AI-based low-altitude traffic situation awareness and early warning method

Abstract

The invention relates to the technical field of AI-based low-altitude traffic situation awareness and early warning methods, particularly discloses an AI-based low-altitude traffic situation awareness and early warning method, and aims to solve the problems that an existing system is insufficient in aircraft intention recognition, inaccurate in interactive prediction and inoperable in early warning. The method comprises the following steps: fusing multi-source heterogeneous data to construct a dynamic traffic situation map; generating individual behavior feature vectors; reasoning a high-order task intention; establishing a multi-agent game deduction model; and quantifying the risk and generating graded early warning. According to the technical scheme, transition from shallow perception to deep cognition is achieved, prediction accuracy and early warning operability in a complex low-altitude scene are remarkably improved, cloud training and edge real-time reasoning are supported, and high compatibility and continuous evolution capacity are achieved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and air traffic management. Specifically, it relates to a system and method that utilizes multi-level artificial intelligence technology to deeply understand and accurately predict the behavior, intentions, and interaction relationships of various manned / unmanned aircraft in complex low-altitude environments, and thereby achieve graded and forward-looking safety early warning. Background Technology

[0002] With the gradual liberalization of low-altitude airspace management policies, new types of aircraft such as drones and electric vertical takeoff and landing (eVTOL) aircraft are being deployed on a large scale in scenarios such as urban logistics, emergency rescue, and aerial inspection. The low-altitude traffic system is rapidly evolving into a complex operating environment characterized by high density, strong heterogeneity, and fast dynamics. Traditional low-altitude situational awareness and early warning systems mainly rely on conflict detection logic based on geometric extrapolation and a preset rule base for risk assessment. Their core assumption is that aircraft motion is deterministic and independent. However, in real low-altitude scenarios, aircraft behavior is driven by multiple factors such as mission objectives, environmental constraints, and interactions with other aircraft, exhibiting highly nonlinear and strategic characteristics. This makes it difficult for perception mechanisms that rely solely on shallow state information such as position and speed to support a deep understanding of the traffic situation and proactive intervention.

[0003] In particular, low-altitude traffic situational awareness urgently needs to leap from "trajectory observation" to "intent cognition." While existing systems can acquire the spatiotemporal trajectory of aircraft through radar, ADS-B, and other means, they cannot analyze the underlying behavioral semantics, such as distinguishing between normal cruising and emergency avoidance, mission execution and uncontrolled drift. Furthermore, current mainstream prediction methods mostly employ constant speed or uniform acceleration models for short-term extrapolation, ignoring the active decision-making capabilities of aircraft as intelligent agents, leading to frequent missed alarms or high false alarm rates in densely overlapping areas. More critically, in complex airspaces with multiple aircraft coexisting, there are interaction modes among individuals, including competitive routes, cooperative avoidance, and even non-cooperative game theory. Traditional technologies treat each target as an isolated moving point, lacking the ability to model the collective intelligent interaction mechanism and failing to accurately predict the global situational evolution caused by local decisions.

[0004] Therefore, there is an urgent need to construct an AI-driven low-altitude traffic situational awareness and early warning method that integrates multi-source heterogeneous data, possesses aircraft intent reasoning capabilities, and can simulate multi-agent game interactions. This method should overcome the limitations of existing technologies in terms of insufficient cognitive depth, lack of interactive modeling, and coarse-grained early warning information. It should achieve a paradigm shift from passive response to proactive prediction, and from binary alarms to tiered and actionable recommendations, providing a reliable safety foundation for future high-density low-altitude operations. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based method for low-altitude traffic situational awareness and early warning, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The AI-based method for low-altitude traffic situational awareness and early warning includes the following specific steps: Step S1: Construct a dynamic traffic situation map by integrating multi-source heterogeneous data: Access radar points, automatic dependent surveillance broadcast messages, data reported by mobile communication networks, airborne visual recognition results, airspace geofence information, flight plan texts and meteorological data, and use spatiotemporal graph neural networks to extract features and align the above multi-source data in a spatiotemporal manner to construct a dynamic low-altitude traffic situation map with aircraft as nodes and spatial proximity, communication correlation or mission coordination as edges; Step S2: Generate individual behavior feature vectors: For each aircraft node in the map, based on its historical trajectory sequence, velocity change rate, heading angular rate and acceleration vector, a lightweight temporal neural network encoder is used to generate a behavior feature vector in real time that represents its micro-maneuvering mode. Step S3: Reasoning higher-order mission intent: The behavioral feature vector, the declared flight plan text, the geometric relationship between the current pose and the preset waypoint or mission area, and the summary status of other aircraft in the same airspace are taken as input and fed into an expert model finely tuned by aviation corpus to output the probability distribution of the mission type currently being performed by the aircraft and its higher-order intent. Step S4: Establish a multi-agent game inference model: Model the entire airspace as a partially observable stochastic game system, where each aircraft is an agent, its objective function is defined by the weighted probability distribution of intent output in Step 3, its strategy is parameterized by a deep reinforcement learning network, and multiple rounds of parallel forward inference are performed on edge computing nodes based on the current situation map and the intent of each agent to generate a set of cooperative evolution trajectories within a future preset time window; Step S5: Quantify risks and generate graded early warnings: Analyze the trajectory obtained from the simulation, identify three types of risk events: trajectory violation, minimum interval intrusion, and interactive deadlock. Quantify the risks from three dimensions: time urgency, conflict severity, and simulation uncertainty. Based on the quantification results, generate structured graded early warning information that includes risk causes, level indicators, and avoidance suggestions.

