Flight control sensor electric signal anomaly detection method and system based on electric signal analysis

Abstract

The invention provides a flight control sensor electric signal abnormity detection method and system based on electric signal analysis, and the method comprises the steps: collecting the voltage and current signal data of an aircraft control sensor in real time, and carrying out the data preprocessing of the voltage and current signal data; inputting the preprocessed electric signal data into a pre-constructed nonlinear four-rotor kinetic model for analogue simulation, and outputting an electric signal data set containing fault information; based on the real-time feature vector in the electric signal data set and a pre-stored historical feature sample set, constructing a global adjacency graph containing real-time nodes and historical nodes; inputting the global adjacency graph into a pre-constructed graph convolutional network model, and extracting space-time fusion features of the real-time nodes; and carrying out classification processing on the space-time fusion features, and outputting an abnormal state category of a sensor electric signal, so as to achieve the purposes of online high-precision diagnosis and autonomous fault-tolerant control of aircraft sensor faults and guarantee safe and reliable operation of a complex equipment system.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of sensor fault diagnosis, in particular to a flight control sensor electric signal anomaly detection method and system based on electric signal analysis. BACKGROUND

[0002] The reliability of the flight control system of an unmanned aerial vehicle is highly dependent on the normal operation of multiple source sensors such as IMU, GPS and barometer. These sensors are prone to performance degradation in complex working environments due to mechanical vibration, electromagnetic interference or device aging, which manifests as faults such as offset, drift or transient failure at the level of electric signals. Traditional fault diagnosis methods have obvious limitations: methods based on physical rules or threshold judgment are difficult to adapt to dynamic flight conditions and complex fault modes; mainstream machine learning methods (such as support vector machines and long short-term memory networks) that rely on post-data processing often fail to adequately mine early weak anomalies and spatiotemporal correlation features among multiple sensors from raw sensor electric signals; and the use of hardware redundancy significantly increases system cost, weight and power consumption. Therefore, the prior art lacks an effective means for high-sensitivity, high-real-time online monitoring and diagnosis directly from the electric interface level of sensors. SUMMARY

[0003] The main purpose of the application is to provide a flight control sensor electric signal anomaly detection method and system based on electric signal analysis, which bypasses the dependence on physical quantities (such as attitude and position) after sensor analysis in traditional methods, directly collects, models and analyzes the raw voltage and current signals output by the sensor, and deeply mines abnormal patterns and spatiotemporal correlation features in the electric signals by combining dynamic simulation and graph convolution networks, to achieve the purpose of early and accurate electrical fault diagnosis.

[0004] To achieve the above purpose, the application provides a flight control sensor electric signal anomaly detection method based on electric signal analysis, comprising the following steps: Real-time collection of voltage and current signal data of an aircraft control sensor, data preprocessing of the voltage and current signal data; Inputting the preprocessed electric signal data into a pre-constructed nonlinear quadrotor dynamics model for simulation, and outputting an electric signal data set containing fault information; Based on the real-time feature vectors in the electric signal data set and the pre-stored historical feature sample set, a global adjacency graph containing real-time nodes and historical nodes is constructed; Inputting the global adjacency graph into a pre-constructed graph convolution network model to extract spatiotemporal fusion features of the real-time nodes; Classifying the spatiotemporal fusion features to output the abnormal state category of the sensor electric signal.

[0005] Further, the step of data preprocessing the voltage and current signal data comprises: normalizing the three-axis acceleration voltage signals of the inertial measurement unit sensor; filtering and denoising the current pulse signals of the global positioning system module; adopting the box plot method to eliminate outliers in the voltage signal.

[0006] Further, the step of inputting the preprocessed electrical signal data into a pre-constructed nonlinear quadrotor dynamics model for simulation and outputting an electrical signal dataset containing fault information comprises: inputting the preprocessed electrical signal data into a pre-constructed nonlinear quadrotor dynamics model; in the software-in-the-loop and / or hardware-in-the-loop simulation mode, executing a pre-set flight control sequence, and simulating sensor faults through a fault injection module during the simulation process; generating the electrical signal dataset based on the data collected during the software-in-the-loop and / or hardware-in-the-loop simulation process, the dataset including sensor raw electrical signals, flight state information, and corresponding fault labels.

