An unmanned aerial vehicle flight state safety supervision method based on artificial intelligence

Abstract

The application relates to the technical field of unmanned aerial vehicle safety control, and discloses an unmanned aerial vehicle flight state safety supervision method based on artificial intelligence, which comprises the following steps: step 1, collecting multi-dimensional state parameters of unmanned aerial vehicle flight, wherein the multi-dimensional state parameters at least include motion state parameters, position state parameters, attitude state parameters and body working condition parameters. The application adopts the method of constructing a multi-dimensional flight state digital twin model and fusing multi-source data to form a flight behavior graph, and then utilizes a graph neural network to deeply analyze the internal correlation between nodes in the graph, so that the technical effect of accurately identifying flight attitude instability precursors from multiple coupling factors is achieved. Compared with the technical scheme of comparing an actual flight track with a preset route in the prior art, the application solves the problem that complex attitude abnormalities caused by the combined action of sudden air flow, signal drift and component recession cannot be effectively perceived.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) safety control technology, specifically to an artificial intelligence-based method for monitoring the flight status of UAVs. Background Technology

[0002] During flight, obstacles in the flight path of a drone can affect its flight status. When the spacing between obstacles in the flight path is insufficient for the drone to fly normally, the original flight path of the drone needs to be adjusted to ensure that the drone maintains a safe and stable state during flight.

[0003] Chinese invention patent with publication number CN116700070B discloses a method and system for safety monitoring of the flight status of a drone. The method simulates and predicts the flight status according to preset state test control commands, generates a simulated flight status, and provides a standard reference status. The drone is driven based on the same state test control commands to obtain the first flight status of the drone. The first flight deviation characteristic is determined based on the difference between the first flight status and the simulated flight status.

[0004] However, drones have the following inherent drawbacks that are difficult to overcome during flight:

[0005] Existing technologies monitor flight status by comparing the actual flight trajectory of a drone with a preset route, focusing on the spatial position deviation of the drone. However, they cannot effectively identify flight attitude instability caused by the coupling of multiple factors such as sudden airflow disturbances, navigation signal drift, and performance degradation of airframe components. Therefore, there is a regulatory blind spot for abnormal attitude phenomena that indicate safety risks.

[0006] Existing regulatory methods lack the ability to analyze the historical operational data of individual drones and cannot establish a dedicated health status model based on the accumulated operational characteristics of specific drones over time. Therefore, they cannot provide predictive warnings of abnormal flight status trends and can only respond passively after malfunctions and severe yaws occur, thus missing the best opportunity for intervention.

[0007] The existing technology has a simple judgment logic, which leads to low accuracy in identifying abnormal flight status in complex environments. Furthermore, the response mechanism has inherent delays and cannot effectively intervene within the safety window, directly reducing the scheduling efficiency of airspace resources and posing a threat to the drones themselves and public safety.

[0008] To address these issues, this invention proposes an artificial intelligence-based method for monitoring the flight status of unmanned aerial vehicles (UAVs) to solve the aforementioned problems. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based method for monitoring the flight status of unmanned aerial vehicles (UAVs) to solve the problems mentioned in the background section.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the flight status of unmanned aerial vehicles (UAVs) based on artificial intelligence, comprising the following steps:

[0011] Step 1: Collect multi-dimensional state parameters of the UAV flight. The multi-dimensional state parameters include at least motion state parameters, position state parameters, attitude state parameters, and airframe condition parameters. Based on the multi-dimensional state parameters, construct a digital twin model of the flight state represented by a graph structure. The graph structure includes state nodes and path edges.

[0012] Step 2: Based on the flight status digital twin model, integrate the historical flight data of the UAV, the cluster status data of other UAVs in the same airspace and meteorological data to construct a flight behavior map model, and standardize the node feature data in the flight behavior map model.

[0013] Step 3: Input the node feature data in the standardized flight behavior map model into the graph neural network model, calculate the anomaly score of the node in the graph through the graph neural network model, and identify the flight skew propagation characteristics based on the anomaly score to determine high-risk flight behavior;

[0014] Step 4: When the high-risk flight behavior is identified, determine the potential anomaly source node that generates the highest anomaly score, call the flight status digital twin model to simulate and predict the flight status of the UAV in the future, and generate risk warning and proactive scheduling suggestions based on the simulation prediction results.

[0015] Step 5: Record the push results of the risk warning and proactive scheduling suggestions, as well as the final actual flight results of the UAV. Use the deviation between the push results and the final actual flight results to retrain the graph neural network model and adaptively adjust the threshold used to determine the high-risk flight behavior.

[0016] Preferably, the flight behavior map model is a heterogeneous map model that includes the target UAV status nodes, neighboring UAV status nodes, and airspace meteorological nodes.

[0017] Preferably, in step 3, the graph neural network model is a network model based on graph attention mechanism; the anomaly score is the reconstruction error between the graph neural network model's predicted value of the node state and the actual observed value.

[0018] Preferably, in step 4, the specific method for simulating and predicting the flight status of the UAV in the future is as follows: the state vector of the potential anomaly source node is used as the initial condition for simulation and prediction.

[0019] The proactive scheduling recommendations include at least one of the following: temporary rerouting instructions, hovering stabilization instructions, or deceleration delay instructions.

[0020] Preferably, in step 5, samples identified as false alarms or missed alarms in the deviation between the push result and the final actual flight result are used as hard samples to retrain the graph neural network model.

