An enemy aircraft intention recognition method and system based on multi-source data fusion and LSTM+reinforcement learning

Abstract

The application provides an enemy aircraft intention recognition method and system based on multi-source data fusion and LSTM+reinforcement learning, which comprises a multi-source sensor network, a data processing center and a command decision system. The method is multi-source data acquisition and fusion, LSTM time sequence feature extraction, preliminary intention decision by reinforcement learning, dynamic real-time reasoning, intention output and confidence evaluation. Effects: The information integrity is improved through multi-source data fusion, the behavior understanding ability is enhanced through LSTM time sequence modeling, autonomous optimization is realized through reinforcement learning, and real-time performance is ensured through a dynamic reasoning mechanism. The overall technical scheme can realize high-accuracy, low-latency and strong-adaptive enemy aircraft intention recognition.

Description

Technical Field

[0001] This invention belongs to the fields of air defense and artificial intelligence, and relates to enemy aircraft intent recognition technology. Background Technology

[0002] In the field of border air defense, enemy aircraft intent identification is a crucial step in operational decision-making. Existing technologies primarily employ the following solutions:

[0003] Option 1: Intent recognition based on rule and template matching

[0004] This method identifies enemy aircraft intentions using a predefined rule base and intention templates defined by military experts. The specific process is as follows: First, target motion characteristics (such as altitude, speed, heading, acceleration, etc.) are extracted from radar data; then, the extracted characteristics are matched against predefined intention templates; finally, the type of enemy aircraft intention is determined based on the matching degree. For example, if an enemy aircraft approaches a friendly target at high speed, it is determined to have an "attack" intention; if an enemy aircraft hovers at low altitude, it is determined to have a "reconnaissance" intention.

[0005] Option 2: Probabilistic Inference Based on Bayesian Networks This method constructs a Bayesian network model for intent recognition, inferring enemy aircraft intentions through conditional probability relationships. Network nodes include nodes representing target motion characteristics, electromagnetic characteristics, environmental characteristics, and intent types. Observation nodes receive sensor data, while intent nodes are calculated through probabilistic inference.

[0006] Option 3: Machine Learning Based on Support Vector Machine (SVM) This method models intent recognition as a classification problem, using historical training data to train an SVM classifier. The input is the target feature vector, and the output is the intent category label.

[0007] The existing technical solutions have the following technical problems:

[0008] 1. Limited data source and low information utilization. Existing solutions primarily rely on single radar data, failing to fully utilize information from multiple sensors such as electro-optical, infrared, and radio sensors. Different sensors are complementary: radar can accurately measure distance and velocity, electro-optical sensors provide target imaging information, infrared sensors can detect target thermal radiation, and radio sensors can identify target communication characteristics. A single data source limits information completeness, resulting in limited accuracy in intent recognition.

[0009] 2. Lack of time series modeling capabilities Rule-based and template-matching methods primarily rely on static feature matching, failing to capture the temporal dependencies in enemy aircraft behavior. Enemy aircraft intentions are typically manifested through a series of sequential actions, such as approaching at low altitude, then circling, and finally accelerating to penetrate defenses. Existing methods struggle to extract meaningful patterns from these temporal behaviors.

[0010] 3. Insufficient real-time performance Bayesian network probabilistic inference and traditional machine learning methods are computationally intensive when dealing with complex features, and the processing latency is usually more than 5 seconds, which makes it difficult to meet the real-time decision-making requirements of modern air defense operations (which usually require millisecond-level response).

[0011] 4. Weak adaptability Existing methods use static models, which cannot autonomously adapt to changes in the battlefield environment and the evolution of enemy aircraft tactics. Once enemy aircraft adopt new tactics or deception methods, the system's recognition accuracy will drop significantly, requiring manual retraining or adjustment of the model. Summary of the Invention

[0012] To address the above technical problems, this invention proposes an enemy aircraft intent recognition method and system based on multi-source data fusion and LSTM+reinforcement learning. By utilizing multi-source sensor data and employing deep learning and reinforcement learning techniques, it achieves high accuracy, low latency, and strong adaptability in enemy aircraft intent recognition.

[0013] The technical solution of this invention: A method for enemy aircraft intent recognition based on multi-source data fusion and LSTM+reinforcement learning, comprising the following steps: S1, Multi-source data acquisition and spatiotemporal calibration Periodically collect observation data from four types of sensors: radar, photoelectric, infrared, and radio; use BeiDou / GPS timing to unify the timestamps of the data from each sensor; convert the coordinates of each sensor to the WGS84 geographic coordinate system to achieve spatial alignment; use a joint probabilistic data association algorithm to associate different sensor data of the same target with the same target ID, and output a spatiotemporally aligned multi-source target data stream;

[0014] S2, Multimodal Feature Layer Fusion The multi-source target data stream output in step S1 is layered, and features are extracted and fused: the target motion features extracted by radar and photoelectric are fused by weighted average; the target morphology features extracted by photoelectric and infrared are fused by image feature matching; signal features are extracted from radio data; all heterogeneous features are normalized to the [0,1] interval to generate a multi-dimensional fused target feature vector.