[0007] Preferably, in step 1, the spatiotemporal graph neural network adopts a multi-layer spatiotemporal attention mechanism. Each layer includes a spatial message passing module and a temporal state update module. The spatial message passing module dynamically allocates edge weights according to the Euclidean distance between the spacecraft. The temporal state update module uses a gated recurrent unit to fuse historical node states. The final output node embedding vector has a predetermined dimension.

[0008] Preferably, in step 2, the lightweight temporal neural network encoder adopts a bidirectional long short-term memory network structure, the input window length is a predetermined time period, the sampling frequency is a preset frequency, the hidden layer dimension is a predetermined value, and the output behavior feature vector is used for subsequent intent reasoning after layer normalization processing.

[0009] Preferably, in step 3, the expert model is a sequence-to-sequence model based on the Transformer architecture. Its vocabulary is expanded with aviation operation terminology. The input sequence is composed of structured behavioral features and natural language flight plans. The output is a probability distribution on a predefined intent label set, which includes various task types and high-order intent descriptions.

[0010] Preferably, in step 4, the state space of the partially observable random game system is defined by a dynamic traffic situation map, the action space is a discretized set of maneuver instructions, including basic actions such as flying straight, turning left, turning right, climbing, and descending, and the reward function comprehensively considers task completion, maintaining safe intervals, and energy efficiency, and its weight coefficients are dynamically adjusted by the intention type.

[0011] Preferably, in step 4, the number of parallel forward simulations is not less than a preset number, the duration of each simulation is a predetermined time period, the time step is a preset time step, environmental disturbance noise is introduced during the simulation to simulate wind shear and sensor error, and the standard deviation of the disturbance is set to the preset values ​​of position and velocity.

[0012] Preferably, in step 5, trajectory violation refers to the aircraft's simulated trajectory intruding into a geofence or no-fly zone; minimum interval intrusion refers to the minimum distance between any two aircraft in three-dimensional space being less than a preset safety threshold, which is dynamically set according to the aircraft type and takes a value within a specific range; interactive deadlock refers to two or more aircraft continuously avoiding each other during simulation, resulting in the inability to reach a consensus on passage.

[0013] Preferably, in step 5, the time urgency is defined as the time difference from the current moment to the predicted moment of conflict, the severity is quantified by the product of the predicted minimum interval and the relative speed, and the uncertainty is measured by the standard deviation of the frequency of conflict events in multiple simulations. The three factors are normalized and then weighted to obtain a comprehensive risk index.

[0014] Preferably, the structured hierarchical early warning information includes three levels: L1 alert level warning is triggered when the comprehensive risk index is greater than a first preset threshold and is mainly driven by intentional trends, and the content includes potential risk areas and associated aircraft identifiers; L2 warning level warning is triggered when the comprehensive risk index is greater than a second preset threshold and the conflict probability exceeds a preset probability threshold, and includes the spatiotemporal coordinates of the conflict, the aircraft involved, suggested avoidance actions, and expected effects; L3 emergency level warning is triggered when the comprehensive risk index is greater than a third preset threshold or when extreme abnormal maneuvers such as dives or spirals are detected, immediately activating the emergency protocol and broadcasting the highest priority alarm to all relevant parties.

