Multi-source sensing data fused low-altitude flight track offset intelligent repairing method

Abstract

The invention relates to the technical field of intelligent flight control, in particular to a multi-source sensing data fused low-altitude flight track offset intelligent repair method, which comprises the following steps of: performing jitter correction on a visual image stream through a motion compensation module, performing resampling on an ADS-B positioning signal by adopting a Lagrange interpolation method, and performing multi-source sensing data fusion on the ADS-B positioning signal; the radar point cloud is mapped to an image coordinate system, and a fusion track representation vector with time-space synchronization is generated; extracting a curvature abrupt change point set, an environment shielding factor and an airspace visibility gradient based on the vector, and constructing a multi-dimensional offset feature tensor in combination with aircraft dynamics constraints; inputting the tensor into a pre-trained space-time diagram neural network, and outputting a risk grading restoration decision comprising a restoration angle, a climbing rate and a speed compensation amount; injecting the repair decision into a flight control simulator for verification, and outputting a final repair instruction when the trajectory deviation value is smaller than a safety threshold value; the method has high precision and closed-loop execution flight path repairing capability, and is suitable for a flight control system in a complex low-altitude environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent flight control, and in particular to a low-altitude flight track deviation intelligent repair method fusing multi-source sensing data. BACKGROUND

[0002] With the gradual opening of urban low-altitude airspace and the rapid development of unmanned aerial vehicle group operation, low-altitude flying vehicles are widely used in logistics distribution, city inspection, environmental monitoring and other scenarios. However, due to the complex factors such as building obstruction, electromagnetic interference and meteorological mutation in the low-altitude flight environment, the actual track of the flying vehicle often deviates to varying degrees, which easily leads to path drift, heading error and even safety accidents. Therefore, the real-time repair of the flight path, the correction of the track deviation and the risk avoidance have become the key research direction in the current low-altitude track management. With the application and maturity of multi-sensor fusion technology and graph neural network algorithm, the intelligent repair of track deviation by using multi-source data fusion analysis has gradually become a core means to improve the flight stability and safety.

[0003] However, there are still several outstanding problems in the actual application of the prior art: firstly, some methods only rely on a single data source (such as GPS or image sequence), which has poor robustness in obstruction scenarios or signal failure conditions, and it is difficult to obtain complete track state; secondly, multi-source data fusion often lacks a unified spatio-temporal alignment mechanism, resulting in information redundancy or conflict between data, affecting the accuracy of repair judgment; thirdly, the existing repair strategies mostly rely on static path planning or rule matching, which is difficult to respond to time-varying risk factors such as dynamic obstacles and meteorological mutations, and lacks adaptability to complex environments; fourthly, the repair strategy lacks a closed-loop verification mechanism before execution, and cannot dynamically evaluate its executability and stability in the flight control system. SUMMARY

[0004] The present application provides a low-altitude flight track deviation intelligent repair method fusing multi-source sensing data, which has high robustness, can realize dynamic reasoning and closed-loop verification, and effectively solves the bottleneck problem of current low-altitude track management technology.

[0005] A low-altitude flight track deviation intelligent repair method fusing multi-source sensing data, comprising the following steps: S1: receiving a radar point cloud sequence, an ADS-B positioning signal and a visual image stream, performing jitter correction on the visual image stream through a motion compensation module, time resampling the ADS-B signal by using Lagrange interpolation method, mapping the radar point cloud sequence to a synchronous image coordinate system based on timestamp, and outputting a spatio-temporal synchronous fusion track representation vector; S2: based on the fusion trajectory feature vector, extract the curvature mutation point set, environmental shielding factor, and airfield visibility gradient, and combine the aircraft dynamics constraint to construct a multi-dimensional offset feature tensor; S3: input the multi-dimensional offset feature tensor into a pre-trained spatio-temporal graph neural network, associate the dynamic obstacle motion vector with the airfield topology grid, and output a risk classification repair decision including the repair angle, climb rate, and speed compensation amount; S4: inject the risk classification repair decision into a flight control simulator for back propagation verification, generate a final repair instruction when the trajectory deviation value is less than a safety threshold, otherwise return to S3 to adjust the decision parameters.

