOL flight control state anomaly detection method based on dynamic graph cross-modal reconstruction and prototype comparison learning. Background Technology

[0002] With the rapid development of the low-altitude economy, electric vertical take-off and landing (eVTOL) aircraft have shown broad application prospects in urban air traffic, emergency rescue, logistics transportation, and inspection due to their advantages such as vertical take-off and landing, environmental friendliness, and flexible deployment. However, eVTOL flight involves multiple key subsystems, including power systems, energy systems, navigation systems, and flight control systems, and these systems have complex coupling relationships; their operational status directly affects flight safety.

[0003] To ensure the safe and stable operation of aircraft, flight control anomaly detection has become a key technology in eVTOL operational status monitoring and health management. However, most existing anomaly detection methods are based on time-series prediction frameworks, which use historical flight state sequences to predict future states and determine whether anomalies have occurred based on prediction errors. But in real-world complex flight environments, factors such as airflow disturbances, sensor noise, system parameter drift, and dynamic changes in multi-subsystem coupling can easily lead to error accumulation in prediction models. This makes the mapping relationship between prediction errors and actual abnormal states unstable, reducing the accuracy and robustness of anomaly detection.

[0004] Furthermore, existing methods typically rely on single-modal data or employ simple feature stitching for multi-source information fusion, failing to fully model the inherent physical relationships between the flight control system, propulsion system, energy system, and navigation system, resulting in insufficient collaborative representation capabilities among multimodal information. Therefore, how to fully utilize the correlations between multimodal sensor data to construct a high-precision, robust flight control anomaly detection model, and achieve anomaly identification and fault location in complex flight environments, has become a pressing technical problem to be solved in the field of eVTOL intelligent monitoring. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an eVTOL flight control state anomaly detection method based on dynamic graph cross-modal reconstruction and prototype comparison learning, which solves the problems of insufficient multimodal correlation modeling of aircraft, significant accumulation of prediction errors, and limited ability to distinguish abnormal states.

[0006] To achieve the above objectives, this invention provides an eVTOL flight control state anomaly detection method based on dynamic graph cross-modal reconstruction and prototype comparison learning, comprising: Acquire multimodal sensor data during eVTOL flight, preprocess it, and construct a fused feature sequence; A dynamic graph structure is constructed based on the fused feature sequence, and a dynamic graph neural network is used to extract flight state representations; Modality masking is performed on the fused feature sequence, and the masked modality is reconstructed using a cross-modal reconstruction network to generate a reconstructed state representation; Calculate multi-scale reconstruction error based on the real state representation and reconstructed state representation corresponding to the flight state representation; A normal flight state prototype library is constructed based on the set of flight state representations, and the state representation space composed of flight state representations is optimized by using a prototype contrastive learning method. Anomaly scores are calculated based on the deviation between the distribution of true states in the true state representation set and the distribution of reconstructed states in the reconstructed state representation set. Anomaly detection and anomaly localization are performed based on anomaly scoring, and the anomaly detection results of the eVTOL flight control status are output.

[0007] In one embodiment of the present invention, the multimodal sensor data includes one or more of the following: video image data, inertial measurement unit data, global positioning system data, motor operating parameters, battery operating parameters, flight control commands, wind speed sensor data, barometric pressure sensor data, and attitude angle data.

[0008] In one embodiment of the present invention, the method of acquiring multimodal sensor data during eVTOL flight and performing preprocessing to construct a fused feature sequence includes: Outlier removal and missing value completion are performed on the multimodal sensor data to obtain the data for each modality; The data is classified and processed based on the data type of each modality. Continuous data is normalized using Min-Max, and discrete / categorical data is represented by one-hot encoding or embedding. Time synchronization is performed based on a unified time base. The synchronized modal data are spliced ​​and fused in chronological order to obtain a fused feature sequence.

[0009] In one embodiment of the present invention, a dynamic graph structure is constructed based on fused feature sequences, and a dynamic graph neural network is used to extract flight state representations, including: Modal features corresponding to the flight control system, power system, energy system and navigation system are extracted from the fused feature sequence and used as the initial state features of the graph structure nodes respectively; Fixed directed edge connections are constructed based on the physical coupling relationships between the systems; Based on the graph attention mechanism, the association weights between nodes are dynamically learned according to the current features of each node, and the node representation is iteratively updated by aggregating neighbor information; A dynamic graph neural network is used to output a flight state representation that integrates spatiotemporal information.

[0010] In one embodiment of the present invention, a modality masking operation is performed on the fused feature sequence, and a cross-modal reconstruction network is used to reconstruct the masked modality to generate a reconstructed state representation, including: Randomly select at least one modality to perform a masking operation; A cross-modal reconstruction network is used to reconstruct the masked modes and output the reconstructed modal features; Reconstruction state representation is generated based on reconstruction modal characteristics.

[0011] In one embodiment of the present invention, multi-scale reconstruction error is calculated based on the real state representation and the reconstructed state representation corresponding to the flight state representation, including: Calculate point-level reconstruction error, window-level reconstruction error, and trend-level reconstruction error; A comprehensive reconstruction error is constructed based on point-level reconstruction error, window-level reconstruction error, and trend-level reconstruction error.

[0012] In one embodiment of the present invention, a normal flight state prototype library is constructed based on a set of flight state representations, and a prototype contrastive learning method is used to optimize the state representation space constituted by the flight state representations, including: Clustering algorithms were used to perform cluster analysis on samples in normal flight conditions; State prototypes are constructed based on each cluster center. The state prototypes include takeoff state prototype, hovering state prototype, cruise state prototype, and landing state prototype. Based on state prototypes, the flight state representation is constrained by prototype contrast loss, which enables normal states to cluster into corresponding state prototypes, thereby forming state prototype features with prototype distinguishability in the state representation space.

[0013] In one embodiment of the present invention, based on state prototypes, constraint learning is performed on the flight state representation through prototype contrast loss, causing normal states to cluster towards corresponding state prototypes, thereby forming state prototype features with prototype discriminativeness in the state representation space, including: A state prototype library was constructed based on normal flight state samples; Calculate flight state characterization With state prototype Cosine similarity between them; A prototype contrast loss function is constructed based on the cosine similarity between the flight state representation and the state prototype.