[0015] Preferably, this method supports separate deployment of cloud training and edge inference. The expert model and the game inference model are continuously iterated and optimized in the cloud using historical running data, while the edge node only loads the inference engine. The time taken for a single complete inference is less than the preset time threshold, which meets the real-time requirements.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. Achieving a paradigm shift from shallow perception to deep cognition: Through the collaboration of a behavioral encoder and a domain expert model, this invention can analyze the semantic intent behind an aircraft's trajectory, accurately distinguishing between task execution and abnormal states, thus solving the fundamental flaw of traditional systems that "see the form but not the meaning." In tests conducted in typical urban logistics scenarios, the accuracy of intent recognition is significantly better than methods that rely solely on trajectory extrapolation.

[0017] 2. Improve prediction accuracy in complex interactive scenarios: A multi-agent game theory framework is introduced, modeling the aircraft as a goal-oriented intelligent decision-making agent to fully simulate its competition, cooperation, and avoidance behaviors in dense airspace. Experiments show that in areas where multiple UAVs converge, the root mean square error of trajectory prediction in this invention is significantly reduced compared to the constant velocity model, and the false alarm rate is significantly decreased.

[0018] 3. Provide an interpretable and operable tiered early warning mechanism: The early warning information is derived from the quantitative analysis of the results of multi-step game theory. It not only includes the risk level, but also clearly points out the causes and specific avoidance suggestions, which greatly improves the response efficiency and decision-making quality of controllers and autopilot systems.

[0019] 4. High compatibility and strong scalability: This method is integrated into the existing low-altitude monitoring platform in the form of software modules without the need to replace the underlying hardware; its data-driven characteristics support continuous optimization as operational data accumulates, and can adaptively add new aircraft types, mission modes and airspace rules, providing a continuously evolving intelligent foundation for future low-altitude traffic management. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical architecture of the AI-based low-altitude traffic situational awareness and early warning method proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the multi-agent game inference model in this invention; Figure 3 This is a logical flowchart of the multi-source heterogeneous data fusion and dynamic traffic situation map construction in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the generation of individual behavior feature vectors of aircraft and the reasoning of higher-order mission intentions in this invention; Figure 5 This is a schematic diagram comparing the core principle of this invention with existing technologies in terms of risk quantification. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0022] Example 1 In the aforementioned AI-based low-altitude traffic situational awareness and early warning method, step 1, which integrates multi-source heterogeneous data to construct a dynamic traffic situation map, is implemented as follows: The system first accesses seven types of heterogeneous data sources through standardized interfaces, including point data output by primary / secondary radars, Automatic Dependent Surveillance-Broadcast (ADS-B) messages, aircraft status messages reported by 4G / 5G mobile communication networks, target detection and semantic segmentation results output by the airborne visual recognition module, airspace geofence vector data provided by Geographic Information System (GIS), structured flight mission text submitted by the flight plan declaration system, and real-time wind speed, wind direction, visibility, and turbulence index data pushed by the meteorological service interface. All data streams are synchronized with a unified timestamp, with a time synchronization accuracy of no less than 10 milliseconds, and cross-device clock alignment is achieved using the IEEE 1588 Precise Time Protocol (PTP).

[0023] The incoming data is then fed into a spatiotemporal graph neural network for feature extraction and fusion. This ST-GNN employs a four-layer stacked architecture, with each layer containing a spatial message passing module and a temporal state update module. In the spatial message passing module, the system first determines the current three-dimensional coordinates of all spacecraft. Calculate any two aircraft and Euclidean distance between .like Less than the preset spatial proximity threshold (Typical value is 2000 meters), then an undirected edge is established in the graph. Edge weight Dynamically calculated using radial basis functions (RBF): ,in This is the distance attenuation factor, taken as 500 meters. Furthermore, if the aircraft... and If they belong to the same operator or share the same communication channel, an additional weight is added for communication relevance; if their flight plans share a common mission area or cooperative waypoints, an additional weight is added for mission coordination. The final edge weight is the weighted sum of the three, with weight coefficients of 0.6, 0.25, and 0.15, respectively.

[0024] In the time status update module, each spacecraft node status Updated at time step t by the gated loop unit (GRU): in, Let i represent the set of neighbors of node i. The learnable message aggregation function is implemented using a multilayer perceptron (MLP), with the input being a concatenated vector of neighbor node states and edge features. After four layers of spatiotemporal message passing, each node outputs a 128-dimensional embedding vector. This vector integrates the aircraft's own historical state, the dynamics of neighboring aircraft, and environmental context information, forming a dynamic low-altitude traffic situation map. ,in For the set of aircraft nodes, Let be the set of edges. Embed the node matrix.