[0015] S3. Temporal Feature Extraction Based on LSTM + Self-Attention A 30-second sliding time window is constructed based on the current time, with a window step size of 100ms and a total of 300 time steps. The continuous temporal multidimensional fusion target feature vector is input into an LSTM network. The LSTM network uses a 3-layer LSTM layer with 128 hidden units per layer and an attention layer with 8 attention heads to perform forward computation to update the temporal information. The key time step features of the temporal information are weighted using a self-attention mechanism. The output is a 64-dimensional temporal feature vector.

[0016] S4. Initial Intent Decision Based on DQN Reinforcement Learning Enemy aircraft intent recognition is modeled as a Markov Decision Process (MDP): A 64-dimensional temporal feature vector forms the state space S, and six intent types—attack, penetration, reconnaissance, feint, jamming, and retreat—form the action space A. A composite reward function Rt = α·accuracy is constructed. t -β·delay t Where accuracy is weighted by α=10, latency by β=0.1, and accuracy... t To identify correct and incorrect indicators, delay t To address latency, a Deep Q-Network (DQN) agent was constructed. The network input layer receives 64-dimensional temporal features, and the three fully connected hidden layers have 256, 128, and 64 neurons with ReLU activation function. The output layer has 6 neurons that output Q-values ​​for each intent category. Training phase collection (s t , a t , r t , s t+1 The quadruplets are stored in the experience pool, and 32 experiences are randomly sampled to form a mini-batch. Every 1000 steps, the current network parameters are copied to the target network. An ε-greedy exploration strategy is adopted, with an initial ε=1.0, a decay coefficient of 0.995, and a lower limit of 0.1. During online inference, the temporal features are input into the trained DQN agent, and the intent corresponding to the maximum Q value is selected as the preliminary recognition result.

[0017] S5. Decision-making level fusion and confidence calculation based on DS evidence theory The preliminary identification results and rule-based identification results are fused together to define an identification framework and construct two sets of basic probability allocation functions. The two sets of evidence are fused using Dempster synthesis rules. The final confidence score is calculated. The final intent identification result after fusion and the confidence score in the range of 0 to 1 are output.

[0018] S6, Pipeline Parallel Real-Time Inference Acceleration The end-to-end inference process is broken down into three parallel pipeline stages: data preprocessing, feature extraction, and intent recognition. The neural network model is quantized from FP32 to INT8 format. CUDA parallel computing is used. Computational resources are dynamically allocated according to the target threat priority. The end-to-end total inference latency is ≤150ms, and the recognition results are pushed to the command and decision-making end.

[0019] Preferably, in step S2, the radar weight is 0.6 and the photoelectric weight is 0.4.

[0020] Furthermore, it also includes adaptive optimization steps for strong electromagnetic interference scenarios:

[0021] In step S1, adaptive filtering and Kalman filtering are added to suppress radar noise, and the data quality of each sensor is evaluated in real time. In step S2, the fusion weights of various features are dynamically adjusted based on the sensor data quality scores. When the radar data quality is poor, the weights of photoelectric, infrared, and radio data are automatically increased, and abnormal data is detected and removed. In step S4, during the training phase, electromagnetic interference scene samples are expanded to the experience pool, and the reward weights corresponding to correct identification in electromagnetic interference scenes are increased simultaneously.

[0022] Preferably, the LSTM network training parameters are set as follows: learning rate 0.001, total number of training rounds 1000; and the DQN reinforcement learning parameters are set as follows: experience pool size 10000, discount factor γ=0.99, network learning rate 0.0001.