[0007] Further, the construction of the nonlinear quadrotor dynamics model comprises: establishing a nonlinear quadrotor dynamics model based on the Newton-Euler equation in the simulation platform; integrating a fault injection module into the nonlinear quadrotor dynamics model for dynamically modifying model parameters to simulate sensor offset, drift, and failure faults.

[0008] Further, based on the real-time feature vectors in the electrical signal dataset and the pre-stored historical feature sample set, the step of constructing a global adjacency graph containing real-time nodes and historical nodes comprises: extracting real-time feature vectors from the electrical signal dataset; calculating the Euclidean distance between samples based on the real-time feature vectors and the historical feature sample set; based on the K-nearest neighbor algorithm, taking the real-time feature vectors and historical feature samples as nodes and distance relationships as edges to construct a global adjacency graph.

[0009] Further, the step of inputting the global adjacency graph into a pre-constructed graph convolution network model to extract the spatio-temporal fusion features of the real-time nodes comprises: inputting the adjacency matrix and node feature matrix of the global adjacency graph into a pre-trained graph convolution network model; The at least one graph convolution layer of the graph convolution network model aggregates neighbor node feature information of a real-time node, fuses time sequence change characteristics of an electric signal and spatial topological relations between sensors, and generates a space-time fusion feature vector of the real-time node.

[0010] Further, the construction step of the graph convolution network model comprises: constructing an initial global adjacency graph based on a historical sensor electric signal sample set; dividing the historical sensor sample set into a training set and a test set; The at least one graph convolution layer of the graph convolution network model aggregates neighbor node feature information of a real-time node, fuses time sequence change characteristics of an electric signal and spatial topological relations between sensors, and generates a space-time fusion feature vector of the real-time node updating node representation by aggregating global neighbor features through a plurality of space-time fusion feature extraction layers; calculating cross-entropy loss of nodes in the training set using a Softmax classification layer; updating model parameters using a stochastic gradient descent algorithm; verifying model accuracy based on nodes in the test set and deploying.

[0011] Further, the step of classifying the space-time fusion feature and outputting an abnormal state category of the sensor electric signal comprises: classifying the space-time fusion feature using a Softmax classification layer in the graph convolution network model and outputting a current state category of the flight control sensor, wherein the state category comprises normal, offset, drift or failure.

[0012] Further, after the step of classifying the space-time fusion feature and outputting an abnormal state category of the sensor electric signal, the step further comprises: extracting original voltage and current signals within a corresponding time window according to the abnormal state category; packaging a fault message, which at least contains a fault type, a confidence, original signal waveforms and a timestamp; sending the fault message to a flight system for triggering a corresponding fault-tolerant control strategy.

[0013] The application also provides a flight control sensor electric signal anomaly detection system based on electric signal analysis, comprising: a real-time acquisition unit configured to acquire voltage and current signal data of an aircraft control sensor in real time and perform data preprocessing on the voltage and current signal data; a modeling simulation unit configured to input the preprocessed electric signal data into a pre-constructed nonlinear quadrotor dynamics model for simulation and output an electric signal data set containing fault information; a structure graph construction unit configured to construct a global adjacency graph containing real-time nodes and historical nodes based on real-time feature vectors in the electrical signal dataset and a pre-stored historical feature sample set; a feature extraction unit configured to input the global adjacency graph into a pre-constructed graph convolution network model to extract spatio-temporal fusion features of the real-time nodes; a state determination unit configured to perform classification processing on the spatio-temporal fusion features to output an abnormal state category of the sensor electrical signal.

