A global monitoring black flight management system for unmanned aerial vehicles

Abstract

The application discloses a kind of unmanned vehicle global monitoring black flight management and control system, including multi-source perception fusion module, behavior intention prediction engine, multi-dimensional context collaborative decision engine, strategy execution module, online learning and strategy self-optimization engine, airspace collaborative management module and defense posture visualization module.Uniform target track is generated by multi-source perception fusion module;Behavior intention prediction engine predicts motion trajectory based on Transformer and outputs uncertainty measure;Multi-dimensional context collaborative decision engine dynamically adjusts prediction feature weight through feature gating mechanism, and generates active defense strategy;Strategy execution module issues control instruction;Online learning and strategy self-optimization engine continuously optimize strategy based on reinforcement learning;Airspace collaborative management module provides airspace modeling and query;Visualization module shows defense posture.The application realizes complete closed-loop active control from prediction to decision, execution, feedback, significantly improves the robustness and adaptability of black flight defense in complex scenarios.

Description

Technical Field

[0001] This invention belongs to the field of low-altitude safety management technology for unmanned aerial vehicles (UAVs), specifically relating to a comprehensive monitoring and control system for unauthorized UAV flights. Background Technology

[0002] In recent years, the civilian drone industry has experienced explosive growth, with frequent incidents of unauthorized and reckless drone flights and malicious intrusions posing a serious threat to the security of key targets such as airports, nuclear power plants, large event venues, and military facilities. Currently, the technical means to control unauthorized drone flights still have the following significant shortcomings:

[0003] 1. Primarily reactive, lacking proactive control capabilities based on predicting behavioral intentions.

[0004] Most existing mainstream control systems adopt a passive "detect first, judge later, and then handle" approach. The system only begins tracking after the drone enters the monitoring range, and only triggers alarms and countermeasures after it actually intrudes into the no-fly zone or reaches the vicinity of sensitive targets, resulting in a delayed response. Although some solutions (such as CN120995391A) propose threat intent prediction methods based on deep learning, there is a lack of effective linkage between the prediction results and control strategies. After the threat is predicted, manual intervention is still required to decide on countermeasures, which cannot achieve a complete closed loop from prediction to automatic strategy generation, making it difficult to meet the real-time proactive defense requirements in highly dynamic scenarios.

[0005] 2. The control strategy lacks a dynamic coordination mechanism based on multi-dimensional contextual information, and its optimization dimension is singular.

[0006] Existing systems often consider only a single piece of information (such as position or speed) when formulating countermeasures. However, different types of airspace (no-fly zones, key control zones, airway zones, etc.) have drastically different control objectives, and the threat level and appropriate response measures for the same behavior vary greatly in different airspaces. Furthermore, current systems mostly employ a uniform countermeasure logic, lacking the ability to differentiate strategies based on multi-dimensional contextual information such as the target's airspace type, behavioral intent, historical patterns, and real-time prediction confidence.

[0007] In addition, existing optimization schemes (such as CN121585309A and CN122239764A) only optimize single-dimensional continuous variables (such as interference power) or the aircraft's own waypoints, without involving joint optimization of the complete four-dimensional action space of "surveillance-warning-drive-interception". Moreover, the strategy rules are mostly statically configured and lack adaptive evolution capabilities based on actual combat feedback.

[0008] 3. Lack of adaptive adjustment mechanisms for quantifying forecast uncertainty and decision weights.

[0009] The confidence scores output by prediction models are often used directly, but these models may make overconfident misjudgments when faced with out-of-distribution data or insufficient training. Existing systems do not deeply couple prediction confidence scores with the decision-making process, preventing the decision-making system from dynamically adjusting its dependence on prediction features based on its own confidence in the prediction, thus reducing the robustness of decision-making in complex scenarios.

[0010] To address the aforementioned issues, this invention proposes a comprehensive monitoring and control system for unauthorized drone flights. Through the deep integration of behavioral intent prediction and multi-dimensional collaborative decision-making, reinforcement learning self-optimization of a four-dimensional complete action space, and a feature gating mechanism driven by prediction confidence, it achieves closed-loop proactive control from perception to prediction, decision-making, execution, and feedback. Summary of the Invention

[0011] This invention provides a comprehensive monitoring and control system for unauthorized drone flights. It generates a unified target trajectory through a multi-source perception fusion module; a behavior intent prediction engine predicts the trajectory based on Transformer and outputs an uncertainty metric; a multi-dimensional contextual collaborative decision-making engine dynamically adjusts the predicted feature weights through a feature gating mechanism to generate an active defense strategy; a strategy execution module issues control commands; an online learning and strategy self-optimization engine continuously optimizes the strategy based on reinforcement learning; an airspace collaborative management module provides airspace modeling and querying; and a visualization module displays the defense posture, thus addressing the problems mentioned in the background section.

[0012] To achieve the above objectives, the present invention provides the following technical solution: a comprehensive monitoring and control system for unmanned aerial vehicles (UAVs) operating illegally, comprising:

[0013] The multi-source perception fusion module is used to receive UAV monitoring data from multiple sensing devices and perform data fusion processing to generate target trajectory data.