[0023] An enemy aircraft intent recognition system based on the aforementioned enemy aircraft intent recognition method includes: a multi-source sensor network, a data processing center, and a command and decision-making system; The multi-source sensor network includes a remote phased array radar, an optoelectronic tracking system, an infrared thermal imaging system, and a radio spectrum monitoring device, which are used to simultaneously collect raw data on the motion trajectory, optical, thermal radiation, and radio signals of airspace targets, and transmit the data to the data processing center via optical fiber. The data processing center is deployed with a multi-core server cluster and has five built-in functional modules for executing enemy aircraft intent recognition methods: (1) Spatiotemporal calibration and data association module: Configure BeiDou / GPS timing unit, coordinate transformation unit, and JPDA association unit to perform multi-source data time synchronization, spatial alignment, and target data association; (2) Adaptive feature fusion module: realizes weighted fusion of motion, morphology and signal multimodal features, feature normalization, and dynamically adjusts fusion weights in interference scenarios; (3) LSTM temporal feature extraction module: equipped with a 3-layer 128-unit LSTM network and an 8-head self-attention layer, extracting 64-dimensional high-level temporal features by extracting a 30-second temporal window; (4) DQN reinforcement learning decision module: It has built-in MDP modeling unit, DQN agent network and experience playback storage unit, and outputs the initial intention recognition result of enemy aircraft; (5) Real-time reasoning and decision fusion module: integrates pipeline parallel scheduling unit, model quantization unit, GPU acceleration unit, and DS evidence fusion calculation unit to complete multi-source decision fusion and confidence calculation, and achieve ≤150ms millisecond-level reasoning output; The command and decision-making system is connected to the data processing center via a local area network, providing a human-computer interaction interface to receive and display the target airspace situation, enemy aircraft identification intentions and corresponding confidence levels, assisting commanders in making combat decisions.

[0024] The time synchronization accuracy of the spatiotemporal calibration and data association module is better than 1ms, and the transmission delay from the sensor to the data processing center is less than 10ms.

[0025] The data processing center also has a built-in anti-interference extension submodule, which includes an adaptive noise filtering unit and a sensor quality assessment unit. These are used to dynamically adjust the feature fusion weights when the radar is subjected to electromagnetic interference, and to expand the training samples of electromagnetic interference scenarios to optimize the DQN agent.

[0026] The real-time inference and decision fusion module executes the following acceleration strategies: three-stage pipeline parallel processing, FP32 to INT8 model quantization, CUDA parallel computing to accelerate neural network inference, and dynamic computing power scheduling based on target threat priority.

[0027] Beneficial effects of this invention: Effect 1: Significantly improved recognition accuracy Principle: Multi-source data fusion makes full use of the complementary information from different sensors, LSTM time-series modeling captures the long-term dependence of behavior, and reinforcement learning achieves policy optimization; Data support: Compared to the 72% accuracy of traditional methods, this invention achieves over 94% accuracy; Analysis: Data fusion improves feature completeness, temporal modeling enhances behavioral understanding, and reinforcement learning optimizes decision-making strategies. The synergistic effect of these three factors significantly improves accuracy.

[0028] Effect 2: Significantly enhanced real-time performance Principle: Pipeline parallel processing increases throughput, model quantization reduces computation, and GPU acceleration utilizes parallel computing capabilities. Data support: End-to-end inference latency has been reduced from more than 5 seconds in traditional methods to less than 150ms; Analysis: The multi-layered optimization measures together ensure millisecond-level response, meeting the real-time requirements of modern air defense operations.

[0029] Effect 3: Significantly improved adaptability Principle: Reinforcement learning continuously optimizes its strategies through interaction with the environment, adapting to new scenarios without human intervention; Data support: In the new tactical scenarios, the system's recognition accuracy decreased by less than 5%, while the traditional method decreased by more than 20%. Analysis: Reinforcement learning's self-learning ability enables the system to continuously optimize and adapt to changes in the battlefield environment.

[0030] Effect 4: Enhanced system reliability Principle: The decision-making level adopts the DS evidence theory, which can handle uncertainty and conflicting information; Data support: In complex electromagnetic environments, the false alarm rate is only 0.1-0.3%; Analysis: Multi-source information fusion reduces the impact of single sensor failure or interference, and improves the overall reliability of the system. Attached Figure Description

[0031] Figure 1 This is a block diagram of the enemy aircraft intent recognition system of the present invention; Figure 2 This is a flowchart of the enemy aircraft intent recognition method of the present invention. Detailed Implementation

[0032] This invention applies to border air defense systems, and the enemy aircraft intent recognition system architecture includes: 1. Multi-source sensor network: including long-range phased array radar, photoelectric tracking system, infrared thermal imaging system, and radio spectrum monitoring equipment; 2. Data Processing Center: Deployed with a multi-core server cluster, equipped with GPU accelerator cards; 3. Command and Decision System: Connected to the data processing center, providing a human-computer interaction interface.

[0033] Functions of each device: Long-range phased array radar: detects target range, speed, azimuth, and altitude, and tracks target trajectory; Photoelectric tracking system: acquires optical images of the target and identifies the target type and external features; Infrared thermal imaging system: detects target thermal radiation and identifies engine status and activity characteristics; Radio spectrum monitoring equipment: monitors target communication signals, identifies IFF (Identification Friend or Foe) responses, and detects electronic interference; Data Processing Center: Executes data fusion, feature extraction, and intent recognition algorithms; Command and Decision System: Displays situational information, receives intent recognition results, and assists commanders in making decisions.