[0014] The flight control sensor electrical signal anomaly detection method and system based on electrical signal analysis provided by the application have the following beneficial effects: the application directly monitors from the sensor electrical signal level, captures earlier and more subtle abnormal phenomena before physical quantity analysis, and has high diagnostic sensitivity; the graph convolution network is used to simultaneously model the time sequence evolution law of the electrical signal and the spatial coupling relationship between multiple sensors, and has stronger representation ability for complex fault modes; based on the lightweight graph model and the preprocessing process, millisecond-level online diagnosis can be realized; at the same time, a large amount of fault data is generated by using the dynamic model to train the model, and the generalization ability for unseen fault conditions is improved; a complete closed loop from signal collection, fault diagnosis to message reporting is provided, which can be directly embedded in the existing flight control system, provides key input for autonomous fault-tolerant control, and improves the safety and reliability of the aircraft. BRIEF DESCRIPTION OF DRAWINGS

[0015] The application can be further understood from the following description in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but emphasis is placed on showing the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0016] Figure 1 is a structural schematic diagram of the flight control sensor electrical signal anomaly detection method based on electrical signal analysis according to an embodiment of the application; Figure 2 is a structural schematic diagram of the flight control sensor electrical signal anomaly detection method based on electrical signal analysis according to an embodiment of the application; Figure 3 is a structural schematic diagram of the flight control sensor electrical signal anomaly detection method based on electrical signal analysis according to an embodiment of the application; Figure 4 is a structural schematic diagram of the hardware-in-the-loop (HIL) simulation process of the nonlinear quadrotor dynamics model in an embodiment of the application; Figure 5 is a structural schematic diagram of three graph construction methods on the rotor electrical signal dataset in an embodiment of the application; Figure 6 is a structural schematic diagram of a two-layer GCN architecture model in an embodiment of the application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with its embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the protection scope of the present application.

[0018] Referring to Figure 1 A flowchart of a flight control sensor electric signal anomaly detection method based on electric signal analysis according to the present application is shown in the figure, which comprises the following steps: S1, real-time acquisition of voltage and current signal data of an aircraft control sensor, data preprocessing of the voltage and current signal data; S2, input of the preprocessed electric signal data into a pre-constructed nonlinear quadrotor dynamics model for simulation and output of an electric signal data set containing fault information; S3, based on the real-time feature vector in the electric signal data set and the pre-stored historical feature sample set, construction of a global adjacency graph containing real-time nodes and historical nodes; S4, input of the global adjacency graph into a pre-constructed graph convolution network model to extract the spatio-temporal fusion features of the real-time nodes; S5, classification processing of the spatio-temporal fusion features to output the abnormal state category of the sensor electric signal.

[0019] In one embodiment, for step S1, The step of data preprocessing of the voltage and current signal data comprises: Normalizing the three-axis acceleration voltage signal of the inertial measurement unit sensor; Filtering and denoising the current pulse signal of the global positioning system module; Using the box plot method to eliminate outliers in the voltage signal.

[0020] In a specific implementation, the original voltage signal of the three-axis accelerometer is collected (range ±5V), Z-score standardization is used to eliminate the dimension and stabilize the data distribution, and the mean value and standard deviation of each axis are calculated based on the historical calibration data : Perform standardization: For the GPS current pulse signal (baseline noise ± 2mA), a second-order Butterworth low-pass filter is designed to suppress high-frequency interference (such as electromagnetic pulses), and to retain the current step characteristics caused by satellite signal loss and other faults. Boxplot method is used to process voltage signal, and the interquartile range , the rejection condition is: or According to the flight phase, the threshold value is adaptively adjusted (the IQR coefficient is expanded to 2.0 during flight attitude maneuver), and the voltage outliers caused by transient interference are removed.

[0021] In one embodiment, for step S2, The preprocessed electrical signal data is input into the pre-constructed nonlinear quadrotor dynamics model for simulation, and the step of outputting the electrical signal data set containing fault information includes: Real-time acquisition of electrical signal data output by the flight control sensor of the unmanned aerial vehicle, data preprocessing of the voltage and current signal data; The preprocessed electrical signal data is input into the pre-constructed nonlinear quadrotor dynamics model; In the software-in-the-loop and / or hardware-in-the-loop simulation mode, a pre-set flight control sequence is executed, and sensor faults are simulated by a fault injection module during simulation; In the software-in-the-loop and / or hardware-in-the-loop simulation mode, a pre-set flight control sequence is executed, and sensor faults are simulated by a fault injection module during simulation;