[0014] In one embodiment of the present invention, calculating anomaly scores based on the deviation between the true state distribution in the true state representation set and the reconstructed state distribution in the reconstructed state representation set includes: Calculate the Wasserstein distance based on the true state distribution and the reconstructed state distribution; The overall reconstruction error is calculated based on the sum of the multi-scale reconstruction errors. Construct an anomaly scoring function using the Wasserstein distance; According to the anomaly score It is compared with a preset threshold δ, when the condition is met The system determines the current flight status as abnormal and outputs the corresponding abnormality detection result. The anomaly contribution is calculated by using the attention weights of each node in the dynamic graph neural network. The node with the largest contribution is identified as the source node of the anomaly, thus achieving anomaly localization.

[0015] The present invention also provides an electronic device, including a memory and a processor. The memory is a computer-readable storage medium and stores program instructions. The processor is used to run the program instructions to execute the above-described method for detecting anomalies in eVTOL flight control based on dynamic graph cross-modal reconstruction and prototype comparison learning.

[0016] This invention provides an anomaly detection method for eVTOL flight control systems based on dynamic graph cross-modal reconstruction and prototype comparison learning. Compared with existing technologies, it has the following advantages: 1. This invention constructs a dynamic graph structure between the flight control system, power system, energy system and navigation system through a dynamic graph neural network, realizes joint modeling of multimodal state information, can fully explore the dynamic correlation between systems, and improve the flight state representation capability.

[0017] 2. This invention adopts a cross-modal reconstruction mechanism, which utilizes unmasked modes to reconstruct masked modes, effectively learning the cooperative relationship between different modes. Compared with traditional future state prediction methods, it can reduce error accumulation and improve the robustness of anomaly detection.

[0018] 3. This invention constructs a normal flight state prototype library and introduces a prototype comparison learning mechanism, which enables normal states to cluster towards corresponding prototypes, different flight conditions to separate from each other, and abnormal states to be judged because they are far away from all prototypes, thereby enhancing the discriminative ability of the state representation space.

[0019] 4. This invention combines multi-scale reconstruction error and Wasserstein distance to construct an anomaly scoring function, and combines graph node attention weights to locate the anomaly subsystem, thereby achieving accurate identification of anomaly states and location of anomaly sources, significantly improving the accuracy and generalization ability of flight control state anomaly detection. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of the eVTOL flight control state anomaly detection method based on dynamic graph cross-modal reconstruction and prototype comparison learning provided in an embodiment of the present invention; Figure 2 This diagram illustrates the steps for detecting anomalies in eVTOL flight control systems based on dynamic graph cross-modal reconstruction and prototype comparison learning, as provided in this embodiment of the invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This application uses a quadcopter manned eVTOL as the experimental verification platform. The entire machine is equipped with eight types of modal data sources, including IMU, GPS, brushless motor, power battery, wind speed, barometric pressure sensors, and flight control command acquisition module. The onboard edge computing unit has a computing power of 8 TOPS, meeting the requirements for onboard real-time inference. This solution fully covers the two major engineering stages: offline model training and onboard online inference.

[0024] This application provides an eVTOL flight control state anomaly detection method based on dynamic graph cross-modal reconstruction and prototype comparison learning, which specifically addresses the three core defects of existing eVTOL flight control anomaly detection schemes: weak multi-subsystem multi-modal coupling and correlation modeling capability, continuous accumulation of time series prediction framework errors, and insufficient feature representation space discrimination, thereby significantly improving the accuracy of fault identification and location in complex low-altitude scenarios.

[0025] Current mainstream eVTOL flight control anomaly detection solutions use timing prediction errors as the basis for anomaly identification. These solutions exhibit significant shortcomings under complex operating conditions such as urban low-altitude turbulence, sensor measurement noise, and battery / motor parameter drift. These shortcomings can be summarized in three points: (1) Time-series prediction models rely on historical sequences to extrapolate future flight states. Airflow disturbances, sensor noise, and nonlinear coupling of the system will cause prediction errors to accumulate frame by frame. Under normal operating conditions, false alarms are likely to occur, and the identification of minor faults is delayed, resulting in poor stability of anomaly judgment results.

[0026] (2) Traditional multi-source data fusion methods only perform simple dimensional splicing of various sensor features, without explicitly modeling the inherent physical coupling and time-varying correlation of the four major subsystems of flight control, power, energy and navigation, thus losing the constraint relationship between multimodal systems and weakening the effectiveness of feature expression.

[0027] (3) Conventional contrastive learning methods rely on perturbation methods such as random masking and random noise to construct positive and negative samples. They do not combine the four inherent working conditions of eVTOL takeoff, hovering, cruise and landing to divide the samples. The boundaries of the state representation space clusters obtained by training are blurred, and the ability to distinguish between normal samples and fault samples is weak, which directly reduces the anomaly recognition recall rate.

[0028] To address the aforementioned issues, this invention proposes an eVTOL flight control anomaly detection method based on dynamic graph cross-modal reconstruction and prototype comparative learning. This method avoids the accumulation of time-series prediction errors through modal reconstruction, quantifies modal mismatch using point, window, and trend-based error quantification, generates standardized state prototypes based on real flight conditions to optimize the representation space, and finally integrates multi-scale reconstruction errors and distribution distances to calculate anomaly scores. This enables accurate identification of eVTOL flight control anomalies and tracing the source of faulty subsystems. The following sections provide all feasible engineering parameters, value ranges, default configurations, and parameter selection criteria for each step.

[0029] This application addresses the technical pain points of existing multimodal correlation modeling, such as lack of quantitative parameter support, accumulation of time-series errors, and insufficient differentiation of operating conditions. The overall process is divided into the following steps, which refine the engineering implementation and the definition of all parameters step by step.