[0025] In the above method, step 2, which generates individual behavior feature vectors, is implemented as follows: For each aircraft node in the dynamic traffic situation map G... The system extracts kinematic sequence data from its historical trajectory buffer within the most recent T seconds, where T is the input window length, set to 30 seconds, and the sampling frequency is... The frequency is 5 Hz, meaning a data set is collected every 200 milliseconds. Each data set contains a six-dimensional vector: [x, y, z, ... , , [ ] represents the three-dimensional position and three-dimensional velocity components, respectively. By performing a difference operation on this sequence, the system calculates the rate of change of velocity in real time. angular rate of heading (Unit: radians / second) and total acceleration vector magnitude .

[0026] The aforementioned seven-dimensional temporal features (position, velocity, acceleration, and yaw rate) are input into a lightweight temporal neural network encoder. This encoder employs a bidirectional long short-term memory (Bi-LSTM) network structure, comprising two sub-networks: a forward LSTM and a backward LSTM, each with a hidden layer dimension of 64, resulting in a combined hidden state dimension of 128. The input sequence length of the Bi-LSTM is 150 (i.e., 30 seconds × 5Hz), and the output is the combined hidden state at the last time step. This state vector is then processed by layer normalization to eliminate feature scale shifts caused by sensor calibration differences between different aircraft. The final output is a behavioral feature vector. It accurately characterizes the micro-maneuvering patterns of aircraft i over the past 30 seconds, including the combined characteristics of basic flight maneuvers such as acceleration, deceleration, turning, hovering, climbing, and descending.

[0027] In the above method, step 3, which infers the higher-order task intent, is implemented as follows: The system will generate the behavioral feature vector in step 2. Flight plan text Geometric relationship characteristics and the general status of other aircraft in the same airspace. As joint input, an expert model finely tuned from an aviation-related corpus is fed in. This expert model is built on the Transformer architecture, employing an encoder-decoder structure with 6 encoder layers, 3 decoder layers, 8 attention heads, and 512 hidden layer dimensions.

[0028] Flight plan text The text is described in natural language, for example, "Departing from logistics center A, flying along corridor C1 to delivery point B, performing a fixed-point hovering delivery, estimated time 12 minutes." This text is first converted into a sequence of sub-words using a Byte Pair Encoding (BPE) tokenizer and aligned with an expanded vocabulary of aviation operational terminology to ensure that specialized terms such as "no-fly zone," "waiting route," and "emergency return" are fully preserved. Behavioral feature vector. The sequence is linearly projected to the same dimension as the word embeddings (512-dimensional) and concatenated with the text embedding sequence in the sequence dimension to form a mixed-modality input sequence. .

[0029] Geometric relationship feature g_i calculates the current position of the aircraft set of preset waypoints Minimum Euclidean distance between The angle between the current heading and the direction of the next waypoint The quantization is based on whether it is located inside the task region polygon (a Boolean value). This triplet... Encoded as a 16-dimensional vector and appended to the end of the input sequence. Summary status of other aircraft in the same airspace. The average velocity, density distribution entropy, and main intent labels (taken from the previous cycle's inference results) are compressed into a 32-dimensional vector by the MLP and then concatenated to the input.

[0030] The output layer of the expert model is connected to a fully connected classification head, and its output is a predefined set of intent labels. The probability distribution on. It includes 12 task types and 8 higher-order intents, totaling 96 composite labels. The model is trained using a cross-entropy loss function, and outputs the top 3 labels with the highest probabilities and their confidence scores as the final intent inference result.

[0031] In the above method, step 4 establishes a multi-agent game inference model, and the specific implementation process is as follows: The system models the entire low-altitude airspace as a partially observable stochastic game (POSG) system. The state space S of this system is fully defined by the dynamic traffic situation map G constructed in step 1, including the position, speed, behavioral characteristics, and intention labels of all aircraft. The action space A is a discretized set of maneuver instructions, including 5 basic actions: maintain, turn left, etc. Changes in heading angle ),Turn right( , The actions include climbing (vertical velocity +2m / s) and descending (-2m / s). Each agent (i.e., the aircraft) selects one action to execute in each decision cycle.