[0006] Optionally, the S1 comprises: S11: receive a radar point cloud sequence, an ADS-B positioning signal, and a visual image stream, input the visual image stream into a motion compensation module, calculate the motion vector between adjacent frames by optical flow method, correct the jitter of the visual image stream, and output the jitter-corrected visual image stream; S12: perform non-uniform time series analysis on the ADS-B positioning signal, resample the position coordinates at equally spaced time points using the Lagrange interpolation method, and generate resampled ADS-B positioning signal; S13: extract the timestamp of the jitter-corrected visual image stream, map the radar point cloud sequence to the image coordinate system based on the same timestamp, convert the radar three-dimensional point cloud coordinates to two-dimensional image plane coordinates through the perspective transformation matrix, and generate timestamp-aligned radar point cloud mapping data.

[0007] Optionally, after generating the timestamp-aligned radar point cloud mapping data, the resampled ADS-B positioning signal and the timestamp-aligned radar point cloud mapping data are spatially registered in the image coordinate system, the position data is fused through Kalman filtering, and the feature point trajectory of the jitter-corrected visual image stream is combined to generate a spatio-temporally synchronized fusion trajectory feature vector.

[0008] Optionally, the S2 comprises: S21: based on the spatio-temporally synchronized fusion trajectory feature vector, calculate the curvature derivative of the flight path by differential geometry algorithm, extract the flight path points with curvature change rate exceeding a preset threshold, and generate a curvature mutation point set; S22: perform semantic segmentation on the jitter-corrected visual image stream, extract building and terrain edge features, simultaneously analyze the reflection intensity distribution of the timestamp-aligned radar point cloud mapping data, fuse the edge features and reflection intensity distribution to calculate the shielding coverage, and generate an environmental shielding factor; S23: combine the weather code in the resampled ADS-B positioning signal and the attenuation coefficient of the radar point cloud sequence to construct an airfield visibility gradient model, and output a dynamically updated airfield visibility gradient; S24: Stack the curvature discontinuity set, the environmental shielding factor, the airport visibility gradient and the aircraft dynamics constraint parameters by channel to build a multi-dimensional offset feature tensor containing spatial position, environmental risk and ontology constraints.

[0009] Optionally, the S23 comprises: S231: Analyze the weather code in the resampled ADS-B positioning signal, and extract the visibility interference parameters, including cloud height, precipitation intensity and turbulence index; S232: Calculate the attenuation coefficient of the radar point cloud sequence, and quantify the signal attenuation degree through the deviation value of the radar echo power and the theoretical propagation model; S233: Fuse the visibility interference parameters and the attenuation coefficient to build an airport visibility gradient model, which is defined as a three-dimensional function with geographical coordinates and time stamp as input; S234: Dynamically update the parameter weight of the airport visibility gradient model using the time sliding window mechanism, and output the dynamically updated airport visibility gradient.

[0010] Optionally, the pre-trained spatio-temporal graph neural network comprises an input encoding layer, a graph convolution layer, a time convolution layer, a cross-modal fusion layer and a decision output layer.

[0011] Optionally, the S3 comprises: S31: Input the multi-dimensional offset feature tensor into the pre-trained spatio-temporal graph neural network, extract the static structure features of the spatial topology grid through the graph convolution layer, and capture the time evolution pattern of the dynamic obstacle motion vector through the time convolution layer; S32: Based on the static structure features of the spatial topology grid and the time evolution pattern of the dynamic obstacle motion vector, perform cross-modal attention fusion to generate an obstacle threat field that fuses environmental constraints; S33: Calculate the path repair path in the obstacle threat field that fuses environmental constraints, and output the original repair decision vector containing repair angle, climb rate and speed compensation; S34: According to the collision probability of the dynamic obstacle motion vector and the risk level mapping table of the spatial topology grid, risk classification labeling is performed on the original repair decision vector to generate a risk classification repair decision.

[0012] Optionally, the S4 comprises: S41: Inject the risk classification repair decision into the flight control simulator, perform trajectory repair simulation through the six-degree-of-freedom flight dynamics model, and output the simulated flight trajectory data; S42: Calculate the trajectory deviation value based on the Euclidean distance between the simulated flight trajectory data and the reference trajectory, and iteratively optimize the flight control parameters using the backpropagation verification mechanism; S43: judging whether the trajectory deviation value is less than the safety threshold value, if the trajectory deviation value is less than the safety threshold value, generating a final repair instruction, otherwise returning to S3 to adjust the repair angle, the climb rate and the speed compensation parameter in the risk grading repair decision; S44: when the trajectory deviation value is less than the safety threshold value for three consecutive iterations, locking the final repair instruction and transmitting it to the flight control system for execution.