[0014] A behavioral intent prediction engine, connected to the multi-source perception fusion module, is used to predict the future movement trend of the target based on the target trajectory data and output a behavioral intent prediction result containing the predicted trajectory and its uncertainty measure.

[0015] A multi-dimensional contextual collaborative decision engine is connected to the multi-source perception fusion module, the behavior intention prediction engine, and the spatial domain collaborative management module, respectively. It is used to acquire multi-dimensional contextual information, dynamically adjust the decision weight of the predicted trajectory according to the uncertainty metric through a feature gating mechanism, and generate an active defense strategy in combination with the multi-dimensional contextual information.

[0016] The strategy execution module, connected to the multi-dimensional context collaborative decision engine, is used to convert the proactive defense strategy into device control commands and execute them.

[0017] The online learning and strategy self-optimization engine is connected to the multi-dimensional context collaborative decision engine and the strategy execution module, respectively. It is used to calculate the reward value based on the feedback of the strategy execution result and continuously optimize the decision strategy through reinforcement learning algorithm.

[0018] As a further option, the behavioral intent prediction engine includes a trajectory encoding submodule and a spatiotemporal attention prediction submodule built on the Transformer architecture; the spatiotemporal attention prediction submodule uses Monte Carlo Dropout technology to perform multiple random forward propagations during the inference phase and calculates the uncertainty measure based on the statistical distribution characteristics of the multiple output results.

[0019] As a further option, the feature gating mechanism adjusts the feature weights of the predicted trajectory in the input of the decision network by calculating a gating coefficient, which changes continuously from a maximum value to a minimum value within a preset range as the uncertainty metric increases.

[0020] As a further option, the policy space of the multi-dimensional context collaborative decision engine includes four action dimensions: monitoring, warning, expulsion, and interception. Each action dimension has multiple levels to form a hierarchical action space. The multi-dimensional context collaborative decision engine generates the probability distribution of each action dimension through a hierarchical output network and uses a masking mechanism to make the output probability of action combinations that do not meet the preset compatibility constraints zero.

[0021] As a further option, the multi-dimensional context collaborative decision engine includes a context information aggregation submodule, which is used to obtain the airspace context of the target's current airspace, the target context including behavioral intent labels and confidence levels, and the historical context including statistical features of historical black flight events, and encodes various types of context information into a unified multi-dimensional feature vector.

[0022] As a further option, the online learning and policy self-optimization engine updates policy network parameters based on a near-end policy optimization algorithm. The online learning and policy self-optimization engine includes a reward calculation submodule, which calculates a comprehensive reward value based on multiple factors, including whether the defense is successful, whether the prevention is successful, the cost of action execution, false alarms, and missed alarms. The reward calculation submodule also stores experience samples containing state, action, reward, and next state into an experience replay cache pool for policy updates.

[0023] As a further option, the online learning and strategy self-optimization engine also includes a threshold adaptive adjustment submodule, which is used to dynamically adjust the intrusion probability judgment threshold and the strategy adoption confidence threshold based on the defense effect statistics within a preset sliding time window and the F1-score maximization criterion.

[0024] As a further option, the behavioral intent prediction engine adopts an asynchronous prediction mode, and its output prediction results carry a version number corresponding to the current track data; after receiving the prediction results, the multi-dimensional context collaborative decision engine determines whether to adopt the prediction results based on the difference between the version number and the latest track data version number. If it is within a preset tolerance range, it adopts the prediction results; otherwise, it discards them.

[0025] As a further option, the multi-source perception fusion module uses a cross-identifier matching method based on spatiotemporal correlation to de-identify the same target reported by different sensing devices. The cross-identifier matching method associates targets based on weighted similarity of position differences, speed differences, and heading differences. The multi-source perception fusion module also maintains a mapping relationship cache with a time-sensitive management strategy, marking mapping records that have exceeded their validity period as pending verification and performing secondary confirmation when a match is found.

[0026] As a further option, the airspace collaborative management module includes a spatial modeling submodule for performing three-dimensional digital modeling of various types of airspace and maintaining their status, and a spatial query submodule for querying the airspace type and distance from the boundary based on the target coordinates; the system also includes a defense situation visualization module, which is connected to the multi-source perception fusion module, the multi-dimensional context collaborative decision engine and the airspace collaborative management module, respectively, and is used to generate a visualization interface containing target situation information and strategy decision information on a three-dimensional map based on real-time track data, currently effective strategies and airspace boundary data.

[0027] Compared with the prior art, the beneficial effects of the present invention are:

[0028] 1. By using the feature gating mechanism between the behavioral intent prediction engine and the multi-dimensional context collaborative decision engine, the prediction confidence (uncertainty measure) is directly used to adjust the weight of the prediction feature in the policy network, realizing a complete closed loop of "prediction-uncertainty quantification-weight adaptive adjustment", which effectively suppresses the negative impact of prediction misjudgment on decision-making and significantly improves the robustness of decision-making in complex scenarios.