[0021] The method for adaptively adjusting the threshold used to determine the high-risk flight behavior is as follows: dynamically increasing or decreasing the threshold based on statistical data of early warning accuracy, false alarm rate and missed alarm rate within a preset period.

[0022] Preferably, step 1 further includes:

[0023] Sub-step 1.1: During the execution of the UAV flight mission, at a preset sampling time interval... The sampling is periodic, conducted via an onboard sensor system at each sampling time. Simultaneously collect multi-dimensional state parameters of the UAV, and combine all collected parameters into a multi-dimensional state vector. The multi-dimensional state vector Defined by the following formula:

[0024] ,

[0025] in, , , Let be the velocity components of the UAV in the three-axis coordinate system. , , The acceleration components of the UAV in the three-axis coordinate system. , , The longitude, latitude, and altitude are obtained from the Global Navigation Satellite System. , , The pitch, roll, and yaw angles are obtained from the inertial measurement unit. , The battery voltage and discharge current are obtained from the battery management system;

[0026] Sub-step 1.2, based on the multi-dimensional state vector obtained in sub-step 1.1 , for sampling time The obtained multi-dimensional state vector Create the corresponding state node The state node The attribute information is the multi-dimensional state vector. All parameter values ​​included;

[0027] The adjacent state nodes in the time series and Path edge construction Path edge For drones, there are state nodes To the state node The state transition process;

[0028] Sub-step 1.3 aggregates all state nodes and all path edges generated in sub-step 1.2 to finally form the digital twin model of flight state represented by the graph structure. The flight status digital twin model The structure is defined by the following formula:

[0029] ,

[0030] in, For a set of nodes, , For the final state node, This represents the total number of samples.

[0031] Let be the set of edges. , This is the edge of the final path.

[0032] Preferably, step 2 further includes:

[0033] Sub-step 2.1: Association and node set expansion of multi-source heterogeneous data; Based on the flight state digital twin model constructed in step 1, for this model at the sampling time... State Nodes To correlate multi-source heterogeneous data, first determine the set of nearby drones based on the following distance determination criteria:

[0034] ,

[0035] If the target drone is at time Location Location of another drone Spatial distance Less than the preset distance threshold It identifies another drone as a neighboring drone;

[0036] Then, the core status data of each drone in the set of neighboring drones is obtained to generate neighboring drone status nodes;

[0037] Next, acquire meteorological data corresponding to the gridded airspace where the target drone is located to generate airspace meteorological nodes;

[0038] Finally, the neighboring UAV status nodes and the airspace meteorological nodes are added to the node set of the flight status digital twin model to form an extended node set;

[0039] Sub-step 2.2: Based on the extended node set formed in sub-step 2.1, construct the flight behavior graph model. Flight behavior map model Defined by the following formula:

[0040] ,

[0041] in, For the extended node set, For expanding the edge set;

[0042] Sub-step 2.3 is to eliminate the flight behavior map model constructed in sub-step 2.2. The dimensional differences between the feature data of each node within the set need to be addressed by expanding the node set. Original feature values ​​of the dimensions of the middle node Perform standardization to obtain standardized eigenvalues. Standardization is achieved through the following methods: Standardized formula completed:

[0043] ,

[0044] in, These are the original feature values ​​to be processed. The dimensional feature is the arithmetic mean of the features in the historical data set. denoted as the standard deviation of the dimensional feature in the historical statistical data set.

[0045] Preferably, step 3 further includes:

[0046] Sub-step 3.1: Input the standardized flight behavior map model obtained in step 2 into the graph neural network model, and for any node in the graph... The model calculates the attention coefficient with neighboring nodes using the following formula. Attention coefficient between :

[0047] ,

[0048] in, and For nodes with neighboring nodes The input feature vector, The weight matrix, For a single-layer feedforward neural network, This is a vector concatenation operation. For nodes The set of all neighboring nodes, It is a non-linear activation function;

[0049] Then, the calculated attention coefficient is used The node is obtained by weighted summation of the features of all neighboring nodes. Updated feature vector ;

[0050] Sub-step 3.2 involves updating the feature vectors of the nodes obtained in sub-step 3.1. The input is fed into the decoder to reconstruct the initial feature vector of the node, thus obtaining the reconstructed feature vector. ;

[0051] The node is then calculated using the following reconstruction error formula. Abnormal scores :

[0052] ,

[0053] in, For nodes The final abnormal score, For nodes The original input feature vector, The nodes obtained after encoding and decoding using a graph neural network model Reconstructing the feature vector, The square of the Euclidean distance;

[0054] Sub-step 3.3: Based on the anomaly scores of the nodes in the graph calculated in sub-step 3.2. High-risk flight behaviors are determined using the following high-risk criteria:

[0055] ,

[0056] The high-risk determination criterion is: the existence of a subset of nodes in the flight behavior mapping model. Node subset The conditions to be met include: the anomaly scores of all nodes in the subset. All are greater than the preset node anomaly threshold. subset All nodes within form a connected subgraph Furthermore, the sum of the anomaly scores of all nodes in the subset is greater than the preset propagation risk threshold. When the high-risk determination criteria are met, the node subset will be... The flight behavior was identified as a skewed propagation characteristic, indicating a high-risk flight behavior.