[0013] Optionally, the S42 further comprises: S421: when the trajectory deviation value is less than the safety threshold value for the first time, activating the safety margin compensation module, calculating a dynamic safety margin compensation based on historical trajectory deviation data; S422: superimposing the dynamic safety margin compensation on the repair angle, the climb rate and the speed compensation parameter in the risk grading repair decision, generating a strengthened repair decision; S423: re-injecting the strengthened repair decision into the flight control simulator for back propagation verification, updating the trajectory deviation value.

[0014] The beneficial effects of the present application are: The present application realizes unified processing and spatial registration of three types of sensing information with different sources, inconsistent sampling frequencies and different accuracy levels, by introducing the radar point cloud mapping data with the time stamp aligned with the vision image stream after jitter correction and the ADS-B positioning signal after Lagrange interpolation resampling. Especially, through Kalman filter fusion, combined with the position information supplemented by the visual feature point trajectory, a fusion trajectory representation vector with high robustness and high spatio-temporal consistency is generated, which provides a high-quality basic input for subsequent feature extraction and deviation repair, effectively avoiding flight path judgment errors and false trigger risks caused by a single data source.

[0015] The present application, through the constructed multi-dimensional deviation feature tensor, fuses the curvature mutation point set, the environmental shielding factor, the airfield visibility gradient and the aircraft dynamics constraint parameter, and reflects the current flight path state, environmental threat factors and controllable boundaries in multiple dimensions and multiple channels. In S3, the pre-trained spatio-temporal graph neural network has strong structure modeling and time sequence perception ability, can extract deep features of spatial topology grid and dynamic obstacle motion vector through graph convolution layer and time convolution layer respectively, and construct an obstacle threat field combined with a cross-modal attention fusion mechanism, so as to output a risk grading repair decision with risk level. Compared with traditional rule judgment or two-dimensional path planning method, this mechanism can more effectively adapt to complex changes of flight environment and realize high-precision repair strategy reasoning for the deviated flight path.

[0016] The application introduces a flight control simulator and a six-degree-of-freedom flight dynamics model to simulate the execution of risk classification repair decisions, calculates the deviation value by the Euclidean distance between the trajectory and the reference flight path, and iteratively optimizes the control parameters based on the back propagation verification mechanism to ensure that the repair strategy can be safely executed in the real flight control system. At the same time, a safety margin compensation module is designed, when the deviation value is less than the safety threshold for the first time, the compensation mechanism is automatically activated, the dynamic safety margin compensation amount is calculated combined with the historical flight deviation trend, the repair angle, the climb rate and the speed compensation amount are weighted and adjusted, and the reinforced repair decision is formed. Finally, through multiple rounds of verification and the "satisfy the threshold for three times in succession" criterion, the optimal control command is locked and written into the flight control system, which significantly improves the flight safety and the adaptive robustness of the flight path control. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only illustrate the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0018] Fig. 1 The method flowchart of the embodiment of the present application is shown in the figure. Fig. 2 The S3 flowchart of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0019] The present application will be described in detail below with reference to the drawings and specific embodiments. It should be noted that in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be implemented by those skilled in the art for some known technologies; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the present application.

[0020] It should be noted that in the specification, "one embodiment", "embodiment", "exemplary embodiment", "some embodiments" and the like indicate that the described embodiments can include specific features, structures or characteristics, but not necessarily every embodiment includes the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in combination with an embodiment, it should be within the knowledge of those skilled in the art to realize this feature, structure or characteristic in combination with other embodiments (whether or not explicitly described).

[0021] In general, terminology can be understood at least in part from a context of a use of the terminology. For example, the term “one or more” as used herein, depending at least in part upon a context of a use of the term, can be used to describe any feature, structure, or characteristic in a singular sense or can be used to describe combinations of features, structures or characteristics in a plural sense. Additionally, the term “based on” can be understood as not necessarily being confined to factors that are described and that can refer at least in part to additional factors never specifically described.

[0022] As shown in Figs. 1-2 A low-altitude flight track deviation intelligent repair method for fusing multi-source sensing data, comprising the following steps: S1: receiving radar point cloud sequence, ADS-B positioning signal, visual image stream, correcting the jitter of the visual image stream through the motion compensation module, time resampling the ADS-B signal by using the Lagrange interpolation method, mapping the radar point cloud sequence to the synchronous image coordinate system based on the timestamp, and outputting the spatiotemporal synchronous fusion track representation vector, specifically: S11, visual image stream jitter correction: first, receiving radar point cloud sequence, ADS-B positioning signal and visual image stream from the airborne multi-sensing system, wherein: The radar point cloud sequence is collected by an airborne laser radar or a millimeter wave radar, and contains three-dimensional spatial coordinates and reflection intensity information of the target reflection point; The ADS-B positioning signal is obtained in real time by the airborne ADS-B system, including longitude, latitude, weather code, height, heading angle and other flight parameters; The visual image stream is collected by the embedded airborne camera system at a fixed frame rate, and there is a relatively obvious interframe jitter phenomenon.