[0034] Communication connectivity: Each sensor is connected to the data processing center via a fiber optic network, with a data transmission delay of less than 10ms; the data processing center is connected to the command and decision-making system via a local area network.

[0035] The data processing center has six built-in functional modules: (1) Spatiotemporal calibration and data association module: Configure BeiDou / GPS timing unit, coordinate transformation unit, and JPDA association unit to perform multi-source data time synchronization, spatial alignment, and target data association; (2) Adaptive feature fusion module: realizes weighted fusion of motion, morphology and signal multimodal features, feature normalization, and dynamically adjusts fusion weights in interference scenarios; (3) LSTM temporal feature extraction module: equipped with a 3-layer 128-unit LSTM network and an 8-head self-attention layer, extracting 64-dimensional high-level temporal features by extracting a 30-second temporal window; (4) DQN reinforcement learning decision module: It has built-in MDP modeling unit, DQN agent network and experience playback storage unit, and outputs the initial intention recognition result of enemy aircraft; (5) Real-time reasoning and decision fusion module: integrates pipeline parallel scheduling unit, model quantization unit, GPU acceleration unit, and DS evidence fusion calculation unit to complete multi-source decision fusion and confidence calculation, and achieve ≤150ms millisecond-level reasoning output; (6) Anti-interference extension submodule: includes an adaptive noise filtering unit and a sensor quality assessment unit, which are used to dynamically adjust the feature fusion weights when the radar is subjected to electromagnetic interference, and to expand the training samples of electromagnetic interference scenarios to optimize the DQN agent.

[0036] Method Example 1 An enemy aircraft intent recognition method based on multi-source data fusion and LSTM+reinforcement learning. Scenario description: An enemy aircraft intent recognition system of this invention has been deployed in a border area, with a monitoring range of 200 kilometers. The system monitors the activities of multiple enemy aircraft in real time and needs to dynamically identify their intent.

[0037] Technical process: Step S1: Multi-source data acquisition and spatiotemporal calibration Triggering condition: Any sensor detects a target and periodically triggers data acquisition at a frequency of 10Hz.

[0038] Execution entity: Data acquisition module of the data processing center.

[0039] Processing actions: 1. Receive radar data: including target position (longitude, latitude, altitude), velocity (vx, vy, vz), acceleration, heading angle, and azimuth angle; 2. Receive photoelectric data: including target optical images and target type identification results (such as fighter jets, reconnaissance aircraft, and transport aircraft); 3. Receive infrared data: including target thermal imaging images, engine temperature, and activity status; 4. Receiving radio data: including communication frequency, modulation method, signal strength, and IFF (Identification Friend or Foe) response status;

[0040] Spatiotemporal calibration processing: 1. Timestamp synchronization: The GPS / BeiDou timing system is used to unify the timestamps of data from all sensors, with a synchronization accuracy better than 1ms; 2. Spatial coordinate transformation: Transform the data from various sensors to a unified WGS84 geographic coordinate system, and achieve spatial alignment through a coordinate transformation matrix; 3. Data Association: Based on target location information, the Joint Probabilistic Data Association (JPDA) algorithm is used to associate different sensor data of the same target with the same target ID.

[0041] Processing result: Generates a spatiotemporally aligned multi-source target data stream, with each target containing a multi-dimensional feature vector.

[0042] Use of the results: as input for subsequent feature extraction and intent recognition.

[0043] Function of this step: To solve the problem of spatiotemporal inconsistency of multi-source heterogeneous data and lay the foundation for feature layer fusion.

[0044] Step S2: Multimodal feature layer fusion Triggering condition: Triggered after the spatiotemporal alignment is completed in step S1.

[0045] Execution entity: Adaptive feature fusion module of the data processing center.

[0046] Processing actions: 1. Motion feature fusion: The target motion features extracted from radar and electro-optical data are fused, including position, velocity, acceleration, heading angle, rate of climb, etc., and a weighted average method is used, with radar data weighted at 0.6 and electro-optical data weighted at 0.4. 2. Morphological feature fusion: Fusion of target morphological features extracted from photoelectric and infrared data, including target size, shape, number of engines, external mounts, etc., based on image feature matching and fusion; 3. Signal Feature Fusion: Fusion of signal features extracted from radio data, including frequency, modulation scheme, pulse characteristics, etc. 4. Feature normalization: Normalize features of different dimensions and map them to the [0,1] interval.

[0047] Processing result: A multi-dimensional fused target feature vector is generated, with a dimension of 128.

[0048] Use of the results: as input features for LSTM networks.

[0049] The purpose of this step is to improve the integrity and robustness of features and enhance the accuracy of intent recognition.

[0050] Step S3: Temporal feature extraction based on LSTM + self-attention Triggering condition: Triggered after step S2 outputs the multidimensional fused target feature vector.