[0057] Preferably, step 4 further includes:

[0058] Sub-step 4.1: After step 3 determines that a high-risk flight behavior has occurred, immediately apply the following to the subset of nodes corresponding to the identified flight skewness propagation characteristics. In this process, the potential anomaly source node that generates the highest anomaly score is determined using the following maximum value index function. :

[0059] ,

[0060] in, For the identified potential source nodes of anomalies, For a subset of nodes, For a subset of nodes Middle node Abnormal scores; The function is used to return the abnormal score. The node that obtains the maximum value ;

[0061] Sub-step 4.2 calls the flight state digital twin model built for the UAV in step 1 as the simulation engine, and identifies the potential anomaly source nodes in sub-step 4.1. The corresponding multi-dimensional state vector As the initial state predicted in the simulation;

[0062] Through state transition function For the future continuum The flight state at each time step is iteratively extrapolated to generate a prediction sequence composed of future state vectors. The predicted sequence is defined by the following formula:

[0063] ,

[0064] in, For the future The predicted state vector at each time step, and the initial state. equal ;

[0065] The state transition function is derived from the digital twin model of flight state. This is the preset total number of prediction steps;

[0066] Sub-step 4.3, the predicted sequence generated in sub-step 4.2 Perform a multi-condition risk assessment. When the state vector in the predicted sequence meets any of the following risk conditions, generate a corresponding risk warning and proactive scheduling suggestion:

[0067] Condition 1: If the predicted sequence There exists a minimum one-bit predicted state vector. attitude angle or The absolute value exceeds the preset attitude stability threshold. It generates a "high risk of attitude instability" warning and a "perform hovering stability" scheduling suggestion;

[0068] Condition 2: If the predicted sequence There exists a minimum one-bit predicted state vector. Location coordinates Spatial distance to the point corresponding to the planned route greater than the preset route deviation threshold It generates "route deviation risk" warnings and "implement temporary rerouting" scheduling suggestions;

[0069] Condition 3: If the predicted sequence The last predicted state vector Battery voltage Below the minimum voltage threshold required for safe flight It generates a "risk of insufficient battery range" warning and a "deceleration delay" scheduling suggestion.

[0070] Preferably, step 5 further includes:

[0071] Sub-step 5.1: After the risk warning and proactive scheduling suggestion generated in step 4 are executed, the system captures the final actual flight result of the UAV and encapsulates the risk warning, scheduling suggestion, the final actual flight result of the UAV, and the flight behavior map model when the warning occurs into a structured feedback data unit.

[0072] The feedback data units are then classified according to the following deviation determination criteria to determine the deviation type:

[0073] Condition 1: If the system generates a risk warning and the actual flight results verify the existence of the risk, then the deviation type of the feedback data unit will be marked as "accurate warning".

[0074] Condition 2: If the system generates a risk warning but the final actual flight results show that no risk occurred, the deviation type of the feedback data unit will be marked as "false alarm".

[0075] Condition 3: If the system does not generate a risk warning but the drone subsequently experiences a flight anomaly that fails to pass the warning, the deviation type of the feedback data unit will be marked as "missed report".

[0076] Sub-step 5.2: From the feedback data units accumulated in sub-step 5.1, select all feedback data units whose deviation type is marked as "false alarm" or "missed alarm" as a hard sample set;

[0077] When the preset retraining period arrives, the graph neural network model from step 3 is incrementally retrained using the hard sample set. The following weighted loss function is used during the retraining process. Perform model parameter optimization:

[0078] ,

[0079] in, The final weighted loss used for model optimization, The total number of training samples. For a single sample, the original loss function is... For sample weights;

[0080] Sub-step 5.3: Using a preset evaluation period as the unit, calculate the false alarm rate within that period based on the deviation type marked in sub-step 5.1. With the false negative rate ;

[0081] Then, the node anomaly threshold used to determine high-risk flight behavior in step 3 is adjusted according to the following threshold adjustment function. With transmission risk threshold Perform adaptive adjustments to generate updated thresholds. and :

[0082] ,

[0083] ,

[0084] in, and The target false positive rate and the target false negative rate are preset. This is the false alarm rate adjustment factor. This is an adjustment factor for the false negative rate;

[0085] When the actual false positive rate is higher than the target value, the threshold is raised to reduce false positives; when the actual false negative rate is higher than the target value, the threshold is lowered to improve recall.

[0086] This invention provides an artificial intelligence-based method for monitoring the flight status of unmanned aerial vehicles (UAVs). It offers the following advantages:

[0087] 1. This invention employs a method of constructing a multi-dimensional digital twin model of flight status and fusing multi-source data to form a flight behavior map. Then, it utilizes graph neural networks to deeply analyze the intrinsic correlation between nodes in the map. This achieves the technical effect of accurately identifying the precursors of flight attitude instability from multiple coupling factors. Compared with the existing technology that compares the actual flight trajectory with the preset route, this invention solves the problem of not being able to effectively perceive complex attitude anomalies caused by the combined effects of sudden airflow, signal drift, and component degradation.

[0088] 2. This invention employs a method that identifies flight deviation propagation characteristics and combines them with a digital twin model to simulate and predict future flight states. This achieves the technical effect of proactively warning and providing intervention strategies before a failure occurs. Compared with existing technologies that passively respond after a failure or severe yaw occurs, this invention addresses the shortcomings of lacking predictive monitoring capabilities and experiencing delayed responses that lead to missed optimal intervention opportunities.