[0030] like Figure 1 As shown in the figure, an eVTOL flight control state anomaly detection method based on dynamic graph cross-modal reconstruction and prototype comparison learning provided by an embodiment of the present invention includes: S100 acquires multimodal sensor data during eVTOL flight, performs preprocessing, and constructs a fused feature sequence.

[0031] Specifically, the system acquires multimodal sensor data from IMU, GPS, motors, batteries, and flight control commands; performs Min-Max normalization on continuous data and one-hot encoding on discrete data; synchronizes each mode with a high-precision clock and removes misaligned frames; and constructs a fused feature sequence by splicing synchronization features over time using a fixed-length sliding window.

[0032] S200: A dynamic graph structure is constructed based on the fused feature sequence, and a dynamic graph neural network is used to extract flight state representations.

[0033] Specifically, by mapping flight control, power, energy, and navigation systems to graph nodes, directed edges are established based on physical dependencies; the graph attention mechanism is used to dynamically calculate edge weights based on the current features of nodes, aggregate neighbor information, and iteratively update node representations, and finally, the dynamic graph neural network outputs a flight state representation that integrates spatiotemporal information.

[0034] S300. Perform modal masking operation on the fused feature sequence, and use a cross-modal reconstruction network to reconstruct the masked modality and generate a reconstructed state representation.

[0035] Specifically, by randomly masking at least one modality (setting it to zero), the masked fused features are input into the cross-modal reconstruction network, and the masked modality is reconstructed using the unmasked modality. The reconstructed modal features are then combined with the unmasked real modal features to generate a complete reconstructed state representation.

[0036] S400: Calculate multi-scale reconstruction error based on the real state representation and reconstructed state representation corresponding to the flight state representation.

[0037] Specifically, based on the real state representation and the reconstructed state representation, the point-level, window-level, and trend-level reconstruction errors are calculated respectively, and then weighted and summed according to preset weights (α, β, γ) to obtain the multi-scale comprehensive reconstruction error, which is used to characterize the degree of modal mismatch from three levels: instantaneous, local, and trend.

[0038] S500 constructs a normal flight state prototype library based on the set of flight state representations, and uses a prototype comparison learning method to optimize the state representation space composed of flight state representations.

[0039] Specifically, by collecting a large number of normal flight state representations, a state prototype library covering takeoff, hovering, cruise, landing and other operating conditions is generated using a clustering algorithm; by calculating the cosine similarity between the state representations and each prototype and minimizing the prototype contrast loss, normal states are clustered toward their corresponding prototypes and different operating conditions are separated from each other, thus optimizing the distinguishability of the state representation space.

[0040] S600. Calculate the anomaly score based on the deviation between the true state distribution in the true state representation set and the reconstructed state distribution in the reconstructed state representation set.

[0041] Specifically, based on the comprehensive reconstruction error, the Wasserstein distance between the true state representation distribution and the reconstructed state representation distribution is calculated; the two are then fused through a balancing coefficient λ to construct an anomaly scoring function that simultaneously reflects local reconstruction bias and global distribution deviation.

[0042] S700 performs anomaly detection and anomaly localization processing based on anomaly scoring, and outputs the eVTOL flight control status anomaly detection results.

[0043] Specifically, by comparing the anomaly score with a preset threshold δ, anomalies are determined to be abnormal if they exceed the threshold. The attention weight of the dynamic graph nodes is used to locate the subsystem with the highest contribution as the fault source, and the final output is a structured detection result containing the anomaly type, timestamp, level, confidence level and faulty subsystem.

[0044] This invention uses a dynamic graph modeling and cross-modal reconstruction collaborative framework to learn the modal relationships within a flight system without the need for future state prediction, and combines prototype comparative learning to enhance the ability to distinguish between normal and abnormal states, thereby improving anomaly detection performance.

[0045] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0046] In step S100, multimodal sensor data during eVTOL flight is acquired, preprocessed, and a fused feature sequence is constructed.

[0047] First, it should be noted that this invention is applicable to all types of vertical takeoff and landing (VTOL) aircraft, and is not limited to a specific eVTOL model. For example, the multimodal sensing data used in this embodiment includes: inertial measurement unit (IMU) three-axis acceleration and angular velocity data; GPS positioning and ground speed data; motor three-phase current, bus voltage, speed, and winding temperature operating parameters; power battery cell voltage, total circuit current, and SOC status parameters; flight control PWM control commands and flight gear discrete commands; external wind speed sensor data; barometric altitude sensor data; and three-axis attitude angle data of the airframe.

[0048] Step S100 also includes the following steps.

[0049] S110. Perform outlier removal and missing value completion on the multimodal sensor data to obtain the data for each modality.

[0050] That is, for the acquired raw flight data, outlier removal and missing value completion are performed first to obtain a complete data sequence for subsequent processing steps.

[0051] S120. Classify the data based on the data type of each modality, use Min-Max normalization for continuous data, use one-hot encoding or embedding for discrete / categorical data, and perform time synchronization operation according to a unified time base.

[0052] Specifically, different modal data are classified and processed according to their data types. For example, continuous data (IMU, flight speed, motor current, bus voltage, etc.) are normalized using the Min-Max normalization method, mapping each dimension of data to the [0,1] interval to eliminate the interference of differences in sensor dimensions on model training and feature calculation; discrete / categorical data (flight control gear commands, flight status labels) are represented by one-hot encoding or low-dimensional embedding, without performing normalization scaling operations.

[0053] Subsequently, full-modal timing synchronization was performed based on the unified high-precision hardware time base of the digital twin. The timing alignment error tolerance threshold was set to ±1ms, and frames with timing misalignment exceeding the threshold were marked as invalid samples and discarded directly. After timing alignment was completed, timing samples were constructed using a fixed-length sliding window method. The mathematical expression of the window sample is as follows: in, The sliding time window length; taking the project configuration in this embodiment as an example, the window length is taken as... , corresponding to 1 second of flight data, for the first Each sampling time step is used to construct a window sample. Standardize the subscript representation; all time series samples are represented by the current time. Cut off the window from the end point. A continuous historical sampling point.

[0054] S130. The synchronized modal data are spliced ​​and fused in chronological order to obtain a fused feature sequence.