[0051] Execution Module: LSTM temporal feature extraction module in the data processing center. LSTM Network Structure: Input Layer: 128-dimensional feature vector input; LSTM Layer: 3 LSTM layers, 128 hidden units per layer, learning rate 0.001; Attention Layer: Self-attention mechanism, 8 attention heads; Output Layer: Fully connected layer, outputting a 64-dimensional temporal feature vector.

[0052] Processing actions: 1. Construct a time window: Using the current moment as a baseline, backtrack through the historical data of the past 30 seconds to form a time series, with a time step of 100ms and a total of 300 time steps; 2. Forward computation: Forget Gate Calculation: Controls the information discarded from the state; symbol meaning: --Forget gate output, value [0,1]; σ--Sigmoid activation function; --Forget gate weight matrix; --Vector concatenation, combining the hidden states from the previous time step and current input Combine them into a single long vector; --Forget gate bias term; Input gate calculation: , controls the input of new information, symbol meaning: --Input gate control coefficient, range (0,1); σ--Sigmoid activation function; --Input gate weight matrix; --Input gate bias term; Candidate state calculation: , Symbol meaning: -- Candidate cell state; tanh -- Hyperbolic tangent activation function, output range [−1,1]; --The weight matrix of the candidate states; --Bias terms corresponding to candidate states; Cell status update: , This indicates the cell state at the previous moment; This indicates the long-term cell state at the current moment; Output gate calculation: Symbol meaning: --Output gate, value range 0~1; σ--Sigmoid activation function; --Output gate weight matrix; --Output gate bias; Hidden output: , --The short-term hidden state at the current moment.

[0053] 3. Attention mechanism: Self-attention is applied to the time series features output by LSTM, automatically weighting the key time step features.

[0054] Processing result: Outputs a 64-dimensional temporal feature vector containing long-term dependency information on enemy aircraft behavior.

[0055] Use of the results: as state input for reinforcement learning agents.

[0056] Function of this step: To capture the temporal patterns and long-term dependencies of enemy aircraft behavior, thereby improving the depth of intent understanding.

[0057] Step S4: Preliminary Intent Decision Based on DQN Reinforcement Learning Triggering condition: Triggered after the timing characteristics are output in step S3.

[0058] Execution entity: Reinforcement learning decision module of the data processing center.

[0059] Processing actions: 1. Environmental Modeling (Markov Decision Process) State space S: 64-dimensional temporal feature vectors; Action Space A: 6 types of intent (attack, penetration, reconnaissance, feint, interference, retreat); Transition probability P: The environmental state transition is determined by the actual enemy aircraft behavior, which is difficult to model, so a model-free method is adopted; Reward function R design: Rt = α·accuracy t -β·delay t Among them: accuracy t To identify correct and incorrect indicators, accuracy t =1, intent recognition is correct; otherwise, 0; delay t To handle latency (seconds), the accuracy weight α = 10 and the latency weight β = 0.1.

[0060] 2. Network Intelligent Agent Structure Input layer: 64-dimensional state vector; Hidden layers: 3 fully connected layers with 256, 128, and 64 neurons respectively, and ReLU activation function; Output layer: 6 neurons, outputting the Q-value of each action, with a linear activation function; Network parameters: Optimized through empirical replay and gradient descent.

[0061] 3. Training Process Collecting experience: The agent interacts with the environment and collects (s) t , a t , r t , s t+1 The quadruple is stored in the experience pool; t : Current state at time t, a t In state s t The action to be executed, r t Instant rewards obtained after performing an action, s t+1 : The next state after performing an action; Experience pool size: 10000; Experience replay: 32 experiences are randomly sampled from the experience pool to form a mini-batch, and the number of training rounds is 1000; Gradient update: Calculate the loss function L -- loss function, mean squared error loss; Q(s,a;θ): Q value of the current network (main network) in state s and action a, θ is the main network parameter; r -- instant reward; γ -- discount factor, value 0.99; --The target network takes the maximum Q value in the next state s′ and the next action a′, θ − The target network parameters are frozen; the network parameters are updated using gradient descent. Target network update: Copy the current network parameters to the target network every 1000 steps; Exploration strategy: An ε-greedy strategy is adopted, with an initial exploration rate ε=1.0 and an exploration rate ε decay rate of 0.995, which gradually decays to 0.1 during training.

[0062] 4. Online reasoning The temporal features output in step S3 are used as state inputs to the trained network agent. The network agent outputs the Q-values ​​of each intent category, and the intent category with the largest Q-value is selected as the recognition result. Processing result: Output preliminary identification result (one of 6 intent categories).

[0063] Use of results: As output, it serves as a tool for subsequent confidence calculations.