[0032] Each agent The policy is parameterized by a deep reinforcement learning network that employs an attention-based policy network. The input is the agent's own observations. (Including its own state and the states of its k nearest neighbor aircraft,) The probability distribution of intent output from step 3. The network first calculates the relevance weights between itself and neighboring aircraft using a multi-head attention mechanism, then aggregates neighbor information in a weighted manner, and finally outputs the action probability distribution through a fully connected layer. .

[0033] reward function The design is a weighted sum of three terms: .in, A positive reward is given for task completion; if the agent successfully approaches the target point as intended, a positive reward is given. The reward for maintaining a safe distance is defined as the reciprocal of the distance to the nearest neighbor aircraft; the smaller the distance, the greater the penalty. The energy efficiency bonus is inversely proportional to the square of the speed. Weighting coefficient. Dynamically adjusted based on intent type: For example, when the intent is "emergency return", =0.7, =0.25, =0.05; when the intention is "fixed-point inspection", =0.4, =0.5, =0.1.

[0034] On edge computing nodes, the system performs multiple rounds of parallel forward inference based on the current situation map and the intentions of each agent. The number of inferences N is no less than 100, each inference lasts for 120 seconds (T_sim), and the time step Δt is 1 second. At time step t of each inference, all agents sample actions according to their policy network and update their states. To enhance robustness, environmental disturbance noise is introduced during the inference process: position disturbances. velocity disturbance ,in rice, The speed is measured in meters per second to simulate wind shear and sensor measurement errors. After 100 simulations, the system obtains 100 sets of co-evolutionary trajectories for the next 120 seconds. .

[0035] In the above method, step 5 quantifies the risk and generates a tiered early warning system. The specific implementation process is as follows: The system first analyzes the set of projected trajectories and identifies three types of risk events: A. Track Violation: If any aircraft intrudes into a geofence or no-fly zone within any simulated trajectory, it is marked as a track violation event. Geofence data is stored in GeoJSON format and includes polygon vertex coordinates and height limits.

[0036] B. Minimum Interval Intrusion: Calculate the minimum three-dimensional distance between any two aircraft i and j across all simulated trajectories. .like Less than the preset safety threshold If it does, it is marked as a minimum interval intrusion event. Dynamically set according to aircraft type: 50 meters between large eVTOLs, 20 meters between small drones, and 35 meters between eVTOLs and drones.

[0037] C. Interactive deadlock: If two or more aircraft continuously avoid each other for more than 10 consecutive time steps (i.e., 10 seconds) and the relative distance does not change significantly (the rate of change is <0.5 m / s), it is determined to be an interactive deadlock event.

[0038] The system quantifies the identified risk events from three dimensions: Time urgency Defined as starting from the current time. To the predicted time of conflict The time difference, i.e. If the conflict occurs at different times in multiple simulations, a weighted average is taken, with the weight being the probability of the conflict occurring in that simulation.

[0039] Severity : By the minimum prediction interval With relative velocity The product quantization, i.e. This indicator comprehensively reflects the collision kinetic energy and the difficulty of avoidance.

[0040] Uncertainty Frequency of conflict events in 100 simulations The standard deviation is measured as follows: .like ,but =0.1.

[0041] The three factors, after normalization, are weighted and summed to obtain the comprehensive risk index. : in, =120 seconds, For the maximum possible severity (taken as 1000 m) 2 / s), =0.4, =0.4, =0.2.

[0042] in accordance with Value and Conflict Probability The system generates a three-level structured early warning: L1 alert level warning: When >0.3 and Triggered when <0.3. Content includes potential risk areas (expressed in latitude and longitude ranges), associated aircraft identifiers (such as UAV-2024), and a description of the intended trend.

[0043] L2 warning level: When >0.6 and Triggered when >0.5. Content includes conflicting spatiotemporal coordinates (e.g., (The aircraft involved, suggested evasive maneuvers, and expected effects.)

[0044] L3 Emergency Level Warning: When >0.9 or detects a dive (vertical velocity <-5m / s for 3 seconds) or a spin (yaw rate > The system is triggered during extreme and abnormal maneuvers such as rapid altitude descent. It immediately activates preset emergency protocols, including broadcasting the highest priority alarm to all relevant parties (monitoring platform, operations center, and the aircraft involved) and automatically uploading data from the aircraft's black box.