[0055] Finally, all modal features at the same time step within the window are horizontally concatenated and fused in chronological order to generate a standardized fused feature sequence. This serves as the input data for subsequent dynamic graph network modeling.

[0056] In step S200, a dynamic graph structure is constructed based on the fused feature sequence, and a dynamic graph neural network is used to extract flight state representations. Step S200 also includes the following steps.

[0057] S210. Extract the modal features corresponding to the flight control system, power system, energy system and navigation system from the fused feature sequence, and use them as the initial state features of the graph structure nodes.

[0058] Specifically, the eVTOL flight operation comprises four core subsystems: flight control system, power system, energy system, and navigation system. These subsystems are strongly physically coupled. Exemplary coupling logic includes: changes in flight control commands directly adjust the motor output speed; motor speed changes lift and torque, thus affecting the aircraft's flight attitude; the aircraft attitude synchronously updates the navigation and positioning status; and the continuous operation of the motors consumes battery power. Traditional one-dimensional time-series models can only independently model single-modal time-series changes and cannot explicitly depict the dynamic coupling relationships across subsystems. Therefore, this invention models the flight system at any given time as a time-varying dynamic graph topology. Represents the set of graph nodes. This represents the set of directed edges between nodes.

[0059] Specifically, the flight control system, propulsion system, energy system, and navigation system are mapped as nodes in a graph structure: in, Indicates the flight control node; Indicates nodes in the power system; Represents nodes in an energy system; This represents a navigation system node.

[0060] S220. Construct fixed directed edge connections based on the physical coupling relationships between the systems.

[0061] Specifically, the edges between nodes are used to describe the physical dependencies between different systems. In this embodiment, the fixed set of directed edges is: The physical meaning of the directed edge is as follows: the flight control system sends control commands to drive the power motor to run; the motor works continuously and consumes the power battery energy; the motor outputs power to change the aircraft's attitude and navigation position.

[0062] S230. Based on the graph attention mechanism, the association weights between nodes are dynamically learned according to the current features of each node, and the node representation is iteratively updated by aggregating neighbor information.

[0063] Specifically, the coupling strength between subsystems changes in real time under different operating conditions such as takeoff, hovering, cruise, and landing. A graph attention mechanism is introduced to dynamically update the edge weights between nodes step-by-step. Let the... At any given moment, nodes in the graph structure The corresponding original state feature vector is ,node The corresponding original state feature vector is For learnable query, key projection weight matrix, This serves as the unified dimension for the node feature vectors. The range of engineering values ​​is The default configuration in this embodiment This is used to balance airborne inference computing power and feature representation capability; then the first... Time Node With nodes The formula for calculating the original attention score of the edges between them is: The original attention score is normalized based on the node neighbor set to obtain Time Node Normalized dynamic correlation weights between The calculation method is as follows: in: Represents a node At any moment Corresponding state characteristics; and This represents the learnable parameter matrix. Then, a graph attention mechanism is used to aggregate neighbor node information: in: Represents a node The corresponding set of neighboring nodes; Represents the feature mapping matrix; This represents a nonlinear activation function. After iteration through a multi-layer dynamic graph neural network, the flight state representation is obtained: in: Indicates time The corresponding dynamic graph state representation.

[0064] In this way, the dynamic coupling strength between the flight control, power, energy and navigation systems can be plotted in real time, so that the node features can adaptively reflect the interaction between the subsystems under the current operating conditions, thereby enhancing the spatiotemporal sensitivity and modeling accuracy of the flight state representation.

[0065] S240 uses a dynamic graph neural network to output a flight state representation that integrates spatiotemporal information.

[0066] Specifically, by using the node features of the flight control, propulsion, energy, and navigation systems as initial inputs, a graph attention mechanism is used to dynamically calculate the attention weights between nodes at each time step, capturing the real-time changes in the coupling strength between subsystems. Neighbor information is aggregated and node representations are updated through multi-layer graph convolution iterative processing. Finally, the features of each node are fused using a global readout function (such as mean pooling) to generate a flight state representation (i.e., a dynamic graph state representation) that includes both the spatial topological relationships of multiple systems and the temporal evolution of states. In other words, the dynamic graph neural network can fully utilize the correlations between different systems, improving the ability to model flight states.

[0067] In step S300, a modality masking operation is performed on the fused feature sequence, and the masked modality is reconstructed using a cross-modal reconstruction network to generate a reconstructed state representation. Step S300 also includes the following steps.

[0068] S310. Randomly select at least one modality to perform a masking operation.

[0069] S320. A cross-modal reconstruction network is used to reconstruct the masked modes and output the reconstructed modal features.

[0070] S330. Generate a reconstruction state representation based on the reconstruction modal characteristics.

[0071] Specifically, during normal flight, stable correlations exist between different modes. For example, there are kinematic constraints between IMU data and GPS data; energy coupling between battery status and motor operating status; and control response between flight control commands and attitude changes. Therefore, by utilizing multiple unmasked modes to jointly reconstruct the masked mode, the intrinsic correlations between modes can be learned.

[0072] First, construct a multimodal state representation: .in: Indicates the number of modes; Indicates the first Each modality corresponds to a feature. Then, a modality is randomly selected for masking: where, This represents the total number of sensing modes in the entire device; the default configuration in this embodiment. ; Representing the The standardized feature vectors corresponding to each modality at time $t$. Then, a single modality is randomly selected for masking, with the masking positions uniformly filled with zero vectors. After obtaining the masked input feature set: The masked state representation is then input into the cross-modal reconstruction network: in: This represents a Cross-modal Consistency Learning Network. This represents the reconstructed modal characteristics. The reconstruction error is calculated using the mean squared error. When the flight condition is normal, the correlation between modes is stable, resulting in a small reconstruction error. However, when the flight condition is abnormal, the correlation between different modes is disrupted, leading to a significant increase in reconstruction error. Based on the reconstruction results, a reconstructed state representation is generated. Subsequent anomaly analysis is performed using the real state representation and the reconstructed state representation. The reconstructed modal features are directly replaced at the original masking positions and combined with the unmasked real modal features in the original modal order to form a reconstructed state representation with the same dimension as the real state representation.