[0064] Step function: Through reinforcement learning, autonomous optimization is achieved, improving adaptability and recognition accuracy.

[0065] Step S5: Decision-making level fusion and confidence calculation Triggering condition: Triggered after the preliminary identification results are output in step S4.

[0066] Execution entity: Real-time reasoning and decision fusion module of the data processing center.

[0067] Processing actions: 1. Multi-source decision fusion: The preliminary identification results are fused with the rule-based identification results.

[0068] 2. Application of DS Evidence Theory: Define the identification framework: Θ = {Attack, Penetration, Reconnaissance, Feint, Interference, Retreat}; Constructing basic probability assignments: recognition confidence based on LSTM + reinforcement learning and rule-based matching degree; Evidence synthesis: Multi-source evidence is fused using Dempster's synthesis rules.

[0069] 3. Confidence level calculation: ,in Let be the probability of the i-th type of intent after fusion; There are j classes of probabilities for the merged intent.

[0070] Processing results: Output the final intent recognition result and confidence score (between 0 and 1).

[0071] Use of results: To be transmitted to the command and decision-making system.

[0072] Function of the steps: Further improve the reliability of identification through decision fusion, and assist commanders in making judgments through confidence assessment.

[0073] Step S6: Pipeline Parallel Dynamic Real-Time Inference Triggering condition: Continuously triggered after receiving new multi-source data.

[0074] Execution entity: Real-time reasoning and decision fusion module of the data processing center.

[0075] Processing actions: 1. Pipeline parallel processing: The inference process is divided into three pipeline stages (data preprocessing, feature extraction, and intent recognition), with different target data being processed in parallel at each stage; 2. Model Quantization: Quantize the trained neural network model from FP32 to INT8, reducing model size and computation by 75%; 3. GPU acceleration: Utilizes CUDA parallel computing to accelerate neural network inference; 4. Dynamic scheduling: Dynamically allocate computing resources based on the target threat priority, and process high-priority targets first.

[0076] Processing result: End-to-end inference latency within 150ms was achieved.

[0077] Application of results: The identification results are pushed to the command and decision-making end to meet the real-time requirements.

[0078] Purpose of this step: To ensure the availability of the system in real-time scenarios.

[0079] Method Example 2 (Extended Example)

[0080] Scenario Description: In complex electromagnetic environments, the enemy employs electronic jamming tactics, resulting in a decline in radar data quality, necessitating enhanced anti-jamming capabilities of the system.

[0081] Differences in technical processes: Compared with Example 1, this example has been improved in the following aspects: Step S1 Improvement: Anti-interference Data Preprocessing Add an anti-interference filtering module: Adaptive filtering and Kalman filtering are used to suppress noise and interference in radar data; Add signal quality assessment: assess the data quality of each sensor.

[0082] Step S2 Improvement: Adaptive Feature Fusion Adaptive weighted fusion is adopted: the weight coefficients are dynamically adjusted according to the quality of sensor data, and the weights of photoelectric, infrared and radio data are automatically increased when the radar data quality is poor. Add anomaly detection: detect and remove abnormal data to prevent interfering data from affecting the fusion results.

[0083] Step S4 Improvement: Robust Reinforcement Learning Reward function adjustment: Increase the reward weight for accurate identification under electromagnetic interference environment; Enhanced Experience Replay: Electromagnetic interference scenario samples are added to the training data experience pool to improve the system's anti-interference capability.

[0084] Beneficial effects: In environments with strong electromagnetic interference, the identification accuracy increased from 94% in Example 1 to 97%. The false alarm rate decreased from 0.3% to 0.1%. The system robustness was significantly enhanced.

[0085] Summary of key technical points: 1. An intent recognition algorithm architecture based on the combination of LSTM and reinforcement learning Technical approach: This invention combines LSTM temporal feature extraction with reinforcement learning decision-making to form an end-to-end intent recognition framework. LSTM is responsible for extracting high-level features from multi-source temporal data, while reinforcement learning is responsible for intent classification and policy optimization based on these features. The two are connected through feature vectors to form a complete recognition process.

[0086] Beneficial Effects: Compared to traditional rule-based or statistical learning methods, this method can automatically learn complex intent patterns without the need for manual rule design. LSTM can capture long-term dependencies, and reinforcement learning can continuously optimize the strategy; the combination of the two significantly improves recognition accuracy. Experiments show that this architecture achieves an accuracy of over 94% on enemy aircraft intent recognition tasks, a 22 percentage point improvement over traditional methods.

[0087] 2. Multi-level fusion strategy for multimodal data fusion Technical approach: This invention proposes a three-layer fusion strategy consisting of a data layer, a feature layer, and a decision layer. The data layer performs spatiotemporal calibration and correlation, the feature layer performs weighted fusion of multi-source features, and the decision layer performs result fusion based on DS evidence theory. The three layers work together to form a complete fusion system.