[0029] 2. The decision engine deeply integrates multi-dimensional information such as airspace type, target attributes, historical statistics, and behavioral intent, and generates differentiated proactive policies based on a deep policy network. The policy network adopts a four-dimensional hierarchical action space of "monitoring-warning-deterrence-interception" and utilizes layered output and masking mechanisms to naturally satisfy actual constraints (such as warning downgrading during interception), realizing joint policy generation for the complete control action space.

[0030] 3. The online learning and strategy self-optimization engine uses the PPO algorithm to perform global optimization of the complete four-dimensional action space. It is supplemented by F1-score maximization-driven threshold adaptive adjustment and rule-guided cold start scheme, which constitute a systematic strategy self-optimization framework, enabling the system to continuously evolve and adapt to changing environments based on actual combat feedback.

[0031] 4. From behavioral intent prediction, multi-dimensional contextual collaborative decision-making, strategy execution, to effect evaluation and strategy self-optimization, each stage is automatically linked through a closed-loop feedback mechanism. The asynchronous prediction and version number matching mechanism balances real-time performance and prediction accuracy, ensuring stable operation of the system in high-concurrency scenarios. Attached Figure Description

[0032] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0033] In the attached diagram:

[0034] Figure 1 This is a schematic diagram of the overall system architecture according to an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of the data processing flow of the behavior intent prediction engine in an embodiment of the present invention.

[0036] Figure 3 This is a schematic diagram of the network structure of the Transformer spatiotemporal attention prediction submodule in an embodiment of the present invention.

[0037] Figure 4 This is a schematic diagram of the strategy generation process of the multi-dimensional context collaborative decision engine in an embodiment of the present invention.

[0038] Figure 5 This is a timing diagram of the prediction-decision-execution-feedback closed-loop linkage mechanism in an embodiment of the present invention.

[0039] Figure 6 This is a schematic diagram of the training process of the online learning and strategy self-optimization engine in an embodiment of the present invention.

[0040] Figure 7 This is a schematic diagram of the layout of the control situation visualization interface according to an embodiment of the present invention.

[0041] The diagram shows: 1. Multi-source perception fusion module; 2. Behavioral intent prediction engine; 3. Multi-dimensional contextual collaborative decision-making engine; 4. Policy execution module; 5. Online learning and policy self-optimization engine; 6. Airspace collaborative management module; 7. Defense posture visualization module. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] In the various embodiments of the present invention, unless otherwise specified, the English abbreviations involved have the following meanings: GIS stands for Geographic Information System, API stands for Application Programming Interface, JSON stands for JavaScript Object Notation, RESTful stands for Representational State Transfer Architecture, SN stands for Device Serial Number, Transformer stands for Deep Learning Architecture Based on Self-Attention Mechanism, and PPO stands for Proximal Policy Optimization Algorithm.

[0044] It should be noted that the preset thresholds, preset time windows, network structure parameters, etc., involved in the various embodiments of the present invention can be reasonably set by those skilled in the art through a limited number of experiments or based on experience, according to the accuracy requirements of the actual application scenario, specific airspace management specifications, and defense needs. The specific values ​​or ranges listed below are only used as illustrative examples of the embodiments of the present invention and should not be construed as any improper limitation on the scope of protection of the present invention.

[0045] Example 1: As Figure 1 As shown, this embodiment provides a comprehensive monitoring and control system for unauthorized drone flights, including a multi-source perception fusion module 1, a behavior intent prediction engine 2, a multi-dimensional contextual collaborative decision-making engine 3, a policy execution module 4, an online learning and policy self-optimization engine 5, an airspace collaborative management module 6, and a defense posture visualization module 7. The modules interact with each other via RESTful API interfaces and message queues (such as RabbitMQ) to achieve loosely coupled collaboration.

[0046] The specific implementation methods of each module are described in detail below.

[0047] I. Multi-source sensing fusion module

[0048] The multi-source perception fusion module 1 is used to receive UAV monitoring data from various sensing devices (including but not limited to message detection devices, IoT sensing devices, radar detection devices, 5G-A sensing devices, photoelectric detection devices, etc.), and generate unified target trajectory data after data standardization, spatiotemporal alignment, target association and location fusion.

[0049] To address data ambiguity caused by inconsistent identifier fields when reporting the same target from different sensing devices (e.g., the same drone being recorded with different identifiers such as SN, flight control serial number, and RFID number on different devices), this module employs a cross-identifier matching method based on spatiotemporal correlation. Specifically, when two data packets carry different target identifiers but their time difference is within a preset window and their spatial distance is less than a preset threshold, the similarity of motion features between targets from different data sources is calculated based on the target's position, velocity, heading, and other motion characteristics. This similarity is calculated using the F1 function.

[0050] Where Δp, Δv, and Δh are the normalized differences in position, velocity, and heading, respectively (each component has been normalized to [0,1] according to its respective feature range), ωp, ωv, and ωh are the corresponding weights, and ωp + ωv + ωh = 1. For example, ωp can be set to 0.5–0.7, ωv to 0.2–0.3, and ωh to 0.1–0.2. When F1 is less than the preset similarity threshold τsim (exemplarily 0.65–0.85), it is determined to be the same target, and a cross-identifier mapping relationship is established and stored in the cache.