[0023] The above visual image stream is input into the motion compensation module, the optical flow method (such as Lucas-Kanade or Farneback method) is used to calculate the motion vector between adjacent frames in the image sequence, and then the motion compensation processing is performed on the image frame by frame, realizing dynamic jitter repair. The jitter-corrected visual image stream output will be used for subsequent timestamp synchronization and spatial registration operations.

[0024] S12, Lagrange resampling of ADS-B positioning signal: non-uniform time sequence analysis is performed on the received ADS-B positioning signal, because the sampling time interval is not equal, it is difficult to directly time four with the image stream. The original ADS-B track coordinate sequence is interpolated and reconstructed by using the Lagrange interpolation method, and each target timestamp (usually aligned with visual image frame timestamp) is the interpolated point, based on the longitude and latitude coordinates of the neighboring ADS-B points, second or third order interpolation is performed to generate a set of equally spaced resampled ADS-B positioning signals. The interpolation sequence has a stable time step, which is convenient for subsequent joint processing with other sensor data.

[0025] S13, radar point cloud timestamp mapping and perspective conversion: extract the timestamp of each frame of the jitter-corrected visual image stream, filter the radar point cloud sequence at the corresponding time point, perform timestamp-based point cloud filtering, and then construct a perspective transformation matrix . The matrix is generated according to the internal and external parameter calibration results of the airborne radar, which completes the conversion from radar three-dimensional coordinate points ) to image two-dimensional pixel coordinate system ): ; wherein, is the image intrinsic matrix, is the rotation and translation matrix from radar to camera. The conversion result is the radar point cloud mapping data on each frame of image, each point of which contains the original three-dimensional position and corresponding image position coordinates , and retains the reflection intensity value and other attributes of the radar point.

[0026] After completing image jitter correction and radar point cloud mapping, the resampled ADS-B positioning signal and the timestamp-aligned radar point cloud mapping data are registered in the unified image coordinate system. Kalman filter is used for multi-source position estimation fusion, wherein the prediction state comes from the ADS-B positioning signal, the observation state is constructed by the radar point cloud data, and the filter iteratively updates the position state: ; wherein, is the current estimated position, is the observation value (radar mapping point), is the Kalman gain, is the observation matrix.

[0027] At the same time, sparse feature points (such as corner points, edge points) in the jitter-corrected visual image stream are extracted, which are used to further assist in position information fusion and trajectory smoothing.

[0028] Finally, the fused ADS-B trajectory points, radar reflection points and image feature trajectories are arranged according to the unified timestamp and spatial coordinate system, and a spatiotemporally synchronized fusion trajectory representation vector with the following structure is constructed: ; wherein, represents the filtered ADS-B position point, reprojecting position of the point cloud in image coordinates, a set of image feature point trajectories.

[0029] S2: Based on the fusion trajectory feature vector, a set of curvature mutation points, an environmental shielding factor, and an airfield visibility gradient are extracted, and a multi-dimensional offset feature tensor is constructed in combination with the aircraft dynamics constraints, specifically: S21, curvature mutation point set extraction: First, the trajectory sequence corresponding to the ADS-B positioning signal filtering result in the spatiotemporally synchronized fusion trajectory feature vector is subjected to difference processing, and the curvature derivative between consecutive trajectory points is calculated based on the difference geometry algorithm, which is used to capture the local change trend of the flight path in three-dimensional space.

[0030] Specifically, let the resampled ADS-B trajectory sequence be: The spatial curvature of each point is calculated as: wherein, is the velocity vector, is the acceleration vector; Further derivation gives the curvature derivative: When exceeds the set curvature change rate threshold (e.g., 0.5 m ), it is determined that the point is a curvature mutation point. A set of curvature mutation points is finally formed, which is used to identify the key turning points where flight deviation may occur.

[0031] S22, environmental shielding factor calculation: Perform a semantic segmentation algorithm such as DeepLabV3+ or SegNet on the jitter-corrected visual image stream to identify building outlines, terrain undulation edges, and obstacle boundaries in the image, obtaining an image-level shielding area label map.