[0022] In specific implementation, real sensor fault samples are high-cost and dangerous, and a dynamics model and fault injection can generate a data set covering multiple fault modes; SIL / HIL dual-mode verification can ensure the consistency of the model in virtual and physical environments, and improve the diagnostic generalization ability. In the simulation platform, a nonlinear quadrotor dynamics model is established based on Newton-Euler equation, a fault injection module is designed to dynamically modify model parameters (such as accelerometer bias, gyroscope drift), and sensor bias (Bias), drift (Drift), failure (Failure) and other fault modes are simulated to accurately describe the dynamic relationship between the attitude, position and motor speed of the unmanned aerial vehicle. Dual-mode simulation is performed, software-in-the-loop (SIL) is pure digital simulation, and fault parameters (such as Figure 3 ) are injected through the control sequence; hardware-in-the-loop (HIL) is connected to Pixhawk flight control hardware, and a fault injection module (such as Figure 4Parameters can be dynamically modified via API, such as adding a fixed offset to the IMU angular velocity signal (analog offset), an offset that increases linearly with time (analog drift), or setting the signal to zero (analog failure). This is demonstrated in both Software-in-the-Loop (SIL) and Hardware-in-the-Loop (HIL) modes connected to the actual flight controller hardware (Pixhawk). Figure 3 , Figure 4 The system performs hovering, maneuvering, and other flight tasks and triggers faults, simultaneously recording the raw voltage / current signals, flight status, and fault labels of all sensors to form a multi-dimensional electrical signal fault dataset: flight information (attitude angles) angular velocity The technical steps include: ground data (manually labeled fault types), telemetry logs (timing sequences of raw sensor voltage / current signals), and flight logs (timing sequences of environmental parameters (temperature, air pressure) and control commands). This approach overcomes the bottleneck of traditional methods that rely on limited real fault data, providing high-fidelity input for global adjacency graph construction.

[0023] In one embodiment, the construction of a nonlinear quadrotor dynamics model includes: A nonlinear quadrotor dynamics model was established in the simulation platform based on the Newton-Euler equations. A fault injection module is integrated into the nonlinear quadrotor dynamics model to dynamically modify model parameters to simulate sensor offset, drift, and failure.

[0024] Specifically, in the RflySim platform, a nonlinear quadrotor dynamics model is constructed based on the Newton-Euler equations: In the formula, Let g be the mass of the drone, and g be the acceleration due to gravity. For position coordinates, This refers to the air drag coefficient; For total lift, Indicates rotor lift ( The lift coefficient, (This refers to the rotor speed).

[0025] Rotational motion follows Euler's equations of motion ( In the formula, (This is the moment of inertia matrix), where the torque is generated by the difference in rotor speed: In the formula, The distance from the rotor center to the center of mass. This is the inverse torque coefficient.

[0026] The attitude kinematic equations are constructed using Euler angles: In the formula, The angular velocity of the body axis, .

[0027] The constructed nonlinear quadrotor dynamics model is uniformly expressed in matrix form, and the transformation matrix ensures the completeness of the attitude solution: Based on this physical model, a fault injection module is designed to dynamically modify parameters: rotor failure is detected. Implementation; sensor deviation modeling as sum; motor torque reduction using The simulation model parameters were calibrated using measured data from the Pixhawk flight controller (moment of inertia). With key parameter errors <3%, fault boundaries conform to the DO-178C safety standard, and SIL / HIL dual-mode fault injection supported by the RflySim hard real-time kernel (latency ≤10ms), the physical fidelity and operational condition coverage of the dataset generated in step S1 are ensured.

[0028] In one embodiment, for step S3, The step of constructing a global adjacency graph containing real-time nodes and historical nodes based on the real-time feature vectors in the electrical signal dataset and the pre-stored historical feature sample set includes: Extract the real-time feature vector of the electrical signal dataset; Based on the real-time feature vector and the historical feature sample set, calculate the Euclidean distance between samples; Based on the K-nearest neighbor algorithm, a global adjacency graph is constructed using the real-time feature vector and historical feature samples as nodes and distance relationships as edges.