[0051] The cross-identity mapping relationship cache adopts a time-decay-based validity management strategy. Each mapping record includes a creation timestamp and a most recent hit timestamp, and has a preset validity period (10-30 minutes for example). The system scans the cache through a periodic inspection thread, marking records that have exceeded their validity period as "pending verification" instead of deleting them directly. When a subsequent matching request hits, the F1 score is recalculated for secondary confirmation. If the confirmation is successful, the validity period is reset; otherwise, if it fails or fails for a preset number of consecutive times, it is removed.

[0052] The fused target trajectory data is published in a structured format (including fields such as target unique identifier, timestamp, three-dimensional position coordinates, speed, heading, confidence level, etc.) via a message queue for the behavioral intent prediction engine 2 to subscribe to and consume.

[0053] To ensure real-time performance in high-concurrency scenarios, a parallel processing architecture with a data pipeline is adopted among the modules: the multi-source perception fusion module processes perception data with the highest priority, and the timestamp of its output location data packet serves as the time base for the entire system; the behavior intent prediction engine adopts an asynchronous prediction mode; the multi-dimensional context collaborative decision engine adopts a combination of event triggering and periodic triggering (targets above the preset intent level are triggered by events, and those below or equal to the preset intent level are triggered by periods).

[0054] II. Behavioral Intent Prediction Engine

[0055] The behavioral intent prediction engine 2 is connected to the multi-source perception fusion module 1. Based on the target's historical trajectory sequence, it uses a time-series prediction model to predict the target's movement trajectory and intrusion probability within a preset future time period, outputting the behavioral intent prediction result. For example... Figure 2 As shown, its core processing flow is as follows:

[0056] (I) Trajectory Encoding Submodule

[0057] The trajectory encoding submodule receives target trajectory data, groups it according to the target's unique identifier, and maintains a historical trajectory sequence for each target for the most recent preset number of moments (30-100 for example, corresponding to 30-100 seconds). Each moment contains six-dimensional features: longitude, latitude, altitude, speed, heading, and timestamp.

[0058] Preprocessing steps: Irregular time intervals are interpolated to fixed intervals (e.g., 1 second) using linear interpolation; missing values ​​are filled using a combination of forward padding and linear interpolation; Z-score normalization is applied to each feature dimension. The preprocessed six-dimensional features are mapped to high-dimensional feature vectors (128-512 dimensions for example) through a fully connected embedding layer, which serve as the input to the spatiotemporal attention prediction submodule.

[0059] (II) Spatiotemporal Attention Prediction Submodule

[0060] This submodule is built on the Transformer architecture and is used to capture long-range dependencies in historical trajectory sequences and predict future trajectories. For example... Figure 3 As shown, its network structure is as follows:

[0061] The encoder consists of multiple (exemplary 4 to 8 layers) stacked Transformer coding layers, each containing a multi-head self-attention sublayer (4 to 16 heads) and a feedforward neural network sublayer (containing ReLU activation). Residual connections and layer normalization are applied after each layer.

[0062] The decoder takes the historical semantic representation output by the encoder and the state at the last moment as input, and generates a predicted trajectory sequence for a preset number of future moments (10 to 30 for example) moment by moment through a masked multi-head attention mechanism. The decoder also consists of multiple stacked layers, each containing masked self-attention, encoder-decoder cross-attention, and a feedforward network.

[0063] Training employs a teacher-mandated strategy, with a smoothed L1 loss function. Training data sources include historical flight records, simulated synthetic tracks (covering maneuvering modes such as constant speed, turning, acceleration, deceleration, hovering, climbing, and diving), and publicly labeled datasets. These data are divided into training, validation, and test sets in a 7:1.5:1.5 ratio, and an early stopping method is used to prevent overfitting.

[0064] After prediction is completed, the output includes the predicted location, timestamp, and a self-evaluation metric for prediction confidence. The confidence evaluation employs Monte Carlo Dropout technology: during the inference phase, multiple random forward propagations are performed (10-30 times in the example, dynamically configurable), collecting statistics on the location distribution at each prediction time point, and calculating the standard deviation as the uncertainty measure U. To ensure real-time performance, multiple forward propagations are executed in parallel on the GPU. The system monitors the computation time in real time, and if it exceeds a preset proportion of the decision cycle time (e.g., 50%), the number of forward propagations is automatically reduced.

[0065] The uncertainty metric U (the average of the standard deviations at each time step) is output to the multi-dimensional context collaborative decision engine as supplementary information for prediction confidence.

[0066] (III) Intrusion Probability Calculation Submodule

[0067] This submodule performs spatial calculations on the predicted flight path sequence and the airspace boundary data provided by the airspace collaborative management module 6. For each predicted location, it uses the ray casting method to determine whether it is within each type of airspace. It counts the number and distribution of predicted points falling into each airspace, and calculates the intrusion probability using a logistic regression model, combining the current heading, speed, and altitude characteristics. This probability is calculated by the function F2.