[0088] Beneficial effects: Compared to single-source methods, multi-level fusion fully utilizes the complementary information from different sensors, improving feature integrity. When sensors fail or are interfered with, the three-layer fusion can compensate for each other, enhancing system robustness. Experiments show that multi-source fusion improves recognition accuracy by 15-20 percentage points compared to single-source methods.

[0089] 3. Dynamic real-time reasoning mechanism Technical means: This invention employs multiple technical means such as pipelined parallel processing, model quantization, hardware acceleration, and dynamic scheduling to construct a dynamic real-time inference mechanism. Pipeline parallelism improves throughput, model quantization reduces computational load, GPU acceleration utilizes parallel computing, and dynamic scheduling optimizes resource allocation.

[0090] Beneficial effects: Compared to traditional serial processing methods, real-time inference mechanisms reduce processing latency from seconds to milliseconds. In multi-target scenarios, pipelined parallel processing ensures that the recognition latency for each target does not exceed 150ms. Dynamic scheduling mechanisms ensure that high-priority targets are processed first, improving the system's task completion rate.

[0091] 4. Self-optimization mechanism based on reinforcement learning Technical means: This invention models intent recognition as a Markov decision process, constructs an agent, environment, and reward function, and uses the DQN algorithm for policy optimization. The agent learns continuously through interaction with the environment and can adapt to new scenarios without human intervention.

[0092] Beneficial effects: Compared to traditional static models, reinforcement learning mechanisms enable the system to autonomously adapt to changes in the battlefield environment and the evolution of enemy aircraft tactics. When new threats emerge, the system can learn and adapt quickly, with a significantly smaller drop in recognition accuracy compared to traditional methods. It reduces the cost of manual maintenance and model updates, improving the system's maintainability.

Claims

1. A method for enemy aircraft intent recognition based on multi-source data fusion and LSTM+reinforcement learning, characterized in that, Includes the following steps: S1, Multi-source data acquisition and spatiotemporal calibration It periodically collects observation data from four types of sensors: radar, photoelectric, infrared, and radio. The BeiDou / GPS timing system is used to unify the timestamps of data from all sensors; the coordinates of all sensors are uniformly converted to the WGS84 geographic coordinate system to achieve spatial alignment; a joint probabilistic data association algorithm is used to associate different sensor data of the same target with the same target ID, and output a spatiotemporally aligned multi-source target data stream. S2, Multimodal Feature Layer Fusion The multi-source target data stream output in step S1 is layered, and features are extracted and fused: the target motion features extracted by radar and photoelectric are fused by weighted average; the target morphology features extracted by photoelectric and infrared are fused by image feature matching; signal features are extracted from radio data; all heterogeneous features are normalized to the [0,1] interval to generate a multi-dimensional fused target feature vector. S3. Temporal Feature Extraction Based on LSTM + Self-Attention A 30-second sliding time window is constructed based on the current time, with a window step size of 100ms and a total of 300 time steps. The continuous temporal multidimensional fusion target feature vector is input into an LSTM network. The LSTM network uses a 3-layer LSTM layer with 128 hidden units per layer and an attention layer with 8 attention heads to perform forward computation to update the temporal information. The key time step features of the temporal information are weighted using a self-attention mechanism. The output is a 64-dimensional temporal feature vector. S4. Initial Intent Decision Based on DQN Reinforcement Learning Enemy aircraft intent recognition is modeled as a Markov Decision Process (MDP): A 64-dimensional temporal feature vector forms the state space S, and six intent types—attack, penetration, reconnaissance, feint, jamming, and retreat—form the action space A. A composite reward function Rt = α·accuracy is constructed. t -β·delay t Where accuracy is weighted by α=10, latency by β=0.1, and accuracy... t To identify correct and incorrect indicators, delay t To address latency, a Deep Q-Network (DQN) agent was constructed. The network input layer receives 64-dimensional temporal features, and the three fully connected hidden layers have 256, 128, and 64 neurons with ReLU activation function. The output layer has 6 neurons that output Q-values ​​for each intent category. Training phase collection (s t , a t , r t , s t+1 The quadruplets are stored in the experience pool, and 32 experiences are randomly sampled to form a mini-batch. Every 1000 steps, the current network parameters are copied to the target network. An ε-greedy exploration strategy is adopted, with an initial ε=1.0, a decay coefficient of 0.995, and a lower limit of 0.