[0073] In step S400, the multi-scale reconstruction error is calculated based on the actual state representation and the reconstructed state representation corresponding to the flight state representation. Step S400 also includes the following steps.

[0074] S410, calculate point-level reconstruction error, window-level reconstruction error, and trend-level reconstruction error.

[0075] Specifically, after step S3, the true state representation can be obtained. and reconstructed state representation It should be noted that relying solely on the reconstruction error at a single moment is insufficient to fully reflect abnormal flight conditions. For example, when subjected to short-term wind disturbances, a significant error may occur at a particular moment, while the aircraft as a whole remains in a normal state; conversely, for slowly evolving system faults, the error at a single moment may not be significant. Therefore, this invention analyzes the reconstruction error at three scales: point level, window level, and trend level.

[0076] First, calculate the point-level reconstruction error, which is expressed as: in, Indicates the number of samples; Represents the characteristics of the true state; This represents the characteristics of the reconstructed state. Point-level reconstruction error is used to characterize the degree of difference between the current real state and the reconstructed state.

[0077] The window-level reconstruction error is then calculated and expressed as follows: in, Indicates the length of the sliding time window; Indicates the first The point-level reconstruction error corresponds to each time step. The window-level reconstruction error can reflect the state changes within a local time period, thereby reducing the impact of random noise.

[0078] Further calculation of the trend-level reconstruction error, which is expressed as: in, This indicates the length of the time series. Trend-level reconstruction error is mainly used to measure the degree of deviation between the actual trend of change and the trend of change of the reconstructed state.

[0079] S420. Construct a comprehensive reconstruction error based on point-level reconstruction error, window-level reconstruction error, and trend-level reconstruction error.

[0080] Specifically, after obtaining the above three types of errors, a comprehensive reconstruction error is constructed, expressed as: Where α, β, and γ represent weighting coefficients, and satisfy: By comprehensively reconstructing the error, information on instantaneous changes, local changes, and long-term changes can be considered simultaneously, thereby improving the stability and robustness of anomaly detection.

[0081] Under normal flight conditions, there is a high degree of consistency between the actual state and the reconstructed state, resulting in a small overall reconstruction error. However, under abnormal flight control conditions, the multimodal correlation is disrupted, leading to a significant increase in the overall reconstruction error.

[0082] In step S500, a normal flight state prototype library is constructed based on the set of flight state representations, and the state representation space constituted by the flight state representations is optimized using a prototype contrastive learning method. Step S500 also includes the following steps.

[0083] S510. Clustering algorithm is used to perform cluster analysis on samples of normal flight status.

[0084] S520. Construct state prototypes based on each cluster center. The state prototypes include takeoff state prototype, hovering state prototype, cruise state prototype, and landing state prototype.

[0085] S530. Based on state prototypes, the flight state representation is constrained by prototype contrast loss, so that normal states are clustered into corresponding state prototypes, thereby forming state prototype features (i.e., state representation structures) with prototype distinguishability in the state representation space.

[0086] Specifically, traditional contrastive learning methods typically construct positive and negative samples using random perturbations, such as random cropping, random noise perturbation, and feature masking. However, for eVTOL flight state data, such random perturbations often fail to accurately reflect the true differences between flight states. Therefore, this invention constructs a prototype flight state library and utilizes a prototype contrastive learning method to optimize the state representation space.

[0087] First, extract flight status representations from normal flight data: in: The normal flight state represents a set. This represents the i-th flight state. This represents the number of normal flight samples. Then, a clustering algorithm is used to perform cluster analysis on the normal flight samples. Let the number of samples obtained after clustering be... A flight state prototype is constructed based on normal flight state samples to create a state prototype library. The state prototype library is represented as follows: in, This represents the total number of prototype working conditions. This is the default configuration in this embodiment. The four prototype types correspond one-to-one with the four basic flight conditions of takeoff, hovering, cruise, and landing; aircraft types performing complex flight path missions such as circling and climbing can be configured... Two new prototype scenarios, climb and hover, have been added. After clustering, the prototype library is stored offline in the onboard Flash storage unit, and can be directly read and called during the onboard inference phase without the need for real-time clustering calculations. In actual flight scenarios, the state prototypes can correspond to: takeoff, hovering, cruise, landing, and other normal flight states.

[0088] Next, the flight state characterization is calculated. With state prototype The cosine similarity between the two conditions is used to quantify the degree of matching between the current flight conditions and standard normal operating conditions. The cosine similarity is expressed as: in, Representation of flight status With state prototype Cosine similarity between them; This indicates that the column vector Transpose to row vector and column vector Perform matrix multiplication; Representing vectors and The L2 norm is the product of the lengths of two vectors.

[0089] Then, a prototype contrastive loss function is further constructed based on the cosine similarity between the flight state representation and the state prototype. It is expressed as: in, This represents the natural exponential function. Indicates the positive sample prototype. This represents the temperature coefficient.

[0090] By minimizing the prototype contrast loss function to optimize the state representation space, samples corresponding to the same flight state are clustered towards the corresponding state prototype, while samples corresponding to different flight states are separated from each other.

[0091] After training, the normal flight state will form multiple stable state cluster structures, while the abnormal state will show obvious deviation in the state space because it cannot match any normal state prototype, thus providing a basis for subsequent anomaly detection.

[0092] In step S600, an anomaly score is calculated based on the deviation between the true state distribution in the true state representation set and the reconstructed state distribution in the reconstructed state representation set. Step S600 also includes the following steps.

[0093] S610. Calculate the Wasserstein distance based on the true state distribution and the reconstructed state distribution.

[0094] S620. The comprehensive reconstruction error based on the sum of multi-scale reconstruction errors. An anomaly scoring function is constructed using the Wasserstein distance.

[0095] S630, Based on anomaly scoring It is compared with a preset threshold δ, when the condition is met The system determines the current flight status as abnormal and outputs the corresponding abnormality detection result.