[0029] In practice, statistical features within a fixed time window are extracted from real-time data streams and historical datasets. (d=100 dimensions, including time-series features such as IMU acceleration / angular velocity, GPS location, and barometric pressure), forming a feature vector. The Euclidean distance between the real-time feature vector and all historical feature vectors is calculated. The K-nearest neighbor algorithm is used to connect each sample (node) to its K most similar samples (K is typically 3-10). Adding real-time nodes to this graph creates a global adjacency graph containing both historical and real-time information. Specifically, pre-stored... Historical sample feature matrix (Save as a .graph file to speed up reading). The similarity between real-time and historical nodes is measured using Euclidean distance: In the formula, The historical sample matrix H is the first Row vectors.

[0030] Construct a global adjacency graph using KNN. Merge the real-time feature r and the historical feature H into an augmented matrix: Recalculate the distances across all samples and run the K-nearest neighbor algorithm (fixed during the diagnostic phase). ): Generate a new adjacency matrix , where nodes For real-time nodes.

[0031] Verification via grid search Experiments show The optimal balance is achieved on the test set (too small a value leads to sensitivity to noise, while too large a value introduces irrelevant nodes). In this embodiment, the historical graph structure is pre-computed during the training phase to avoid repetitive graph construction during diagnosis; by adding real-time nodes that only require local distance calculation, the update time of the adjacency matrix increment is reduced. Compared to Radius Graph (fixed threshold is easily affected by dimensions) and Path Graph (dependent on sequence relationships), KNN is more adaptable to the multimodal distribution characteristics of sensors. A comparison of graph construction methods (such as...) on a quadcopter electrical signal dataset is shown. Figure 5 As shown in Table 1: Note: The test environment was an Intel i7-11800H CPU, and the fault types included offset, drift, and failure.

[0032] In one embodiment, for step S4, The step of inputting the global adjacency graph into a pre-built graph convolutional network model and extracting the spatiotemporal fusion features of the real-time nodes includes: The adjacency matrix and node feature matrix of the global adjacency graph are input into a pre-trained graph convolutional network model; By aggregating the feature information of neighboring nodes of the real-time node through at least one graph convolutional layer of the graph convolutional network model, and fusing the temporal variation features of the electrical signal with the spatial topological relationship between sensors, a spatiotemporal fusion feature vector of the real-time node is generated.

[0033] In specific implementation, the global adjacency graph obtained in step S2 is accepted. This global adjacency graph includes two core components: an adjacency matrix and a node feature set. The adjacency matrix... (N is the number of historical samples, matrix elements) Represents a node With nodes Connected nodes, generated by the K-nearest neighbor algorithm, with the first N rows / columns representing historical nodes and the (N+1)th row / column representing real-time nodes; node feature set. ,in Features of historical nodes (pre-stored in .graph files). This represents real-time node features (current flight sensor data). The above components are organized as the standard input to GCN, where the feature matrix is: Normalized adjacency matrix (Pre-calculation during training phase): in, It is a degree matrix with self-loops.

[0034] After input is received, spatiotemporal feature fusion is performed using a pre-trained graph convolutional network model. First convolutional layer (spatial topology aggregation): In the formula Aggregate the neighbor features of the real-time node r (such as associated IMU-GPS nodes); second convolutional layer (spatiotemporal feature fusion): In the formula By integrating dynamic temporal sequences (historical drift trends) with spatial relationships, historical nodes implicitly reveal fault evolution patterns (such as the increasing gyroscope drift over time). The multi-hop connection allows real-time nodes to absorb the time-series patterns of historical nodes. Real-time node output characteristics: in, The KNN direct neighbors (spatial topology) of the real-time node. This refers to temporally related nodes of historical neighbors (e.g., consecutive time slices of the same sensor). Extraction The last line: Features and connotations, Table 2: In practice, historical node features directly call pre-stored values ​​to avoid real-time calculations; the actual latency measured on the Pixhawk flight control platform is <150ms.