[0032] At the same time, the reflection intensity value of each point is extracted from the timestamp-aligned radar point cloud mapping data to construct a three-dimensional reflectivity distribution map. This intensity value is usually related to the material, density, and shielding degree of the object.

[0033] Fuse the image semantic edge features with the radar reflection intensity map, and use a pixel-level joint mask calculation method to calculate the shielding coverage rate in each frame of image, i.e., the proportion of the shielding area in the entire image. Express this coverage rate in numerical form, define it as the environmental shielding factor, and its range is 0-1. The closer to 1, the worse the visibility and the more serious the shielding.

[0034] S23, airfield visibility gradient construction: Combine the weather encoding in the resampled ADS-B positioning signal and the attenuation coefficient of the radar point cloud sequence to construct an airfield visibility gradient model, outputting a dynamically updated airfield visibility gradient. Its processing flow includes:​​​ S231: Extract the weather code information from the resampled ADS-B positioning signal, and parse out the key visibility interference parameters, including cloud height, precipitation intensity, and turbulence index.

[0035] S232: Calculate the attenuation coefficient for each point in the radar point cloud sequence according to the following formula: ; wherein, is the radar power under ideal propagation, is the measured echo power, i.e., the degree of signal attenuation.

[0036] S233: Fuse the above-mentioned weather parameters and attenuation coefficients to construct an airfield visibility gradient model defined by the following three-dimensional function: cloud height, precipitation intensity, turbulence index, and attenuation coefficient); wherein, the function is a linear weighted model, and the parameter weights are obtained by training historical flight big data, reflecting the actual influence weight of each factor on visibility.

[0037] S234: Introduce a time sliding window mechanism (e.g., updated every 10 seconds) to dynamically adjust the model parameter weights, so that the visibility gradient model can be updated in real time with changes in flight path and weather, and finally output the dynamically updated airfield visibility gradient.

[0038] S24, multi-dimensional offset feature tensor construction: stack and integrate the various features extracted in the above sub-steps in a unified structure. The construction method is as follows: Channel 1: Binary label matrix of curvature mutation points (1 represents mutation point, 0 represents non-mutation); Channel 2: Environmental shielding factor mapping into a shielding heat map according to time sequence; Channel 3: Visibility distribution tensor formed by sampling the airfield visibility gradient according to ( ); Channel 4: Aircraft dynamics constraint parameters from the flight control system, such as maximum climb rate, maximum turn radius, maximum speed, etc., which are statically filled with codes. Align the above-mentioned channel features according to the time axis, stack them along the depth dimension (channel dimension), and generate the final multi-dimensional offset feature tensor; ; wherein, is the trajectory length, is the spatial dimension resolution, is the number of channels (≥4), which will be the core input of the pre-trained spatio-temporal graph neural network in S3.

[0039] S3: input the multi-dimensional offset feature tensor into the pre-trained spatio-temporal graph neural network, associate the dynamic obstacle motion vector with the airspace topology grid, and output the risk classification repair decision including the repair angle, climb rate, and speed compensation amount, specifically: S31, graph neural network input and time series feature extraction: input the multi-dimensional offset feature tensor output by S2 as input, and load the pre-trained spatio-temporal graph neural network model that has completed offline training. The tensor includes four channels (curvature mutation point set, environmental shielding factor, airspace visibility gradient, and aircraft dynamics constraint parameter), wherein: The curvature mutation point set is used to describe the change of the trajectory curve; The environmental shielding factor and the airspace visibility gradient depict the environmental risk; The aircraft dynamics constraint parameter is used to physically limit the repair action.

[0040] The input data stream first enters the input encoding layer of the model, in which the tensor is decoupled into two sub-streams: One is the airspace topology grid, and the structure is derived from the spatial connection relationship formed by the curvature mutation and the environmental factor; The other is the dynamic obstacle motion vector, which represents the relative position and speed of various nearby flying bodies or risk targets appearing within the trajectory range.

[0041] Subsequently, the airspace topology grid data stream enters the graph convolution layer, which extracts the local structure features (static connection) of each node using its adjacency matrix; the dynamic obstacle motion vector data stream enters the time convolution layer, which extracts the time dimension features of the node attributes through one-dimensional convolution operation, and outputs the corresponding time evolution pattern.

[0042] The pre-trained spatio-temporal graph neural network structure includes the following processing layers: Input encoding layer: decoupling the multi-dimensional offset feature tensor into two sub-streams of dynamic obstacle motion vector and airspace topology grid.