[0096] S640. Calculate the anomaly contribution using the attention weights of each node in the dynamic graph neural network, and identify the node with the largest contribution as the source node of the anomaly to achieve anomaly localization.

[0097] Specifically, an anomaly scoring function is constructed based on the comprehensive reconstruction error and the state distribution deviation, and anomaly detection and localization are then performed. After the above steps, the comprehensive reconstruction error can be obtained. Flight status characterization ; Reconstructing state representation .

[0098] Furthermore, this embodiment uses the Wasserstein distance to measure the degree of difference between the true state distribution and the reconstructed state distribution. Let the true state distribution be represented as... The reconstructed state distribution is represented as follows: The Wasserstein distance is then expressed as: In the formula, It represents the set of all feasible joint distributions between the true state distribution and the reconstructed state distribution; Represents any set of joint distribution mappings; For mathematical expectation calculations, a fast approximation algorithm is used to solve the distance problem during the airborne online inference phase, with a single-frame computation time of less than 0.5ms, meeting the 50Hz real-time inference timing constraint. A comprehensive anomaly scoring function is constructed by integrating local multi-scale reconstruction errors and global representation distribution biases. Compared with traditional Euclidean or Mahalanobis distances, Wasserstein distance can more accurately characterize the overall difference between two probability distributions, thus being more suitable for anomaly detection tasks under complex flight conditions. The anomaly scoring function is then constructed: Where: λ represents the balance coefficient; Indicates the deviation in state distribution; This indicates the overall reconstruction error. After an anomaly is triggered, the airborne edge computing unit synchronously outputs a structured fault detection report to the lower-level flight controller and ground remote control and maintenance terminal via the CAN bus. The report includes: a high-precision anomaly trigger timestamp, the current flight status of the aircraft (takeoff / hovering / cruise / landing), and quantitative indicators (overall anomaly score, adaptive judgment threshold). (Three-layer hierarchical reconstruction error, Wasserstein distribution distance), fault tracing subsystem type, and corresponding subsystem anomaly contribution. Simultaneously, based on the anomaly score, a three-level fault handling logic is defined: Level 1 Minor Fault: It only stores fault logs locally and has no onboard audible or visual alarms. Level 2 moderate fault: Airborne audible and visual alarms, and real-time push of all fault data to the ground operation and maintenance platform; Level 3 critical fault: The flight control system automatically triggers deceleration and nearby emergency landing protection logic.

[0099] Finally, by integrating all quantitative indicators and positioning information, a complete eVTOL flight control status anomaly detection result is output, including the anomaly fault type, anomaly occurrence time, fault level, anomaly confidence level, and fault tracing subsystem.

[0100] Furthermore, this invention utilizes node attention weights in a dynamic graph neural network to analyze the source of anomalies. Let the weight corresponding to the i-th node be ωi, then the anomaly contribution is expressed as: The node with the largest contribution is selected as the source node of the anomaly. If the source node corresponds to the power system, it is determined to be an anomaly of the power system; if it corresponds to the energy system, it is determined to be an anomaly of the energy system; if it corresponds to the navigation system, it is determined to be an anomaly of the navigation system; if it corresponds to the flight control system, it is determined to be an anomaly of the flight control system.

[0101] The final output includes the anomaly type, anomaly time, anomaly level, and anomaly confidence level, forming the eVTOL flight control status anomaly detection result.

[0102] The following specific embodiments illustrate the execution and application process of the method of this application.

[0103] This embodiment uses a typical quadcopter manned eVTOL as a verification platform to illustrate the parameter configuration of the entire process from data acquisition to anomaly detection.

[0104] 1. Data Acquisition and Preprocessing (S100) Sensor configuration: Assume that the eVTOL is equipped with 8 types of sensors (M=8), all with a sampling frequency of 50Hz.

[0105] Sliding window: The window length is w=50, corresponding to 2 seconds of flight data. This means that the input for each model analysis is 50 sampling points from all sensors within the past 2 seconds.

[0106] Feature fusion: At each time step t, all eight types of sensor data are concatenated into a multi-dimensional vector. For example, if each type of sensor outputs 4-dimensional features after processing, then The dimension is 8*4=32.

[0107] 2. Dynamic Graph Construction and Feature Extraction (S200) Nodes and Edges: Construct 4 nodes: Flight Control ( ),power( ),energy( ),navigation( Fixed directed edge: ( ),( ),( ).

[0108] Node feature dimensions: settings That is, the initial characteristics of each node. It is a 64-dimensional vector.

[0109] Graph Attention Network: A two-layer graph attention network (GAT) is used. The first layer maps 32-dimensional input features to 64-dimensional node features, and the second layer maintains 64 dimensions.

[0110] Output: After GAT, 64-dimensional features of 4 nodes are obtained. Then, a 256-dimensional (4*64) flight state representation z is obtained by using a readout function (such as mean pooling). t .

[0111] 3. Cross-modal reconstruction (S300) Masking strategy: In each training batch, one of the eight modes is randomly selected for masking (replaced with the zero vector). For example, the "motor current" mode is masked in this case.

[0112] Reconstruction Network: A 4-layer Transformer is used as the cross-modal reconstruction network. The input consists of the masked 7 modal features (plus a zero vector as a placeholder), and the output is the reconstructed "motor current" modal feature. .

[0113] Reconstruction error: Calculating the true current characteristics With reconstruction features Mean square error between .

[0114] 4. Multi-scale reconstruction error calculation (S400) Parameter configuration: Point-level error weights Window-level error weights (window (i.e., 0.2 seconds) Trend-level error weights Notice: This indicates that the primary focus is on instantaneous point-level errors, while also considering local windows and trends.

[0115] Overall error: Assuming at a certain moment, =0.1, =0.05, =0.02, then the comprehensive multi-scale reconstruction error =0.5*0.1+0.3*0.05+0.2*0.02=0.069.

[0116] 5. Prototype Comparison Learning (S500) Prototype Construction: The K-means algorithm is used to cluster the state representation z of a large amount of normal flight data, and the following settings are made: These correspond to prototypes p1, p2, p3, and p4 for the four operating conditions: takeoff, hovering, cruise, and landing, respectively.