[0045] This method supports separate deployment of cloud training and edge inference. Expert models and game theory models are continuously iterated and optimized in the cloud using historical runtime data, employing a federated learning framework to protect data privacy. Edge nodes only load the TensorRT-optimized inference engine, performing a single complete inference (…). The time taken is less than 800 milliseconds, which meets the real-time requirements.

[0046] To verify the effectiveness of the above method, a specific application example is constructed: In the airspace of a logistics hub in a certain city, four logistics drones (UAV-1 to UAV-4) are simultaneously performing delivery tasks, two eVTOL drones (eVTOL-A and eVTOL-B) are performing passenger commuting, and one inspection drone (UAV-Ins) is performing routine inspections. The system executes a complete process once per second. At t=0, the behavioral feature vector of UAV-3 shows that its heading angular rate fluctuates drastically over 10 consecutive seconds, and the expert model infers that its higher-order intention is "avoiding an obstacle on the right" (confidence level 85%). Game theory simulation shows that in the next 60 seconds, the trajectories of UAV-3 and eVTOL-A will be at coordinates ( The two points intersect nearby, and in 62 out of 100 simulations, the minimum interval was less than 30 meters. Calculations show... =58 seconds, , =0.12, =0.68. The system triggered an L2 warning, recommending that UAV-3 turn right and climb, while eVTOL-A maintained its current heading. This warning was sent to the logistics dispatch center and the eVTOL autopilot system, which, upon confirmation, executed evasive maneuvers, successfully avoiding a potential conflict.

[0047] Example 2 In another implementation scenario, the present invention employs an alternative technical solution to adapt to edge devices with limited computing resources. In step 2, the lightweight temporal neural network encoder is replaced with a one-dimensional convolutional neural network (1D-CNN) structure, containing three convolutional layers with kernel sizes of 7, 5, and 3, and channel numbers of 32, 64, and 128, followed by a global average pooling layer. This structure compresses the input sequence length from 150 to 1, outputs a 128-dimensional behavioral feature vector, improves inference speed by 3 times, and reduces memory usage by 60%.

[0048] In step 3, the expert model employs knowledge distillation to transfer knowledge from a large Transformer model to a 4-layer MLP network. The probability distribution output by the teacher model serves as a soft label to guide the training of the student model. The student model's input is a concatenation of behavioral feature vectors and one-hot encoded flight plan keywords, and its output is the intent label probability. Although the accuracy decreases slightly (approximately 5%), the inference latency is reduced from 50 milliseconds to 5 milliseconds, making it suitable for high-frequency update scenarios.

[0049] In step 4, the multi-agent game simulation uses Monte Carlo Tree Search (MCTS) instead of a deep reinforcement learning policy network. MCTS constructs a local search tree in each simulation, with a depth of 10 steps, expanding the tree to include 5 most promising actions in each step. This method does not require pre-training of the policy network and can adapt to new aircraft types online, but it has a high computational cost. To balance performance, the system only enables MCTS for aircraft in high-risk areas, while using the pre-trained policy network in other areas.

[0050] Through the above alternatives, the present invention achieves flexible adaptation to different hardware platforms while maintaining core functions, ensuring deployment and operation in a wide range of scenarios from high-performance servers to embedded edge devices.

[0051] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based method for low-altitude traffic situational awareness and early warning, characterized in that, Includes the following steps: Step S1: Construct a dynamic traffic situation map by integrating multi-source heterogeneous data: Access radar points, automatic dependent surveillance broadcast messages, data reported by mobile communication networks, airborne visual recognition results, airspace geofence information, flight plan texts and meteorological data, and use spatiotemporal graph neural networks to extract features and align the above multi-source data in a spatiotemporal manner to construct a dynamic low-altitude traffic situation map with aircraft as nodes and spatial proximity, communication correlation or mission coordination as edges; Step S2: Generate individual behavior feature vectors: For each aircraft node in the map, based on its historical trajectory sequence, velocity change rate, heading angular rate and acceleration vector, a lightweight temporal neural network encoder is used to generate a behavior feature vector in real time that represents its micro-maneuvering mode. Step S3: Reasoning higher-order mission intent: The behavioral feature vector, the declared flight plan text, the geometric relationship between the current pose and the preset waypoint or mission area, and the summary status of other aircraft in the same airspace are taken as input and fed into an expert model finely tuned by aviation corpus to output the probability distribution of the mission type currently being performed by the aircraft and its higher-order intent. Step S4: Establish a multi-agent game inference model: Model the entire airspace as a partially observable stochastic game system, where each aircraft is an agent, its objective function is defined by the weighted probability distribution of intent output in Step 3, its strategy is parameterized by a deep reinforcement learning network, and multiple rounds of parallel forward inference are performed on edge computing nodes based on the current situation map and the intent of each agent to generate a set of cooperative evolution trajectories within a future preset time window; Step S5: Quantify risks and generate graded early warnings: Analyze the trajectory obtained from the simulation, identify three types of risk events: trajectory violation, minimum interval intrusion, and interactive deadlock. Quantify the risks from three dimensions: time urgency, conflict severity, and simulation uncertainty. Based on the quantification results, generate structured graded early warning information that includes risk causes, level indicators, and avoidance suggestions.