[0043] Graph convolution layer: using the adjacency matrix of the airspace topology grid, extract the structure relationship of each position based on the GCN (Graph Convolutional Network) structure.

[0044] Time convolution layer: based on one-dimensional convolution operation to extract the time evolution features of the dynamic obstacle motion vector, processing the motion trend in the time series trajectory.

[0045] Cross-modal fusion layer: using the attention gate mechanism to fuse the spatial structure and time features, and output the semantic enhanced obstacle threat field.

[0046] The decision output layer uses a graph path search algorithm and a risk classification model in the obstacle threat field to output the final risk classification repair decision.

[0047] S32, cross-modal attention fusion: input the feature tensors output by the graph convolution layer and the time convolution layer into the cross-modal fusion layer, and realize the weighted integration of information between different modalities through the attention gating mechanism. The core calculation formula of this mechanism is as follows: ; wherein, is a query vector (such as a static structure feature), is a key vector, is a value vector (such as a dynamic motion feature), is a dimension normalization factor.

[0048] In this way, the combination mode with the highest semantic coupling degree between spatial structure and time evolution is obtained, and an obstacle threat field constrained by fused environmental information is output.

[0049] The obstacle threat field is a three-dimensional tensor, which provides a score set containing spatial risk level and direction repair reference at each trajectory point position.

[0050] S33, original repair decision vector generation: on the basis of the obstacle threat field, the decision output layer in the model calls the reinforcement graph search algorithm, evaluates the optimal path combination of trajectory repair at each high-risk point, calculates the optimal flight path repair action of each node, and generates the original repair decision vector composed of the following three quantities: repair angle: the angle that should be adjusted relative to the current heading direction; climb rate: the required climb / descent rate in the vertical direction; speed compensation: the speed correction amount required during repair.

[0051] The repair decision vector is output in the form of a time series, providing a quantitative basis for subsequent risk classification.

[0052] S34, risk classification repair decision generation: the original repair decision vector needs to be combined with the external risk assessment mechanism for safety level labeling. The specific method is as follows: extract the dynamic obstacle motion vector that coincides with the current trajectory from the multi-dimensional offset feature tensor; calculate the collision probability based on the prediction model, taking into account its speed, direction, and approach rate; compare the position of the trajectory point in the spatial topology grid with the pre-defined risk level mapping table to generate a static risk level (e.g., low, medium, high); fuse the above two risk factors and generate a comprehensive threat coefficient according to the formula: Definition: ; where, is the collision probability, is the static level, is the fusion weight obtained by tuning the historical model.

[0053] Finally, the risk classification repair decision with risk label (such as "low-risk repair", "medium-risk repair", "high-risk avoidance") is output.

[0054] S4: Inject the risk classification repair decision into the flight control simulator for back propagation verification, and generate the final repair instruction when the trajectory deviation value is less than the safety threshold, otherwise return to S3 to adjust the decision parameters, specifically: S41, simulate the repair decision: input the risk classification repair decision output by S3 into the flight control simulation module, which is built based on a high-fidelity flight control simulator and integrates a six-degree-of-freedom flight dynamics model that meets the aviation dynamics standards. The model considers the following 6 state variables: spatial position coordinates; attitude angles (pitch, roll, yaw); velocity vector (linear velocity and angular velocity); power parameters (climb rate and speed compensation); control input (repair angle and other heading adjustment values); external disturbance terms (wind shear, air flow and other environmental disturbances); The simulation trajectory data is output by dynamically solving and running the flight path repair path through the simulator, which includes flight path point sequence, speed change curve, attitude evolution sequence, etc.

[0055] S42, trajectory deviation value calculation and back propagation verification: compare the simulated flight trajectory data with the preset reference flight path, and calculate the deviation value using the point-by-point Euclidean distance difference: ; where, is the position point in the simulation trajectory, is the reference flight path position point, is the trajectory deviation value at time . Accumulate the trajectory deviation values of all time steps to define the overall deviation index: ; If If the deviation exceeds the set safety threshold (e.g., 2.5m), it means that the current risk-level repair decision cannot effectively control the deviation range in the dynamic model. In this case, the backpropagation verification mechanism is executed according to the error gradient to update the flight control parameters in the simulator (e.g., speed compensation ratio, angular velocity limit, pitch adjustment parameters, etc.) to enhance the dynamic adaptability to the original repair action.