[0089] 3. This invention adopts a feedback closed-loop mechanism, which uses the deviation between actual flight results and early warning judgments to incrementally retrain the graph neural network model and adaptively adjust the judgment threshold. This achieves the technical effect of continuously optimizing the model and realizing high-precision personalized supervision. Compared with the existing technology that uses fixed and simple judgment logic for supervision, this invention solves the shortcomings of low recognition accuracy in complex environments and the inability to effectively guarantee airspace scheduling efficiency and operational safety. Attached Figure Description

[0090] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0091] To enable those skilled in the art to understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0092] The present invention will now be described in detail with reference to the accompanying drawings: Example

[0093] Please see the appendix Figure 1 This invention provides an artificial intelligence-based method for monitoring the flight status of unmanned aerial vehicles (UAVs), comprising the following steps:

[0094] Step 1: Collect multi-dimensional state parameters of the UAV flight. The multi-dimensional state parameters include at least motion state parameters, position state parameters, attitude state parameters and airframe operating condition parameters. Based on the multi-dimensional state parameters, construct a digital twin model of the flight state represented by a graph structure. The graph structure contains state nodes and path edges.

[0095] Step 1 further includes:

[0096] Sub-step 1.1: During the execution of the UAV flight mission, at a preset sampling time interval... The sampling is periodic, conducted via an onboard sensor system at each sampling time. Simultaneously collect multi-dimensional state parameters of the UAV, and combine all collected parameters into a multi-dimensional state vector. Multidimensional state vector Defined by the following formula:

[0097] ,

[0098] in, , , Let be the velocity components of the UAV in the three-axis coordinate system. , , The acceleration components of the UAV in the three-axis coordinate system. , , The longitude, latitude, and altitude are obtained from the Global Navigation Satellite System. , , The pitch, roll, and yaw angles are obtained from the inertial measurement unit. , The battery voltage and discharge current are obtained from the battery management system;

[0099] Sub-step 1.2, based on the multi-dimensional state vector obtained in sub-step 1.1 , for sampling time The obtained multi-dimensional state vector Create the corresponding state node State Node The attribute information is the multi-dimensional state vector. All parameter values ​​included;

[0100] The adjacent state nodes in the time series and Path edge construction Path edge For drones, there are state nodes To the state node The state transition process;

[0101] Sub-step 1.3 aggregates all state nodes and all path edges generated in sub-step 1.2 to finally form a digital twin model of the flight state represented by a graph structure. Flight status digital twin model The structure is defined by the following formula:

[0102] ,

[0103] in, For a set of nodes, , For the final state node, This represents the total number of samples.

[0104] Let be the set of edges. , For the final path edge;

[0105] Step 2: Based on the digital twin model of flight status, integrate the historical flight data of UAVs, the cluster status data of other UAVs in the same airspace and meteorological data to construct a flight behavior map model, and standardize the node feature data in the flight behavior map model.

[0106] The flight behavior graph model is a heterogeneous graph model that includes the target UAV state nodes, neighboring UAV state nodes, and airspace meteorological nodes;

[0107] Step 2 further includes:

[0108] Sub-step 2.1: Association and node set expansion of multi-source heterogeneous data; Based on the flight state digital twin model constructed in step 1, for this model at the sampling time... State Nodes To correlate multi-source heterogeneous data, first determine the set of nearby drones based on the following distance determination criteria:

[0109] ,

[0110] If the target drone is at time Location Location of another drone Spatial distance Less than the preset distance threshold It identifies another drone as a neighboring drone;

[0111] Then, the core status data of each drone in the set of neighboring drones is obtained to generate neighboring drone status nodes;

[0112] Next, acquire meteorological data corresponding to the gridded airspace where the target drone is located to generate airspace meteorological nodes;

[0113] Finally, the neighboring UAV status nodes and the airspace meteorological nodes are added to the node set of the flight status digital twin model to form an extended node set;

[0114] Sub-step 2.2: Based on the extended node set formed in sub-step 2.1, construct the flight behavior graph model. Flight behavior map model Defined by the following formula:

[0115] ,

[0116] in, For the extended node set, For expanding the edge set;

[0117] Sub-step 2.3 is to eliminate the flight behavior map model constructed in sub-step 2.2. The dimensional differences between the feature data of each node within the set need to be addressed by expanding the node set. Original feature values ​​of the dimensions of the middle node Perform standardization to obtain standardized eigenvalues. Standardization is achieved through the following methods: Standardized formula completed:

[0118] ,

[0119] in, These are the original feature values ​​to be processed. The dimensional feature is the arithmetic mean of the features in the historical data set. The standard deviation of the dimensional feature in the historical statistical data set;

[0120] Step 3: Input the node feature data in the standardized flight behavior map model into the graph neural network model, calculate the anomaly score of the node in the graph through the graph neural network model, and identify the flight skew propagation characteristics based on the anomaly score to determine high-risk flight behavior;

[0121] In step 3, the graph neural network model is a network model based on the graph attention mechanism; the anomaly score is the reconstruction error between the graph neural network model's predicted value of the node state and the actual observed value;

[0122] Step 3 further includes:

[0123] Sub-step 3.1: Input the standardized flight behavior map model obtained in step 2 into the graph neural network model, and for any node in the graph... The model calculates the attention coefficient with neighboring nodes using the following formula. Attention coefficient between :