[0035] In one embodiment, the steps for constructing a graph convolutional network model include: An initial global adjacency graph is constructed based on a historical set of sensor electrical signal samples; The historical sensor sample set was divided into a training set and a test set; By aggregating the feature information of neighboring nodes of the real-time node through at least one graph convolutional layer of the graph convolutional network model, and fusing the temporal variation features of the electrical signal with the spatial topological relationship between sensors, a spatiotemporal fusion feature vector of the real-time node is generated. Global neighbor features are aggregated through a multi-layer spatiotemporal fusion feature extraction layer to update node representations; The cross-entropy loss of the training set nodes is calculated using the Softmax classification layer; The model parameters are updated using the stochastic gradient descent algorithm. Verify the model's accuracy and deploy it based on the test set nodes.

[0036] Specifically, data construction is performed, loading data from the electrical signal dataset. Historical samples, each sample has 100-dimensional features extracted (time-series statistics such as IMU angular velocity, GPS displacement, and air pressure values), based on historical sensor sample sets ( Construct a global adjacency graph (using samples): Define the feature vector of each node corresponding to a historical sample. (Includes features such as IMU angular velocity mean, GPS displacement variance, and barometric time sequence entropy); edge generation uses the K-nearest neighbor algorithm ( Establish connection relationship: in, For nodes Calculate the distance to the 5th nearest neighbor and generate an adjacency matrix. .

[0037] The historical sample set was divided into a training set of 80,000 nodes (80% randomly sampled) and a test set of 20,000 nodes (the remaining 20%), while preserving the global adjacency structure. Unchanged (including connections between training and testing nodes). Construct a two-layer graph convolutional network model: Input the feature matrix of the training set. and standardized adjacency matrix Initialization parameters: like Figure 6 Global features are aggregated through multi-layer graph convolution, and direct neighbor features (such as associated IMU-GPS node pairs) are aggregated through the first layer (spatial topology aggregation): The second layer (spatiotemporal fusion) fuses the temporal dynamics of multi-hop neighbors (such as drift trend propagation): Fault classification is performed using the Softmax layer: Output a four-dimensional probability vector (Normal / Offset / Drift / Failure). Perform loss optimization to minimize cross-entropy loss: in, One-hot encoding of the real label (normal / offset / drift / invalid).

[0038] The parameters are updated using stochastic gradient descent (SGD), with the following update rule: Among them, the hyperparameter is configured as the learning rate. 0.01, momentum L2 regularization Batch size: 256 nodes / batch. An early stopping mechanism is implemented; training terminates if the loss does not decrease for five consecutive rounds.

[0039] Perform model validation and deployment. Evaluate on the test set: freeze model parameters and input test set features. Calculation accuracy indicators: Deploy the model and solidify the parameters. Lightweight conversion was performed, using the TensorRT engine to compress it to 3.2MB; flight control integration was performed, embedding PX4 firmware and registering it as a real-time diagnostic module.

[0040] Real-time constraint verification is shown in Table 3: Note: The global adjacency graph is built once during the training phase, and after deployment, real-time node connections are only updated incrementally through KNN.

[0041] In one embodiment, for step S5, The steps for classifying spatiotemporal fusion features and outputting the abnormal state categories of sensor electrical signals include: The spatiotemporal fusion features are classified using the Softmax classification layer in the graph convolutional network model, and the current state category of the flight control sensor is output. The state category includes normal, offset, drift, or failure. Extract the raw electrical signals from the flight control sensors corresponding to the state category; The original electrical signals and abnormal status categories of the flight control sensors are sent to the flight control system.

[0042] In practical implementation, fault decision-making and system response are based on the spatiotemporal fusion features generated in step S3. Input spatiotemporal fusion features Calculate the probability of four types of failures: in, Decision-making rules: Note: The threshold was determined through ROC curve optimization to balance the false alarm rate and the false negative rate. In one embodiment, after classifying the spatiotemporal fusion features and outputting the abnormal state category of the sensor electrical signal, the method further includes: Based on the abnormal state category, extract the original voltage and current signals within the corresponding time window; Encapsulate the fault message, which should at least include the fault type, confidence level, original signal waveform, and timestamp; The fault message is sent to the flight system to trigger the corresponding fault-tolerant control strategy.