[0056] Following step S42, after the safety margin compensation mechanism triggers and reinforces the repair decision generation, a key dynamic safety enhancement mechanism is also included to further reduce the cumulative trajectory risk value, as detailed below: Activate the safety margin compensation module: When the detected trajectory deviation value is first less than the safety threshold (i.e. Threshold, or trigger safety margin compensation module, is used to enhance the robustness of the current repair action. This module extracts deviation trend curves from similar scenarios in the historical database, constructs a statistical model, and calculates the dynamic safety margin compensation amount for the target trajectory point. Its definition is: ; in, For the first time in history The deviation value under similar conditions This is an empirical adjustment coefficient.

[0057] S422, Generate Enhanced Repair Decision: Add the above dynamic safety margin compensation amounts to the current repair parameters respectively: Repair angle ; Climb rate ; Speed ​​compensation amount ; A set of enhanced repair decision parameter vectors is obtained to ensure increased margin space within the control boundary and reduced fault tolerance risk.

[0058] S423, Re-injecting the flight control simulator for verification: The enhanced repair decision is re-inputted into the flight control simulator, a complete trajectory simulation and deviation calculation are performed to form enhanced repaired simulated flight trajectory data, and the backpropagation optimization in step S42 is performed again until the final trajectory deviation control requirements are met.

[0059] S43, Trajectory Deviation Threshold Judgment and Feedback Loop: After each simulation calculation, it is determined whether the current trajectory deviation value is less than the safety threshold. If it is satisfied, the current repair action is retained and the next judgment cycle is entered; if it is not satisfied, the error is fed back to S3, the key parameters (repair angle, climb rate, speed compensation amount) in the risk classification repair decision are adjusted, the repair strategy is regenerated and a new round of simulation verification is entered, forming a closed-loop control process.

[0060] The feedback loop is executed at most three times to ensure response efficiency and instruction stability; S44, final repair instruction locking and control writing: when the track deviation value is less than the safety threshold for three consecutive iterations, it is determined that the current repair action has flight control feasibility and safety margin, and the current repair parameter is immediately locked to generate a final repair instruction; ; Wherein, is the final repair angle, is the final climb rate, is the final speed compensation.

[0061] The final repair instruction is written to the flight control system (such as the autopilot / APU system) through the control bus to ensure that the track repair action is accurately executed according to the preset parameters in the actual flight stage.

[0062] The present application covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present application. In order to make the public have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without the description of these details for those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0063] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principle of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be regarded as the protection scope of the present application.

Claims

1. A method for intelligent repair of low-altitude flight track deviation based on fusion of multi-source sensing data, characterized in that, The method comprises the following steps: S1: receiving a radar point cloud sequence, an ADS-B positioning signal, and a visual image stream, performing jitter correction on the visual image stream through a motion compensation module, performing time resampling on the ADS-B signal by using a Lagrange interpolation method, mapping the radar point cloud sequence to a synchronous image coordinate system based on a timestamp, and outputting a time-space synchronous fusion trajectory representation vector; S2: based on the fusion trajectory representation vector, extracting a curvature mutation point set, an environmental shielding factor, and an airfield visibility gradient, and combining aircraft dynamics constraints to construct a multi-dimensional offset feature tensor; S3: inputting the multi-dimensional offset feature tensor into a pre-trained space-time graph neural network, associating dynamic obstacle motion vectors with space domain topology grids, and outputting a risk classification repair decision including a repair angle, a climb rate, and a speed compensation amount; S4: injecting the risk classification repair decision into a flight control simulator for back propagation verification, generating a final repair instruction when a trajectory deviation value is less than a safety threshold, and otherwise returning to S3 to adjust decision parameters. 2.The low-altitude flight track deviation intelligent repairing method of fusing multi-source sensing data according to claim 1, characterized in that, The S1 comprises: S11: receiving a radar point cloud sequence, an ADS-B positioning signal, and a visual image stream, inputting the visual image stream into a motion compensation module, calculating motion vectors between adjacent frames by using an optical flow method, performing jitter correction on the visual image stream, and outputting a jitter-corrected visual image stream; S12: performing non-uniform time sequence analysis on the ADS-B positioning signal, resampling position coordinates at equally spaced time points by using a Lagrange interpolation method, and generating a resampled ADS-B positioning signal; S13: extracting a timestamp of the jitter-corrected visual image stream, mapping the radar point cloud sequence to an image coordinate system based on the same timestamp, converting radar three-dimensional point cloud coordinates into two-dimensional image plane coordinates through a perspective transformation matrix, and generating timestamp-aligned radar point cloud mapping data. 3.The method of claim 2, wherein, After the timestamp-aligned radar point cloud mapping data is generated, the resampled ADS-B positioning signal and the timestamp-aligned radar point cloud mapping data are spatially registered in the image coordinate system, position data is fused through Kalman filtering, feature point trajectories of the jitter-corrected visual image stream are combined, and a time-space synchronous fusion trajectory representation vector is generated.