[0068] Where x=[R ioU V head V spd H feat ], R ioU To predict the spatial overlap between the trajectory and the airspace boundary, V head V is the angle between the heading and the direction of the normal to the nearest boundary of the airspace (normalized). spd H is the ratio of the current speed to the historical average speed. feat The relationship between altitude and airspace altitude range is represented by θ1 to θ4, which are learnable weights, and b is the bias term.

[0069] (iv) Intent Classification Submodule

[0070] Based on the intrusion probability and preset thresholds, the target behavior intent is divided into four levels: "no threat", "under observation", "suspected intrusion" and "confirmed intrusion", and the corresponding labels and intrusion probability values ​​are output and pushed to the multi-dimensional context collaborative decision engine 3 through the message queue.

[0071] III. Multi-dimensional Contextual Collaborative Decision Engine

[0072] The multi-dimensional context collaborative decision-making engine 3 is connected to the multi-source perception fusion module 1, the behavior intent prediction engine 2, and the airspace collaborative management module 6, respectively. It is used to acquire airspace context, target context, and historical context information, and collaboratively fuse the behavior intent prediction results with the multi-dimensional context information to generate an active defense strategy. For example... Figure 4 As shown, its core processing flow is as follows:

[0073] (I) Context Information Aggregation Submodule

[0074] In each decision cycle (exemplary 1-5 seconds), three types of information are aggregated from multiple data sources:

[0075] Airspace Context: Call the airspace collaborative management module 6 to obtain the airspace type and control rules of the target's current location, as well as the list of airspace types and entry times that the predicted trajectory will pass through.

[0076] Target context: Obtain behavioral intent labels and confidence levels, historical violation records, and device model / type.

[0077] Historical context: Obtain the statistical characteristics and occurrence rate of unauthorized flights in the current airspace within a preset window of the same historical period.

[0078] The above information is encoded as a fixed-length multidimensional context feature vector (exemplary 256 to 1024 dimensions), where spatial domain type and intent level are encoded using one-hot encoding, and continuous features are concatenated after Min-Max normalization.

[0079] (ii) Feature gating mechanism driven by prediction confidence

[0080] After receiving the prediction result and uncertainty metric U, the decision engine dynamically adjusts the weights of the predicted trajectory features in the policy network input through a feature gating unit. The gating coefficients are calculated using function F3.

[0081] Where T1 and T2 are the low and high uncertainty thresholds, respectively (for example, corresponding to standard deviations less than 5% and greater than 15% of the position mean, respectively), Gmax=1.0 and Gmin=0 are adjustable. The gating coefficient multiplied by the predicted trajectory features and together with the real-time position features are used as input to the policy network to achieve continuous and smooth adjustment of the decision weights by the prediction confidence.

[0082] (III) Confidence Calibration Mechanism

[0083] To mitigate model overconfidence, the system maintains a confidence calibration cache, recording the average deviation rate B (the spatial distance between the predicted and final observed locations divided by the prediction standard deviation) within different confidence intervals (0.1 intervals) in historical predictions. If B in a certain interval exceeds a preset threshold TB (Example 2.0), the confidence level of that interval is discounted and corrected using the F4 function.

[0084] The calibration cache is updated at a preset period (e.g., hourly), retaining only the records from the most recent 24 hours.

[0085] (iv) Strategy Space Construction Submodule

[0086] The policy space comprises four action dimensions: surveillance, warning, expulsion, and interception, with each dimension having three levels (0 / 1 / 2). The dimension levels are defined as follows:

[0087] Compatibility constraints: When interception is set to "forced interception", warnings are automatically downgraded to "no warning"; when removal is set to "signal interference removal", monitoring is at least "enhanced monitoring".

[0088] (V) Policy Network Submodule

[0089] The policy network is constructed using a deep neural network. The input is a context feature vector, and the output consists of four branches (each branch is a 3-dimensional Softmax) that correspond to the probability distribution of each action dimension. The highest probability of a predetermined number of combinations (e.g., 3 to 5) are selected as candidate combinations.

[0090] Example network configuration: Input layer 512-dimensional, first hidden layer 256 neurons + ReLU + batch normalization + Dropout (0.2), second hidden layer 128 neurons + ReLU + batch normalization + Dropout (0.2), output layer four-branch, each branch 3-dimensional Softmax. Optimizer Adam, initial learning rate 3×10⁻⁶. -4 The weight decays by 0.96 every 5000 steps, with a weight decay of 1×10. -5 .

[0091] A hierarchical output architecture is adopted, with four branches sharing the main feature extraction layer and the output layers operating independently. During training, for combinations that do not meet the compatibility constraints, a masking mechanism is introduced into the loss function—the logits of the action combinations that violate the constraints are pre-set to -∞ in the Softmax layer to ensure that the output probability is zero, thus naturally satisfying the constraints.

[0092] During the cold start phase (when the number of samples in the experience replay pool is less than the preset minimum sample size, such as 1,000 to 5,000), a rule-guided exploration strategy is adopted: based on a two-dimensional lookup table matrix (rows = airspace types sorted by control strictness, columns = intent level), defensive actions are preset, while random exploration is carried out with an exploration probability of (0.1 to 0.3), and experience is collected and stored in the cache pool.