[0124] ,

[0125] in, and For nodes with neighboring nodes The input feature vector, The weight matrix, For a single-layer feedforward neural network, This is a vector concatenation operation. For nodes The set of all neighboring nodes, It is a non-linear activation function;

[0126] Then, the calculated attention coefficient is used The node is obtained by weighted summation of the features of all neighboring nodes. Updated feature vector ;

[0127] Sub-step 3.2 involves updating the feature vectors of the nodes obtained in sub-step 3.1. The input is fed into the decoder to reconstruct the initial feature vector of the node, thus obtaining the reconstructed feature vector. ;

[0128] The node is then calculated using the following reconstruction error formula. Abnormal scores :

[0129] ,

[0130] in, For nodes The final abnormal score, For nodes The original input feature vector, The nodes obtained after encoding and decoding using a graph neural network model Reconstructing the feature vector, The square of the Euclidean distance;

[0131] Sub-step 3.3: Based on the anomaly scores of the nodes in the graph calculated in sub-step 3.2. High-risk flight behaviors are determined using the following high-risk criteria:

[0132] ,

[0133] The high-risk determination criterion is: the existence of a subset of nodes in the flight behavior mapping model. Node subset The conditions to be met include: the anomaly scores of all nodes in the subset. All are greater than the preset node anomaly threshold. subset All nodes within form a connected subgraph Furthermore, the sum of the anomaly scores of all nodes in the subset is greater than the preset propagation risk threshold. When the high-risk determination criteria are met, the node subset will be... The flight behavior was identified as a skewed propagation characteristic and determined to be a high-risk flight behavior.

[0134] Step 4: When high-risk flight behavior is identified, the potential anomaly source node that generates the highest anomaly score is determined, the flight status digital twin model is called to simulate and predict the flight status of the UAV in the future period, and risk warning and proactive scheduling suggestions are generated based on the simulation prediction results.

[0135] In step 4, the specific method for simulating and predicting the flight status of the UAV in the future is as follows: the state vector of the potential anomaly source node is used as the initial condition for simulation and prediction.

[0136] Active scheduling recommendations should include at least one of the following: temporary rerouting instructions, hovering stabilization instructions, or deceleration delay instructions;

[0137] Step 4 further includes:

[0138] Sub-step 4.1: After step 3 determines that a high-risk flight behavior has occurred, immediately apply the following to the subset of nodes corresponding to the identified flight skewness propagation characteristics. In this process, the potential anomaly source node that generates the highest anomaly score is determined using the following maximum value index function. :

[0139] ,

[0140] in, For the identified potential source nodes of anomalies, For a subset of nodes, For a subset of nodes Middle node Abnormal scores; The function is used to return the abnormal score. The node that obtains the maximum value ;

[0141] Sub-step 4.2 calls the digital twin model of the flight state built for the UAV in step 1 as the simulation engine, and identifies the potential anomaly source nodes in sub-step 4.1. The corresponding multi-dimensional state vector As the initial state predicted in the simulation;

[0142] Through state transition function For the future continuum The flight state at each time step is iteratively extrapolated to generate a prediction sequence composed of future state vectors. The predicted sequence is defined by the following formula:

[0143] ,

[0144] in, For the future The predicted state vector at each time step, and the initial state. equal ;

[0145] The state transition function is derived from the digital twin model of flight state. This is the preset total number of prediction steps;

[0146] Sub-step 4.3, the predicted sequence generated in sub-step 4.2 Perform a multi-condition risk assessment. When the state vector in the predicted sequence meets any of the following risk conditions, generate a corresponding risk warning and proactive scheduling suggestion:

[0147] Condition 1: If the predicted sequence There exists a minimum one-bit predicted state vector. attitude angle or The absolute value exceeds the preset attitude stability threshold. It generates a "high risk of attitude instability" warning and a "perform hovering stability" scheduling suggestion;

[0148] Condition 2: If the predicted sequence There exists a minimum one-bit predicted state vector. Location coordinates Spatial distance to the point corresponding to the planned route greater than the preset route deviation threshold It generates "route deviation risk" warnings and "implement temporary rerouting" scheduling suggestions;

[0149] Condition 3: If the predicted sequence The last predicted state vector Battery voltage Below the minimum voltage threshold required for safe flight It generates a "risk of insufficient battery range" warning and a "deceleration delay" scheduling suggestion;

[0150] Step 5: Record the push results of risk warnings and proactive scheduling suggestions, as well as the final actual flight results of the UAV. Use the deviation between the push results and the final actual flight results to retrain the graph neural network model and adaptively adjust the threshold used to judge high-risk flight behaviors.

[0151] In step 5, samples identified as false alarms or missed alarms in the deviation between the push results and the final actual flight results are used as hard samples to retrain the graph neural network model.

[0152] The method for adaptively adjusting the threshold used to determine high-risk flight behavior is to dynamically raise or lower the threshold based on the statistical data of warning accuracy, false alarm rate and missed alarm rate within a preset period.

[0153] Step 5 further includes:

[0154] Sub-step 5.1: After the risk warning and proactive scheduling suggestion generated in step 4 are executed, the system captures the final actual flight result of the UAV and encapsulates the risk warning, scheduling suggestion, the final actual flight result of the UAV, and the flight behavior map model when the warning occurs into a structured feedback data unit.