4. The method of claim 3, wherein, The S2 comprises: S21: based on the time-space synchronous fusion trajectory representation vector, calculating a flight path curvature derivative by using a differential geometry algorithm, extracting flight path points with a curvature change rate exceeding a preset threshold, and generating a curvature mutation point set; S22: performing semantic segmentation on the jitter-corrected visual image stream, extracting building and terrain edge features, simultaneously analyzing reflection intensity distribution of the timestamp-aligned radar point cloud mapping data, fusing edge features and reflection intensity distribution to calculate a shielding coverage, and generating an environmental shielding factor; S23: combining weather coding in the resampled ADS-B positioning signal and an attenuation coefficient of the radar point cloud sequence to construct an airfield visibility gradient model, and outputting a dynamically updated airfield visibility gradient; S24: stacking the curvature mutation point set, the environmental shielding factor, and the airfield visibility gradient with aircraft dynamics constraint parameters according to channels to construct a multi-dimensional offset feature tensor containing spatial position, environmental risk, and ontology constraints.

5. The method of claim 4, wherein, The S23 comprises: S231: Analyze the weather code in the resampled ADS-B positioning signal, extract the visibility interference parameters, including cloud height, precipitation intensity and turbulence index; S232: Calculate the attenuation coefficient of the radar point cloud sequence, and quantify the signal attenuation degree by the deviation value of the radar echo power and the theoretical propagation model; S233: Fuse the visibility interference parameters and the attenuation coefficient, construct the airfield visibility gradient model, and the model is defined as a three-dimensional function with geographical coordinates and time stamp as input; S234: Dynamically update the parameter weight of the airfield visibility gradient model using the time sliding window mechanism, and output the dynamically updated airfield visibility gradient.

6. The method of claim 5, wherein, The pre-trained spatio-temporal graph neural network includes an input encoding layer, a graph convolution layer, a time convolution layer, a cross-modal fusion layer, and a decision output layer.

7. The method of claim 6, wherein, The S3 includes: S31: Input the multi-dimensional offset feature tensor into the pre-trained spatio-temporal graph neural network, extract the static structure features of the spatial topology grid through the graph convolution layer, and capture the time evolution pattern of the dynamic obstacle motion vector through the time convolution layer; S32: Based on the static structure features of the spatial topology grid and the time evolution pattern of the dynamic obstacle motion vector, perform cross-modal attention fusion to generate an obstacle threat field with fused environmental constraints; S33: Calculate the path repair path in the obstacle threat field with fused environmental constraints, and output an original repair decision vector containing the repair angle, climb rate and speed compensation amount; S34: According to the collision probability of the dynamic obstacle motion vector and the risk level mapping table of the spatial topology grid, the original repair decision vector is risk graded and labeled to generate a risk graded repair decision.

8. The method of claim 7, wherein, The S4 includes: S41: Inject the risk graded repair decision into the flight control simulator, perform trajectory repair simulation through the six-degree-of-freedom flight dynamics model, and output simulated flight trajectory data; S42: Calculate the trajectory deviation value based on the Euclidean distance between the simulated flight trajectory data and the reference trajectory, and iteratively optimize the flight control parameters using the back propagation verification mechanism; S43: Determine whether the trajectory deviation value is less than the safety threshold, if the trajectory deviation value is less than the safety threshold, generate the final repair instruction, otherwise return to S3 to adjust the repair angle, climb rate and speed compensation amount parameters in the risk graded repair decision; S44: When the trajectory deviation value is less than the safety threshold for three consecutive iterations, lock the final repair instruction and transmit it to the flight control system for execution.

9. The method of claim 8, wherein, The S42 further includes: S421: When the trajectory deviation value is less than the safety threshold for the first time, activate the safety margin compensation module, and calculate the dynamic safety margin compensation amount based on the historical trajectory deviation data; S422: Add the dynamic safety margin compensation amount to the repair angle, climb rate and speed compensation amount parameters in the risk graded repair decision to generate a reinforced repair decision; S423: Re-inject the reinforced repair decision into the flight control simulator to perform back propagation verification and update the trajectory deviation value.

Images