[0117] Temperature coefficient: Set τ = 0.1.

[0118] Loss Calculation: Assuming the current sample is in a "cruising" state, its representation z_i has a cosine similarity of 0.9 with the cruising prototype p3, and similarities of 0.2, 0.1, and 0.3 with other prototypes, respectively. Then, the prototype contrast loss is calculated as follows: The value will be small, prompting the model to bring samples in the "cruising" state closer to p3.

[0119] 6. Anomaly Detection and Localization (S600) Anomaly scoring: Set the balance coefficient λ = 0.3. Assume that L at the current time... multi =0.069, the Wasserstein distance D is obtained by calculating the set of state representations and the set of reconstructed representations within the most recent time window (e.g., 5 seconds). w =0.5.

[0120] The anomaly score is Score = 0.069 + 0.3 * 0.5 = 0.219.

[0121] Threshold setting: The threshold δ=0.15 was determined using the validation set.

[0122] Judgment: Due to The current flight status is determined to be abnormal.

[0123] Localization: Examine the attention weights of four nodes in a dynamic graph network. Assume a "dynamic system" node is found. Attention weights The highest (for example, If the value is 0.6 and the average value of other nodes is 0.13, then the source of the anomaly is determined to be the power system. Possible causes include motor failure or blade damage.

[0124] Output: Final output: "An abnormal flight control status was detected. Anomaly type: Power system anomaly. Anomaly time: 2026-06-10 14:35:22. Anomaly score: 0.219. Confidence level: 92%".

[0125] Furthermore, the multi-scale reconstruction error is calculated in step S400.

[0126] Suppose that at a certain time t, after the model reconstructs the "motor current" mode, the following intermediate result is obtained (for simplicity, only one sample is shown): Point-level reconstruction error: Here, N=1 (single point), the true feature x=[0.5,0.6,0.4], and the reconstructed features... ,but: =(0.5 0.45) 2+(0.6 0.58) 2 +(0.4 0.38) 2 =0.0025+0.0004+0.0004=0.0033.

[0127] Window-level reconstruction error: Taking a window size W=10, and assuming the point-level errors of the first 9 points are 0.003, 0.004, 0.002, 0.005, 0.003, 0.006, 0.001, 0.007, and 0.002 respectively, and adding the current point's error of 0.0033, then: (0.003+0.004+0.002+0.005+0.003+0.006+0.001+0.007+0.002+0.0033)=0.00363.

[0128] Trend-level reconstruction error: Taking T=10, and assuming the sum of squares of the differences between the actual and reconstructed changes at adjacent time points is 0.05, then: ×0.05=0.00556.

[0129] Overall reconstruction error: Parameter configuration: Point-level error weight α = 0.5, window-level error weight β = 0.3 (window W = 10, i.e., 0.2 seconds), trend-level error weight γ = 0.2. Substituting the values, we get: 0.5 × 0.0033 + 0.3 × 0.00363 + 0.2 × 0.00556 = 0.00165 + 0.001089 + 0.001112 = 0.003851.

[0130] In step S600, anomaly scoring is performed to calculate the Wasserstein distance.

[0131] In practical engineering, the Wasserstein distance is approximated using the Sinkhorn algorithm. Assuming that within the current 5-second window (250 sampling points), the approximate Wasserstein distance between the true state representation set and the reconstructed state representation set is... = 0.25 (This value is calculated using an algorithm; the result is given directly here).

[0132] Anomaly rating: =0.003851 + 0.3 × 0.25 = 0.078851 Threshold setting: The threshold δ = 0.15 was determined using the validation set. Threshold comparison: due to If so, the current state is determined to be normal.

[0133] In summary, the eVTOL flight control state anomaly detection method based on dynamic graph cross-modal reconstruction and prototype contrastive learning provided by this invention fully explores the real-time coupling relationships between multiple subsystems by constructing the flight control, propulsion, energy, and navigation systems as a dynamic graph and extracting spatiotemporal state representations using a graph attention mechanism. It effectively avoids error accumulation in traditional prediction frameworks by using a cross-modal reconstruction network to reconstruct masked modes from multiple unmasked modes. Furthermore, it introduces a normal state prototype library based on flight condition clustering and employs prototype contrastive learning to optimize the state representation space, significantly enhancing the separability between normal and abnormal states. On this basis, it integrates point-level, window-level, and trend-level multi-scale reconstruction errors with Wasserstein distribution distance to construct an anomaly scoring function, and locates faulty subsystems based on graph node attention weights, achieving accurate anomaly identification and intelligent fault tracing for eVTOL flight control states. This method has advantages in detection accuracy, robustness, real-time performance, and generalization ability, providing an efficient and reliable technical solution for eVTOL flight safety monitoring in complex low-altitude environments.

[0134] An electronic device provided in this application includes a memory and a processor; the memory is a computer-readable storage medium used to store computer programs; the processor is used to implement the above-mentioned eVTOL flight control state anomaly detection method based on dynamic graph cross-modal reconstruction and prototype comparison learning when executing the computer program.

[0135] In this embodiment, the beneficial effects of the electronic device and the computer-readable storage medium are similar to those of the above-described method for detecting anomalies in eVTOL flight control based on dynamic graph cross-modal reconstruction and prototype comparison learning, and will not be repeated here.

[0136] The present invention describes electronic devices that can serve as servers or clients of this application, which are examples of hardware devices applicable to various aspects of this application. Electronic devices are intended to represent various forms of digital electronic computer devices. Electronic devices can also represent various forms of mobile devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0137] Electronic devices include a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM can also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0138] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The storage medium can be read-only memory (ROM) or random access memory (RAM), etc. The integrated units described above can be implemented in hardware or as software functional units.