[0093] Policy network training hyperparameters: learning rate 1×10 -4 ~5×10 -4 Batch size 64-256, update frequency triggered by the number of collected experience points or a fixed time interval.

[0094] (vi) Strategy Execution Determination Submodule

[0095] Based on the confidence level (Softmax probability) of the candidate strategy, if the highest probability exceeds a preset threshold (example 0.5~0.7), it is directly adopted; otherwise, it enters the manual confirmation process. The adopted strategy undergoes compatibility verification (as a safety net), and finally, an executable instruction (including target identifier, action combination, execution time, and priority) is generated and sent to the strategy execution module 4 via a message queue.

[0096] IV. Strategy Execution Module

[0097] The strategy execution module 4 connects to the multi-dimensional context collaborative decision engine 3, and has a unified control interface for built-in monitoring resources (radar, photoelectric turntable), warning resources (audible and visual alarms, directional broadcasting equipment), drive-away resources (navigation decoy devices, signal jamming devices), and interception resources (counter-drone, laser interception equipment). It parses strategy instructions into specific device control commands, issues and monitors the execution status, and feeds back the execution results to the online learning and strategy self-optimization engine 5.

[0098] V. Online Learning and Strategy Self-Optimization Engine

[0099] The online learning and policy self-optimization engine 5 is connected to the multi-dimensional context collaborative decision-making engine 3 and the policy execution module 4, respectively, to record complete defense event data and continuously optimize policy network parameters based on reinforcement learning. Figure 6 As shown, its core processing flow is as follows:

[0100] (I) Experience Replay Cache Submodule

[0101] After each policy execution, the state characteristics, policy action, reward value, next state, and termination flag are encapsulated into an experience sample and stored in a priority experience replay pool (capacity example 10). 5 ~10 6 High-reward samples have a higher sampling probability.

[0102] (II) Reward Calculation Submodule

[0103] The total reward value is calculated using function F5:

[0104] The suggested values ​​for each sub-item and its weight are as follows:

[0105] R succ (Defense successful, repulsion / interception successful): +10~+20

[0106] R prev (Prevention successful, suspected cases successfully dissuaded): +5 to +10

[0107] R cst (Cost penalty, the higher the intensity of the action, the more negative): "Forced Interception" -2 to -5, "Enhanced Surveillance" -0.5 to -1

[0108] R fp (False alarm penalty, high intent but legal): -10 to -20

[0109] R fn (Penalty for missed reporting, actual intrusion without warning): -5 to -10

[0110] The weights λ1 to λ5 are configured according to scenario preferences. The reward value is normalized before being stored in the cache.

[0111] (III) Policy Gradient Update Submodule

[0112] The policy gradient update submodule is based on the near-end policy optimization (PPO) algorithm. It samples batch experience data from the experience replay cache submodule, calculates the policy gradient, and updates the network parameters of the policy network submodule.

[0113] The PPO objective function (tailored version) used in this module is defined as function F6:

[0114] The meanings of each symbol and the engineering calculation methods are as follows:

[0115] θ represents the network parameters to be updated in the policy network; D represents the experience replay cache pool.

[0116] Let be the empirical expectation operator, representing the calculation of the arithmetic mean of the current mini-batch samples randomly sampled from the empirical replay buffer D. In engineering implementations, this operator is equivalent to . , where N is the preset batch size (exemplary values ​​are 64 to 256), that is, the loss mean of all samples in the current batch is used to approximate the theoretical overall expectation;

[0117] For the new strategy π θ Compared to the old strategy π old In state s t Take action a t The probability ratio is used to measure the magnitude of policy updates;

[0118] A t The advantage function is used to evaluate the current action a. t The degree of good or bad relative to the average level is calculated using the general dominance estimation method;

[0119] The pruning hyperparameter (exemplarily set to 0.1 to 0.2) is used to limit the variation range of the probability ratio ρt, preventing the policy update step from being too large and causing training instability.

[0120] Trim ρt to [1- ,1+ Within the interval, when ρt exceeds the range, its boundary value is taken.

[0121] The objective function described above, by taking the minimum operation, limits excessive policy changes while encouraging the policy to be updated in a better direction, thereby ensuring the stability and convergence of the training process.

[0122] The policy gradient update submodule performs network parameter updates at a preset frequency (for example, an update is triggered once every preset number of experience samples collected, or a batch update is triggered once every preset time interval). The updated parameters are managed by version numbers and asynchronously pushed to the policy network submodule of the multi-dimensional context collaborative decision engine 3.

[0123] Furthermore, this module employs an ε-greedy exploration strategy: actions are randomly selected from the action space with a preset probability ε (for example, an initial value of 0.3–0.5, gradually decreasing to a final value of 0.05–0.1 with a preset decay rate as the number of training steps increases). The optimal action output by the policy network is executed with a probability of 1–ε, balancing exploration and exploitation. The system can also deploy multiple sets (for example, 3–5 sets) of independently running policy network replicas, each trained in parallel based on different random initialization and empirical sampling strategies. The main policy network aggregates the outputs of each replica using soft voting (probability averaging) to improve the stability and generalization ability of the policy.