[0043] Specifically, after outputting the fault category in step S4, the electrical signal tracing and transmission protocol is executed to lock the fault time window. The backtracking window is defined. (t is the diagnostic time), based on the original signal indexed by the PX4 flight controller ring buffer. Fault message encapsulation is shown in Table 4: Based on the Mavlink protocol, fault messages are sent to the flight control main controller via CAN bus or Mavlink protocol. The flight control system executes preset fault-tolerant strategies according to the message content, such as adjusting control parameters, switching to backup sensors, or entering a safe mode, as shown in Table 5: Reference Figure 2 Here is a structural block diagram of a flight control sensor electrical signal anomaly detection system based on electrical signal analysis according to an embodiment of the present invention, comprising: The real-time acquisition unit is used to acquire voltage and current signal data from the aircraft control sensors in real time and to perform data preprocessing on the voltage and current signal data. The modeling and simulation unit is used to input the pre-processed electrical signal data into a pre-built nonlinear quadrotor dynamics model for simulation and output an electrical signal dataset containing fault information. The structure graph construction unit is used to construct a global adjacency graph containing real-time nodes and historical nodes based on the real-time feature vectors in the electrical signal dataset and the pre-stored historical feature sample set. The feature extraction unit is used to input the global adjacency graph into a pre-built graph convolutional network model to extract the spatiotemporal fusion features of the real-time nodes; The state determination unit is used to classify the spatiotemporal fusion features and output the abnormal state category of the sensor electrical signal.

[0044] For the specific implementation of each unit in the above device example, please refer to the method embodiments described above, and will not be repeated here.

[0045] In summary, this invention acquires voltage and current signal data from aircraft control sensors in real time, preprocesses the voltage and current signal data, inputs the preprocessed electrical signal data into a pre-constructed nonlinear quadrotor dynamics model for simulation, and outputs an electrical signal dataset containing fault information. Based on the real-time feature vectors in the electrical signal dataset and a pre-stored historical feature sample set, a global adjacency graph containing real-time nodes and historical nodes is constructed. The global adjacency graph is input into a pre-constructed graph convolutional network model to extract the spatiotemporal fusion features of the real-time nodes. The spatiotemporal fusion features are classified to output the abnormal state category of the sensor electrical signals, thereby achieving online high-precision diagnosis and autonomous fault-tolerant control of aircraft sensor faults and ensuring the safe and reliable operation of complex equipment systems.

[0046] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementations.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0048] In this invention, the terms "first" and "second" do not represent a specific quantity or order, but are merely used to distinguish names.

[0049] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0050] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for detecting abnormality of electric signal of flight control sensor based on analysis of electric signal, characterized in that, The method comprises the following steps: Real-time acquisition of voltage and current signal data of aircraft control sensors, data preprocessing of the voltage and current signal data; Input the preprocessed electrical signal data into the pre-constructed nonlinear quadrotor dynamics model for simulation, and output the electrical signal data set containing fault information; Based on the real-time feature vector in the electrical signal data set and the pre-stored historical feature sample set, a global adjacency graph containing real-time nodes and historical nodes is constructed; Input the global adjacency graph into the pre-constructed graph convolution network model to extract the spatio-temporal fusion features of the real-time nodes; Classify the spatio-temporal fusion features to output the abnormal state category of the sensor electrical signal.

2. The method of claim 1, wherein the method is based on electrical signal analysis of the sensor electrical signals. The step of data preprocessing of the voltage and current signal data comprises: Normalizing the three-axis acceleration voltage signal of the inertial measurement unit sensor; Filtering and denoising the current pulse signal of the global positioning system module; Using the box plot method to eliminate outliers in the voltage signal.

3. The method of claim 1, wherein the method is based on electrical signal analysis of the sensor electrical signals. The step of inputting the preprocessed electrical signal data into the pre-constructed nonlinear quadrotor dynamics model for simulation to output the electrical signal data set containing fault information comprises: Input the preprocessed electrical signal data into the pre-constructed nonlinear quadrotor dynamics model; In the software-in-the-loop and / or hardware-in-the-loop simulation mode, execute the preset flight control sequence, and simulate the sensor fault in the simulation process through the fault injection module; Generate the electrical signal data set based on the data collected in the software-in-the-loop and / or hardware-in-the-loop simulation process, wherein the data set includes the original electrical signal of the sensor, the flight state information and the corresponding fault label.