[0155] The feedback data units are then classified according to the following deviation judgment criteria to determine the deviation type:

[0156] Condition 1: If the system generates a risk warning and the actual flight results verify the existence of the risk, then the deviation type of the feedback data unit will be marked as "accurate warning".

[0157] Condition 2: If the system generates a risk warning but the final actual flight results show that no risk occurred, the deviation type of the feedback data unit will be marked as "false alarm".

[0158] Condition 3: If the system does not generate a risk warning but the drone subsequently experiences a flight anomaly that fails to pass the warning, the deviation type of the feedback data unit will be marked as "missed report".

[0159] Sub-step 5.2: From the feedback data units accumulated in sub-step 5.1, select all feedback data units whose deviation type is marked as "false alarm" or "missed alarm" as a hard sample set;

[0160] When the preset retraining period arrives, the graph neural network model from step 3 is incrementally retrained using the hard sample set. The following weighted loss function is used during the retraining process. Perform model parameter optimization:

[0161] ,

[0162] in, The final weighted loss used for model optimization, The total number of training samples. For a single sample, the original loss function is... For sample weights;

[0163] Sub-step 5.3: Using a preset evaluation period as the unit, calculate the false alarm rate within that period based on the deviation type marked in sub-step 5.1. With the false negative rate ;

[0164] Then, the node anomaly threshold used to determine high-risk flight behavior in step 3 is adjusted according to the following threshold adjustment function. With transmission risk threshold Perform adaptive adjustments to generate updated thresholds. and :

[0165] ,

[0166] ,

[0167] in, and The target false positive rate and the target false negative rate are preset. This is the false alarm rate adjustment factor. This is an adjustment factor for the false negative rate;

[0168] When the actual false positive rate is higher than the target value, the threshold is raised to reduce false positives; when the actual false negative rate is higher than the target value, the threshold is lowered to improve recall.

[0169] Step 1 involves comprehensively collecting multi-dimensional state parameters of the UAV using high-frequency sampling, and then constructing them in a structured manner based on time series data into an individualized digital twin model of flight state containing state nodes and path edges, thereby achieving a high-fidelity digital mapping of the UAV's flight characteristics throughout its entire lifecycle.

[0170] Step 2 involves placing the digital twin model of a single UAV's flight status within a broad spatiotemporal context, integrating historical data, cluster data, and environmental data to construct a flight behavior graph model containing diverse heterogeneous nodes. This enables the construction of a global and contextualized view of complex flight scenarios, allowing subsequent intelligent analysis to move beyond isolated individual UAVs and fully consider the coupled impact of machine-to-machine and machine-to-environment interactions. Simultaneously, standardized processing eliminates dimensional barriers between multi-source data.

[0171] Step 3 introduces a graph neural network model based on graph attention mechanism to perform end-to-end deep learning analysis on the flight behavior map, thereby realizing intelligent and forward-looking identification of flight safety risks. This method can automatically learn and quantify the complex dependencies between nodes, and then capture the flight skew propagation characteristics caused by multi-factor coupling that are difficult to detect by traditional threshold methods. This completes the fundamental transformation from passive monitoring to active identification, and greatly improves the ability to perceive early and weak fault signs.

[0172] Step 4 involves quickly identifying potential sources of anomalies after determining high-risk flight behavior and using a digital twin model to simulate and predict flight status in future time periods. This achieves a closed-loop generation from risk warning to proactive and executable intervention strategies. The method clarifies the root cause of the risk, predicts the development trend of the situation through forward-looking deduction, and generates specific and clear scheduling suggestions accordingly.

[0173] Step 5 establishes a feedback loop and adaptive optimization mechanism to enable the regulatory methodology system to evolve and iterate continuously. This method utilizes the actual effectiveness of early warnings, especially by using false alarms and missed alarms as hard samples to retrain the neural network model. It also dynamically adjusts the risk judgment threshold in conjunction with macro performance statistics to ensure that the regulatory model can adapt to aging, environmental changes, and the introduction of new drone models, thus guaranteeing the advanced nature and high reliability of the regulatory methodology throughout its entire lifecycle.

[0174] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the flight status of unmanned aerial vehicles (UAVs) based on artificial intelligence, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional state parameters of the UAV flight. The multi-dimensional state parameters include at least motion state parameters, position state parameters, attitude state parameters, and airframe condition parameters. Based on the multi-dimensional state parameters, construct a digital twin model of the flight state represented by a graph structure. The graph structure includes state nodes and path edges. Step 2: Based on the flight status digital twin model, integrate the historical flight data of the UAV, the cluster status data of other UAVs in the same airspace, and meteorological data to construct a flight behavior graph model, and standardize the node feature data in the flight behavior graph model; the flight behavior graph model is a heterogeneous graph model that includes the target UAV status nodes, the neighboring UAV status nodes, and the airspace meteorological nodes. Step 3: Input the node feature data in the standardized flight behavior map model into the graph neural network model, calculate the anomaly score of the node in the graph through the graph neural network model, identify the flight skew propagation characteristics based on the anomaly score, and determine high-risk flight behavior; the anomaly score is the reconstruction error between the graph neural network model's predicted value of the node state and the actual observed value; Step 4: When the high-risk flight behavior is identified, determine the potential anomaly source node that generates the highest anomaly score, call the flight status digital twin model to simulate and predict the flight status of the UAV in the future, and generate risk warning and proactive scheduling suggestions based on the simulation prediction results. Step 5: Record the push results of the risk warning and proactive scheduling suggestions, as well as the final actual flight results of the UAV. Use the deviation between the push results and the final actual flight results to retrain the graph neural network model and adaptively adjust the threshold used to determine the high-risk flight behavior.