[0124] (iv) Threshold Adaptive Adjustment Submodule

[0125] Based on the defense effectiveness statistics (true positive rate, false positive rate, etc.) within a sliding time window (e.g., 24 hours), dynamically adjust the intrusion probability threshold and the strategy determination confidence threshold:

[0126] If the false positive rate is >10%, the threshold should be appropriately increased;

[0127] If the false negative rate is >5%, the threshold should be appropriately lowered.

[0128] If both exceed the limit, the F1-score maximization criterion is adopted: a grid search is performed near the threshold with a step size of 0.05, and the threshold corresponding to the maximum F1 score is selected.

[0129] Each adjustment step size does not exceed the preset maximum step size (e.g., 0.05), and a gradual change strategy is adopted.

[0130] VI. Prediction-Decision-Execution-Feedback Closed-Loop Linkage Mechanism

[0131] The behavioral intent prediction engine 2, the multi-dimensional context collaborative decision-making engine 3, the policy execution module 4, and the online learning and policy self-optimization engine 5 are automatically linked through a closed-loop mechanism (e.g., Figure 5 (as shown in the timing diagram)

[0132] Multi-source perception fusion module 1 continuously outputs target trajectory data;

[0133] The behavioral intent prediction engine 2 performs predictions on active targets at preset cycles, and outputs the predicted trajectory, intrusion probability, and uncertainty metric U. The number of Monte Carlo Dropout forward propagation times is dynamically adjusted according to the remaining time of the current decision cycle.

[0134] Each time the multi-dimensional context collaborative decision engine 3 receives a prediction result, it adjusts the weight of the prediction features by calculating the gating coefficient through F3, re-aggregates the context, and calls the policy network to generate the latest policy.

[0135] Strategy execution module 4 receives and executes instructions, and provides real-time feedback on the execution results;

[0136] The online learning and strategy self-optimization engine 5 calculates the F5 reward value based on the execution results, stores the experience, and triggers policy network updates (F6) at a preset frequency.

[0137] The updated policy network parameters are pushed to the decision engine asynchronously.

[0138] The behavioral intent prediction engine uses an asynchronous prediction mode, and the prediction results are associated with the track data through version numbers. After receiving the prediction results, the decision engine checks the difference between the version number and the version number of the latest track data. If it is within a preset tolerance range (e.g., within 5 consecutive version numbers), it adopts the prediction; otherwise, it discards the prediction.

[0139] VII. Airspace Collaborative Management Module

[0140] The airspace collaborative management module 6 is used to perform three-dimensional digital modeling and status maintenance of various airspaces within the protected area, and to provide query services to the decision engine.

[0141] (I) Airspace Digital Modeling Submodule: Based on a 3D GIS engine, this module provides graphical drawing tools and supports various airspace types, including no-fly zones, enhanced warning zones, low-altitude airways, take-off and landing point protection zones, and key control zones. It allows setting lower / upper airspace altitude limits, stretching 2D graphics into 3D solids, rendering with semi-transparent materials, and storing data in GeoJSON format with a spatial index.

[0142] (II) Dynamic Airspace Status Management Submodule: Supports temporary activation and deactivation of airspace. Operators can set the effective time period, and the system will automatically activate or remove it from the query range. When airspace boundaries change, update events are pushed to the decision engine via message queue.

[0143] (III) Spatial Relationship Query Submodule: Provides a standardized spatial relationship query interface. Input the three-dimensional coordinates of the target and return a list of airspace types (sorted by priority: no-fly zone > key control zone > enhanced warning zone > airway zone > take-off and landing protection zone > suitable airspace), the distance from each airspace boundary, and the predicted entry time (if outside).

[0144] VIII. Defense Situation Visualization Module

[0145] like Figure 7 As shown, the defense posture visualization module 7 displays defense posture information in real time on a 3D GIS map, including four core views:

[0146] Target Situation View: Displays the real-time location of all targets using icons, with added historical trajectories (solid lines), predicted trajectories (dashed lines, with transparency decreasing over time), behavioral intent labels, and an intrusion probability ring indicator.

[0147] Strategy Decision View: The sidebar displays the currently effective strategy action combination, confidence level, and decision basis (expanded by a decision tree to show the contribution weight of each context feature).

[0148] Airspace Resource View: Various airspaces are displayed as semi-transparent 3D objects of different colors, with floating 3D text labels above.

[0149] Historical Events View: The timeline at the bottom displays a list of recent defense events.