4. The method of claim 1, wherein the method is based on electrical signal analysis of the sensor electrical signals. The construction of the nonlinear quadrotor dynamics model comprises: Establishing a nonlinear quadrotor dynamics model based on the Newton-Euler equation in the simulation platform; Integrating a fault injection module in the nonlinear quadrotor dynamics model for dynamically modifying model parameters to simulate the offset, drift and failure faults of the sensor.

5. The method of claim 1, wherein the method is based on electrical signal analysis of the sensor electrical signals. The step of constructing a global adjacency graph containing real-time nodes and historical nodes based on the real-time feature vector in the electrical signal data set and the pre-stored historical feature sample set comprises: Extracting the real-time feature vector of the electrical signal data set; Calculating the Euclidean distance between samples based on the real-time feature vector and the historical feature sample set; Based on the K-nearest neighbor algorithm, taking the real-time feature vector and the historical feature sample as nodes and the distance relationship as edges, a global adjacency graph is constructed.

6. The method of claim 1, wherein the method is based on electrical signal analysis of the sensor electrical signals. The step of inputting the global adjacency graph into the pre-constructed graph convolution network model to extract the spatio-temporal fusion features of the real-time nodes comprises: Input the adjacency matrix and node feature matrix of the global adjacency graph into the pre-trained graph convolution network model; Through at least one graph convolution layer of the graph convolution network model, aggregate the neighbor node feature information of the real-time node, fuse the time sequence change features of the electrical signal and the spatial topological relationship between sensors, and generate the spatio-temporal fusion feature vector of the real-time node.

7. The method of claim 1, wherein the method is based on electrical signal analysis of the sensor electrical signals. The construction step of the graph convolution network model comprises: Construct an initial global adjacency graph based on the historical sensor electrical signal sample set; Divide the historical sensor sample set into a training set and a test set; Through at least one graph convolution layer of the graph convolution network model, neighbor node feature information of a real-time node is aggregated, time sequence change characteristics of electrical signals and spatial topological relationships between sensors are fused, and a spatio-temporal fusion feature vector of the real-time node is generated.

8. The method of claim 1, wherein the method is based on electrical signal analysis of the sensor electrical signals. The step of classifying the spatio-temporal fusion feature and outputting an abnormal state category of the sensor electrical signal comprises: The spatio-temporal fusion feature is classified by using a Softmax classification layer in the graph convolution network model, and a current state category of the flight control sensor is outputted, wherein the state category comprises normal, offset, drift or failure.

9. The method of claim 1, wherein the method is based on electrical signal analysis of the sensor electrical signals. After the step of classifying the spatio-temporal fusion feature and outputting the abnormal state category of the sensor electrical signal, the method further comprises: According to the abnormal state category, original voltage and current signals in a corresponding time window are extracted; A fault message is packaged, and at least contains a fault type, a confidence, original signal waveforms and a timestamp; The fault message is sent to a flight system, for triggering a corresponding fault-tolerant control strategy.

10. A system for detecting abnormality in electric signals of flight control sensors based on analysis of electric signals, characterized in that, The method comprises: A real-time acquisition unit is configured to acquire voltage and current signal data of an aircraft control sensor in real time, and to perform data preprocessing on the voltage and current signal data; A modeling simulation unit is configured to input preprocessed electrical signal data into a pre-constructed nonlinear quadrotor dynamics model for simulation and analog, and to output an electrical signal data set containing fault information; A structure diagram construction unit is configured to construct a global adjacency graph containing real-time nodes and historical nodes based on real-time feature vectors in the electrical signal data set and a pre-stored historical feature sample set; A feature extraction unit is configured to input the global adjacency graph into a pre-constructed graph convolution network model, and to extract spatio-temporal fusion features of the real-time nodes; A state determination unit is configured to classify the spatio-temporal fusion features and output an abnormal state category of the sensor electrical signal.

Images