2. The method for monitoring the flight status of unmanned aerial vehicles based on artificial intelligence according to claim 1, characterized in that, In step 3, the graph neural network model is a network model based on the graph attention mechanism.

3. The method for monitoring the flight status of unmanned aerial vehicles (UAVs) based on artificial intelligence according to claim 1, characterized in that, In step 4, the specific method for simulating and predicting the flight status of the UAV in the future is as follows: the state vector of the potential anomaly source node is used as the initial condition for simulation and prediction. The proactive scheduling recommendations include at least one of the following: temporary rerouting instructions, hovering stabilization instructions, or deceleration delay instructions.

4. The method for monitoring the flight status of unmanned aerial vehicles based on artificial intelligence according to claim 1, characterized in that, In step 5, samples identified as false alarms or missed alarms in the deviation between the push result and the final actual flight result are used as hard samples to retrain the graph neural network model. The method for adaptively adjusting the threshold used to determine the high-risk flight behavior is as follows: dynamically increasing or decreasing the threshold based on statistical data of early warning accuracy, false alarm rate and missed alarm rate within a preset period.

5. The method for monitoring the flight status of unmanned aerial vehicles based on artificial intelligence according to claim 1, characterized in that, Step 1 further includes: Sub-step 1.1: During the execution of the UAV flight mission, at a preset sampling time interval... The sampling is periodic, conducted via an onboard sensor system at each sampling time. Simultaneously collect multi-dimensional state parameters of the UAV, and combine all collected parameters into a multi-dimensional state vector. The multi-dimensional state vector Defined by the following formula: ， in, , , Let be the velocity components of the UAV in the three-axis coordinate system. , , The acceleration components of the UAV in the three-axis coordinate system. , , The longitude, latitude, and altitude are obtained from the Global Navigation Satellite System. , , The pitch, roll, and yaw angles are obtained from the inertial measurement unit. , The battery voltage and discharge current are obtained from the battery management system; Sub-step 1.2, based on the multi-dimensional state vector obtained in sub-step 1.1 , for sampling time The obtained multi-dimensional state vector Create the corresponding state node The state node The attribute information is the multi-dimensional state vector. All parameter values ​​included; The adjacent state nodes in the time series and Path edge construction Path edge For drones, there are state nodes To the state node The state transition process; Sub-step 1.3 aggregates all state nodes and all path edges generated in sub-step 1.2 to finally form the digital twin model of flight state represented by the graph structure. The flight status digital twin model The structure is defined by the following formula: ， in, For a set of nodes, , For the final state node, This represents the total number of samples. Let be the set of edges. , This is the edge of the final path.

6. The method for monitoring the flight status of unmanned aerial vehicles based on artificial intelligence according to claim 1, characterized in that, Step 2 further includes: Sub-step 2.1: Association and node set expansion of multi-source heterogeneous data; Based on the flight state digital twin model constructed in step 1, for this model at the sampling time... State Nodes To correlate multi-source heterogeneous data, first determine the set of nearby drones based on the following distance determination criteria: ≤ ， If the target drone is at time Location Location of another drone Spatial distance Less than the preset distance threshold It identifies another drone as a neighboring drone; Then, the core status data of each drone in the set of neighboring drones is obtained to generate neighboring drone status nodes; Next, acquire meteorological data corresponding to the gridded airspace where the target drone is located to generate airspace meteorological nodes; Finally, the neighboring UAV status nodes and the airspace meteorological nodes are added to the node set of the flight status digital twin model to form an extended node set; Sub-step 2.2: Based on the extended node set formed in sub-step 2.1, construct the flight behavior graph model. Flight behavior map model Defined by the following formula: ， in, For the extended node set, For expanding the edge set; Sub-step 2.3 is to eliminate the flight behavior map model constructed in sub-step 2.

2. The dimensional differences between the feature data of each node within the set need to be addressed by expanding the node set. Original feature values ​​of the dimensions of the middle node Perform standardization to obtain standardized eigenvalues. Standardization is achieved through the following methods: Standardized formula completed: ， in, These are the original feature values ​​to be processed. The dimensional feature is the arithmetic mean of the features in the historical data set. denoted as the standard deviation of the dimensional feature in the historical statistical data set.

7. The method for monitoring the flight status of unmanned aerial vehicles based on artificial intelligence according to claim 1, characterized in that, Step 3 further includes: Sub-step 3.1: Input the standardized flight behavior map model obtained in step 2 into the graph neural network model, and for any node in the graph... The model calculates the attention coefficient with neighboring nodes using the following formula. Attention coefficient between : ， in, and For nodes with neighboring nodes The input feature vector, The weight matrix, For a single-layer feedforward neural network, This is a vector concatenation operation. For nodes The set of all neighboring nodes, It is a non-linear activation function; Then, the calculated attention coefficient is used The node is obtained by weighted summation of the features of all neighboring nodes. Updated feature vector ; Sub-step 3.2 involves updating the feature vectors of the nodes obtained in sub-step 3.