[0150] 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 comprehensive monitoring and control system for unmanned aerial vehicles (UAVs) operating illegally, characterized in that: include: The multi-source perception fusion module (1) is used to receive UAV monitoring data from multiple sensing devices and perform data fusion processing to generate target trajectory data; The behavioral intent prediction engine (2) is connected to the multi-source perception fusion module (1) and is used to predict the future movement trend of the target based on the target trajectory data and output the behavioral intent prediction result containing the predicted trajectory and its uncertainty measure. The multi-dimensional context collaborative decision engine (3) is connected to the multi-source perception fusion module (1), the behavior intention prediction engine (2) and the spatial domain collaborative management module (6) respectively. It is used to obtain multi-dimensional context information, dynamically adjust the decision weight of the predicted trajectory according to the uncertainty measure through the feature gating mechanism, and generate an active defense strategy in combination with the multi-dimensional context information. The strategy execution module (4) is connected to the multi-dimensional context collaborative decision engine (3) and is used to convert the active defense strategy into device control instructions and execute them. The online learning and strategy self-optimization engine (5) is connected to the multi-dimensional context collaborative decision engine (3) and the strategy execution module (4) respectively, and is used to calculate the reward value based on the feedback of the strategy execution result, and continuously optimize the decision strategy through reinforcement learning algorithm.

2. The UAV full-domain monitoring and control system for illegal flights according to claim 1, characterized in that, The behavioral intent prediction engine (2) includes a trajectory encoding submodule and a spatiotemporal attention prediction submodule built on the Transformer architecture; the spatiotemporal attention prediction submodule uses Monte Carlo Dropout technology to perform multiple random forward propagations during the inference phase and calculates the uncertainty measure based on the statistical distribution characteristics of the multiple output results.

3. The UAV full-domain monitoring and control system for illegal flights according to claim 1, characterized in that, The feature gating mechanism adjusts the feature weights of the predicted trajectory in the input of the decision network by calculating a gating coefficient. The gating coefficient changes continuously from the maximum value to the minimum value within a preset range as the uncertainty metric increases.

4. The UAV full-domain monitoring and control system for illegal flights according to claim 1, characterized in that, The strategy space of the multi-dimensional context collaborative decision engine (3) includes four action dimensions: monitoring, warning, expulsion, and interception. Each action dimension has multiple levels to form a hierarchical action space. The multi-dimensional context collaborative decision engine (3) generates the probability distribution of each action dimension through a hierarchical output network and uses a masking mechanism to make the output probability of action combinations that do not meet the preset compatibility constraints zero.

5. The UAV full-domain monitoring and control system for illegal flights according to claim 1, characterized in that, The multi-dimensional context collaborative decision engine (3) includes a context information aggregation submodule, which is used to obtain the airspace context of the target's current airspace, the target context containing behavioral intent labels and confidence levels, and the historical context containing statistical features of historical black flight events, and encodes various types of context information into a unified multi-dimensional feature vector.

6. The UAV full-domain monitoring and control system for illegal flights according to claim 1, characterized in that, The online learning and strategy self-optimization engine (5) updates the strategy network parameters based on the near-end strategy optimization algorithm; the online learning and strategy self-optimization engine (5) includes a reward calculation submodule, which calculates a comprehensive reward value based on multiple factors including whether the defense is successful, whether the prevention is successful, the cost of action execution, false alarms and missed alarms, and stores the experience sample including state, action, reward and next state into the experience replay cache pool for strategy update.

7. The UAV full-domain monitoring and control system for illegal flights according to claim 6, characterized in that, The online learning and strategy self-optimization engine (5) also includes a threshold adaptive adjustment submodule, which is used to dynamically adjust the intrusion probability judgment threshold and the strategy adoption confidence threshold according to the defense effect statistics within the preset sliding time window and the F1-score maximization criterion.

8. The UAV full-domain monitoring and control system for illegal flights according to claim 1, characterized in that, The behavioral intent prediction engine (2) adopts an asynchronous prediction mode, and its output prediction results carry a version number corresponding to the current trajectory data; After receiving the prediction result, the multi-dimensional context collaborative decision engine (3) determines whether to adopt the prediction result based on the difference between the version number and the latest track data version number. If it is within the preset tolerance range, it adopts the prediction result; otherwise, it discards it.

9. The UAV full-domain monitoring and control system for illegal flights according to claim 1, characterized in that, The multi-source sensing fusion module (1) uses a cross-identifier matching method based on spatiotemporal correlation to de-identify the same target reported by different sensing devices. The cross-identifier matching method associates targets based on the weighted similarity of position difference, speed difference and heading difference. The multi-source perception fusion module (1) also maintains a mapping relationship cache with a timeliness management strategy, marks mapping records that have exceeded their validity period as pending verification and performs secondary confirmation when a match is found.

10. The UAV full-domain monitoring and control system for illegal flights according to claim 1, characterized in that, The airspace collaborative management module (6) includes a spatial modeling submodule for performing three-dimensional digital modeling of various types of airspace and maintaining their status, and a spatial query submodule for querying the airspace type and distance from the boundary based on the target coordinates; the system also includes a defense situation visualization module (7), which is connected to the multi-source perception fusion module (1), the multi-dimensional context collaborative decision engine (3) and the airspace collaborative management module (6) respectively, and is used to generate a visualization interface containing target situation information and strategy decision information on a three-dimensional map based on real-time track data, currently effective strategies and airspace boundary data.

Images