1. During online inference, the temporal features are input into the trained DQN agent, and the intent corresponding to the maximum Q value is selected as the preliminary recognition result. S5. Decision-making level fusion and confidence calculation based on DS evidence theory The preliminary identification results and rule-based identification results are fused together to define an identification framework and construct two sets of basic probability allocation functions. The two sets of evidence are fused using Dempster synthesis rules. The final confidence score is calculated. The final intent identification result after fusion and the confidence score in the range of 0 to 1 are output. S6, Pipeline Parallel Real-Time Inference Acceleration The end-to-end inference process is broken down into three parallel pipeline stages: data preprocessing, feature extraction, and intent recognition. The neural network model is quantized from FP32 to INT8 format. CUDA parallel computing is used. Computational resources are dynamically allocated according to the target threat priority. The end-to-end total inference latency is ≤150ms, and the recognition results are pushed to the command and decision-making end.

2. The enemy aircraft intent recognition method based on multi-source data fusion and LSTM+reinforcement learning according to claim 1, characterized in that, In step S2, the radar weight is 0.6 and the photoelectric weight is 0.

4.

3. The enemy aircraft intent recognition method based on multi-source data fusion and LSTM+reinforcement learning according to claim 1, characterized in that, It also includes adaptive optimization steps for strong electromagnetic interference scenarios: In step S1, adaptive filtering and Kalman filtering are added to suppress radar noise, and the data quality of each sensor is evaluated in real time. In step S2, the fusion weights of various features are dynamically adjusted based on the sensor data quality scores. When the radar data quality is poor, the weights of photoelectric, infrared, and radio data are automatically increased, and abnormal data is detected and removed. In step S4, during the training phase, electromagnetic interference scene samples are expanded to the experience pool, and the reward weights corresponding to correct identification in electromagnetic interference scenes are increased simultaneously.

4. The enemy aircraft intent recognition method based on multi-source data fusion and LSTM+reinforcement learning according to claim 1, characterized in that, The LSTM network training parameters were set as follows: learning rate 0.001, total number of training rounds 1000; the DQN reinforcement learning parameters were set as follows: experience pool size 10000, discount factor γ=0.99, network learning rate 0.0001.

5. An enemy aircraft intent recognition system based on the enemy aircraft intent recognition method according to any one of claims 1 to 4, characterized in that, include: Multi-source sensor networks, data processing centers, and command and decision-making systems; The multi-source sensor network includes a remote phased array radar, an optoelectronic tracking system, an infrared thermal imaging system, and a radio spectrum monitoring device, which are used to simultaneously collect raw data on the motion trajectory, optical, thermal radiation, and radio signals of airspace targets, and transmit the data to the data processing center via optical fiber. The data processing center is deployed with a multi-core server cluster and has five built-in functional modules for executing the enemy aircraft intent recognition method according to any one of claims 1 to 4: (1) Spatiotemporal calibration and data association module: Configure BeiDou / GPS timing unit, coordinate transformation unit, and JPDA association unit to perform multi-source data time synchronization, spatial alignment, and target data association; (2) Adaptive feature fusion module: realizes weighted fusion of motion, morphology and signal multimodal features, feature normalization, and dynamically adjusts fusion weights in interference scenarios; (3) LSTM temporal feature extraction module: equipped with a 3-layer 128-unit LSTM network and an 8-head self-attention layer, extracting 64-dimensional high-level temporal features by extracting a 30-second temporal window; (4) DQN reinforcement learning decision module: It has built-in MDP modeling unit, DQN agent network and experience playback storage unit, and outputs the initial intention recognition result of enemy aircraft; (5) Real-time reasoning and decision fusion module: integrates pipeline parallel scheduling unit, model quantization unit, GPU acceleration unit, and DS evidence fusion calculation unit to complete multi-source decision fusion and confidence calculation, and achieve ≤150ms millisecond-level reasoning output; The command and decision-making system is connected to the data processing center via a local area network, providing a human-computer interaction interface to receive and display the target airspace situation, enemy aircraft identification intentions and corresponding confidence levels, assisting commanders in making combat decisions.

6. The enemy aircraft intent recognition system according to claim 5, characterized in that, The time synchronization accuracy of the spatiotemporal calibration and data association module is better than 1ms, and the transmission delay from the sensor to the data processing center is less than 10ms.

7. The enemy aircraft intent recognition system according to claim 5, characterized in that, The data processing center also has a built-in anti-interference extension submodule, which includes an adaptive noise filtering unit and a sensor quality assessment unit. These are used to dynamically adjust the feature fusion weights when the radar is subjected to electromagnetic interference, and to expand the training samples of electromagnetic interference scenarios to optimize the DQN agent.

8. The enemy aircraft intent recognition system according to claim 5, characterized in that, The real-time inference and decision fusion module executes the following acceleration strategies: three-stage pipeline parallel processing, FP32 to INT8 model quantization, CUDA parallel computing to accelerate neural network inference, and dynamic computing power scheduling based on target threat priority.

Images