[0139] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting anomalies in eVTOL flight control systems based on dynamic graph cross-modal reconstruction and prototype comparison learning, characterized in that, include: Acquire multimodal sensor data during eVTOL flight, preprocess it, and construct a fused feature sequence; A dynamic graph structure is constructed based on the fused feature sequence, and a dynamic graph neural network is used to extract flight state representations. A modal masking operation is performed on the fused feature sequence, and the masked modality is reconstructed using a cross-modal reconstruction network to generate a reconstructed state representation; The multi-scale reconstruction error is calculated based on the real state representation corresponding to the flight state representation and the reconstructed state representation. A normal flight state prototype library is constructed based on the set of flight state representations, and the state representation space constituted by the flight state representations is optimized by the prototype contrastive learning method. Anomaly scores are calculated based on the deviation between the true state distribution in the true state representation set and the reconstructed state distribution in the reconstructed state representation set. Based on the anomaly score, anomaly detection and anomaly localization processing are performed, and the eVTOL flight control status anomaly detection result is output.

2. The method as described in claim 1, characterized in that, The multimodal sensor data includes one or more of the following: video image data, inertial measurement unit data, global positioning system data, motor operating parameters, battery operating parameters, flight control commands, wind speed sensor data, barometric pressure sensor data, and attitude angle data.

3. The method as described in claim 1, characterized in that, The acquisition of multimodal sensor data during eVTOL flight, followed by preprocessing and construction of a fused feature sequence, includes: Outlier removal and missing value completion are performed on the multimodal sensor data to obtain the data for each modality; The data is classified and processed based on the data type of each modality. Continuous data is normalized using Min-Max, and discrete / categorical data is represented by one-hot encoding or embedding. Time synchronization is performed based on a unified time base. The synchronized modal data are spliced ​​and fused in chronological order to obtain a fused feature sequence.

4. The method as described in claim 1, characterized in that, A dynamic graph structure is constructed based on the fused feature sequence, and a dynamic graph neural network is used to extract flight state representations, including: Modal features corresponding to the flight control system, power system, energy system and navigation system are extracted from the fused feature sequence and used as the initial state features of the graph structure nodes, respectively. Fixed directed edge connections are constructed based on the physical coupling relationships between the systems; Based on the graph attention mechanism, the association weights between nodes are dynamically learned according to the current features of each node, and the node representation is iteratively updated by aggregating neighbor information; A dynamic graph neural network is used to output a flight state representation that integrates spatiotemporal information.

5. The method as described in claim 1, characterized in that, A modality masking operation is performed on the fused feature sequence, and the masked modality is reconstructed using a cross-modal reconstruction network to generate a reconstructed state representation, including: Randomly select at least one modality to perform a masking operation; A cross-modal reconstruction network is used to reconstruct the masked modes and output the reconstructed modal features; A reconstruction state representation is generated based on the reconstruction modal features.

6. The method as described in claim 1, characterized in that, The multi-scale reconstruction error is calculated based on the actual state representation corresponding to the flight state representation and the reconstructed state representation, including: Calculate point-level reconstruction error, window-level reconstruction error, and trend-level reconstruction error; The point-level reconstruction error is expressed as: The window-level reconstruction error is expressed as: The trend-level reconstruction error is expressed as: The comprehensive reconstruction error is constructed based on the point-level reconstruction error, window-level reconstruction error, and trend-level reconstruction error, and is expressed as follows: in, Indicates the number of samples; Represents the characteristics of the true state; Represents the reconstructed state characteristics; Indicates the length of the sliding time window; Indicates the first The point-level reconstruction error corresponding to each time step; Let represent the length of the time series; α, β, and γ represent the weighting coefficients, and satisfy the following: .

7. The method as described in claim 1, characterized in that, A normal flight state prototype library is constructed based on the set of flight state representations, and the state representation space constituted by the flight state representations is optimized using a prototype contrastive learning method, including: Clustering algorithms were used to perform cluster analysis on samples in normal flight conditions; A state prototype is constructed based on each cluster center. The state prototype includes a takeoff state prototype, a hovering state prototype, a cruise state prototype, and a landing state prototype. Based on the state prototype, the flight state representation is constrained by prototype contrast loss, so that the normal state is clustered into the corresponding state prototype, thereby forming state prototype features with prototype distinguishability in the state representation space.

8. The method as described in claim 7, characterized in that, include: A state prototype library is constructed based on normal flight state samples. The state prototype library is represented as follows: Calculate flight state characterization With state prototype The cosine similarity between them is expressed as: A prototype contrast loss function is constructed based on the cosine similarity between the flight state representation and the state prototype, and it is expressed as follows: in, This represents the total number of prototypes for the working conditions. Representation of flight status With state prototype Cosine similarity between them; This indicates that the column vector Transpose to row vector and column vector Perform matrix multiplication; Representing vectors and The L2 norm, which is the product of the lengths of two vectors; This represents the natural exponential function. Indicates the positive sample prototype. This represents the temperature coefficient.

9. The method as described in claim 1, characterized in that, Anomaly scores are calculated based on the deviation between the true state distribution in the true state representation set and the reconstructed state distribution in the reconstructed state representation set, including: The Wasserstein distance is calculated based on the true state distribution and the reconstructed state distribution. The Wasserstein distance is expressed as: The overall reconstruction error is calculated based on the sum of the multi-scale reconstruction errors. An anomaly scoring function is constructed using the Wasserstein distance, and it is expressed as follows: in, Represented as the true state distribution, Represented as the reconstructed state distribution; It represents the set of all feasible joint distributions between the true state distribution and the reconstructed state distribution; Represents any set of joint distribution mappings; This is a mathematical expectation operation; λ represents the balance coefficient. According to the anomaly score It is compared with a preset threshold δ, when the condition is met The system determines the current flight status as abnormal and outputs the corresponding abnormality detection result. The anomaly contribution is calculated using the attention weights of each node in the dynamic graph neural network. The node with the largest contribution is identified as the source node of the anomaly, thus achieving anomaly localization.

10. An electronic device, characterized in that, The electronic device includes: The memory is a computer-readable storage medium that stores program instructions. A processor, the processor being configured to run the program instructions to execute the eVTOL flight control state anomaly detection method based on dynamic graph cross-modal reconstruction and prototype comparative learning as described in any one of claims 1 to 9.

Images