2. The AI-based low-altitude traffic situational awareness and early warning method according to claim 1, characterized in that, The spatiotemporal graph neural network adopts a multi-layer spatiotemporal attention mechanism. Each layer includes a spatial message passing module and a temporal state update module. The spatial message passing module dynamically allocates edge weights according to the Euclidean distance between the spacecraft. The temporal state update module uses a gated recurrent unit to fuse historical node states. The final output node embedding vector has a predetermined dimension.

3. The AI-based low-altitude traffic situational awareness and early warning method according to claim 1, characterized in that, The lightweight temporal neural network encoder adopts a bidirectional long short-term memory network structure. The input window length is a predetermined time period, the sampling frequency is a preset frequency, the hidden layer dimension is a predetermined value, and the output behavior feature vector is used for subsequent intent reasoning after layer normalization.

4. The AI-based low-altitude traffic situational awareness and early warning method according to claim 1, characterized in that, The expert model is a sequence-to-sequence model based on the Transformer architecture. Its vocabulary is expanded with aviation operation terminology. The input sequence is composed of structured behavioral features and natural language flight plans. The output is a probability distribution on a predefined intent label set, which contains various task types and high-order intent descriptions.

5. The AI-based low-altitude traffic situational awareness and early warning method according to claim 1, characterized in that, The state space of the partially observable stochastic game system is defined by a dynamic traffic situation map, and the action space is a discretized set of maneuver instructions, including basic actions such as flying straight, turning left, turning right, climbing, and descending. The reward function comprehensively considers task completion, maintaining a safe distance, and energy efficiency, and its weight coefficients are dynamically adjusted by the intention type.

6. The AI-based low-altitude traffic situational awareness and early warning method according to claim 1, characterized in that, The number of simulations in the multi-round parallel forward simulation is no less than a preset number. The duration of each simulation is a predetermined time period, and the time step is a preset time step. Environmental disturbance noise is introduced during the simulation to simulate wind shear and sensor error. The standard deviation of the disturbance is set to the preset values ​​of position and velocity.

7. The AI-based low-altitude traffic situational awareness and early warning method according to claim 1, characterized in that, The trajectory violation refers to the aircraft's simulated trajectory intruding into a geofence or no-fly zone. The minimum interval intrusion refers to the minimum distance between any two aircraft in three-dimensional space being less than a preset safety threshold, which is dynamically set according to the aircraft type. The interactive deadlock refers to two or more aircraft continuously avoiding each other during simulation, resulting in the inability to reach a consensus on passage.

8. The AI-based low-altitude traffic situational awareness and early warning method according to claim 1, characterized in that, The time urgency is defined as the time difference from the current moment to the predicted moment of conflict. The severity is quantified by the product of the predicted minimum interval and the relative speed. The uncertainty is measured by the standard deviation of the frequency of conflict events in multiple simulations. The three factors are normalized and then weighted to obtain a comprehensive risk index.

9. The AI-based low-altitude traffic situational awareness and early warning method according to claim 1, characterized in that, The structured hierarchical early warning information includes three levels: L1 alert level early warning is triggered when the comprehensive risk index is greater than the first preset threshold and is mainly driven by intention trend, and the content is potential risk area and associated aircraft identification; The L2 warning level is triggered when the comprehensive risk index is greater than the second preset threshold and the probability of conflict exceeds the preset probability threshold. It includes the spatiotemporal coordinates of the conflict, the aircraft involved, the recommended avoidance maneuvers, and the expected effects. The L3 emergency level is triggered when the comprehensive risk index is greater than the third preset threshold or when extreme abnormal maneuvers such as dives or spirals are detected. It immediately activates the emergency protocol and broadcasts the highest priority alert to all relevant parties.

Images