1. The input is fed into the decoder to reconstruct the initial feature vector of the node, thus obtaining the reconstructed feature vector. ; The node is then calculated using the following reconstruction error formula. Abnormal scores : ， in, For nodes The final abnormal score, For nodes The original input feature vector, The nodes obtained after encoding and decoding using a graph neural network model Reconstructing the feature vector, The square of the Euclidean distance; Sub-step 3.3: Based on the anomaly scores of the nodes in the graph calculated in sub-step 3.

2. High-risk flight behaviors are determined using the following high-risk criteria: ， The high-risk determination criterion is: the existence of a subset of nodes in the flight behavior mapping model. Node subset The conditions to be met include: the anomaly scores of all nodes in the subset. All are greater than the preset node anomaly threshold. subset All nodes within form a connected subgraph Furthermore, the sum of the anomaly scores of all nodes in the subset is greater than the preset propagation risk threshold. When the high-risk determination criteria are met, the node subset will be... The flight behavior was identified as a skewed propagation characteristic, indicating a high-risk flight behavior.

8. The method for monitoring the flight status of unmanned aerial vehicles based on artificial intelligence according to claim 1, characterized in that, Step 4 further includes: Sub-step 4.1: After step 3 determines that a high-risk flight behavior has occurred, immediately apply the following to the subset of nodes corresponding to the identified flight skewness propagation characteristics. In this process, the potential anomaly source node that generates the highest anomaly score is determined using the following maximum value index function. : ， in, For the identified potential source nodes of anomalies, For a subset of nodes, For a subset of nodes Middle node Abnormal scores; The function is used to return the abnormal score. The node that obtains the maximum value ; Sub-step 4.2 calls the flight state digital twin model built for the UAV in step 1 as the simulation engine, and identifies the potential anomaly source nodes in sub-step 4.

1. The corresponding multi-dimensional state vector As the initial state predicted in the simulation; Through state transition function For the future continuum The flight state at each time step is iteratively extrapolated to generate a prediction sequence composed of future state vectors. The predicted sequence is defined by the following formula: ， in, For the future The predicted state vector at each time step, and the initial state. equal ; The state transition function is derived from the digital twin model of flight state. This is the preset total number of prediction steps; Sub-step 4.3, the predicted sequence generated in sub-step 4.2 Perform a multi-condition risk assessment. When the state vector in the predicted sequence meets any of the following risk conditions, generate a corresponding risk warning and proactive scheduling suggestion: Condition 1: If the predicted sequence There exists a minimum one-bit predicted state vector. attitude angle or The absolute value exceeds the preset attitude stability threshold. It generates a "high risk of attitude instability" warning and a "perform hovering stability" scheduling suggestion; Condition 2: If the predicted sequence There exists a minimum one-bit predicted state vector. Location coordinates Spatial distance to the point corresponding to the planned route greater than the preset route deviation threshold It generates "route deviation risk" warnings and "implement temporary rerouting" scheduling suggestions; Condition 3: If the predicted sequence The last predicted state vector Battery voltage Below the minimum voltage threshold required for safe flight It generates a "risk of insufficient battery range" warning and a "deceleration delay" scheduling suggestion.

9. The method for monitoring the flight status of unmanned aerial vehicles based on artificial intelligence according to claim 1, characterized in that, Step 5 further includes: Sub-step 5.1: After the risk warning and proactive scheduling suggestion generated in step 4 are executed, the system captures the final actual flight result of the UAV and encapsulates the risk warning, scheduling suggestion, the final actual flight result of the UAV, and the flight behavior map model when the warning occurs into a structured feedback data unit. The feedback data units are then classified according to the following deviation determination criteria to determine the deviation type: Condition 1: If the system generates a risk warning and the actual flight results verify the existence of the risk, then the deviation type of the feedback data unit will be marked as "accurate warning". Condition 2: If the system generates a risk warning but the actual flight results show that no risk occurred, the deviation type of the feedback data unit will be marked as "false alarm". Condition 3: If the system does not generate a risk warning but the drone subsequently experiences a flight anomaly that fails to pass the warning, the deviation type of the feedback data unit will be marked as "missed report". Sub-step 5.2: From the feedback data units accumulated in sub-step 5.1, select all feedback data units whose deviation type is marked as "false alarm" or "missed alarm" as a hard sample set; When the preset retraining period arrives, the graph neural network model from step 3 is incrementally retrained using the hard sample set. The following weighted loss function is used during the retraining process. Perform model parameter optimization: ， in, The final weighted loss used for model optimization, The total number of training samples. For a single sample, the original loss function is... For sample weights; Sub-step 5.3: Using a preset evaluation period as the unit, calculate the false alarm rate within that period based on the deviation type marked in sub-step 5.

1. With the false negative rate ; Then, the node anomaly threshold used to determine high-risk flight behavior in step 3 is adjusted according to the following threshold adjustment function. With transmission risk threshold Perform adaptive adjustments to generate updated thresholds. and : ， ， in, and The target false positive rate and the target false negative rate are preset. This is the false alarm rate adjustment factor. This is an adjustment factor for the false negative rate; When the actual false positive rate is higher than the target value, the threshold is raised to reduce false positives; when the actual false negative rate is higher than the target value, the threshold is lowered to improve recall.

Images