Unmanned aerial vehicle multimode detection and collaborative countering system based on mooring platform

Abstract

The invention belongs to the technical field of unmanned aerial vehicle defense, and particularly discloses an unmanned aerial vehicle multimode detection and collaborative countering system based on a mooring platform. The system is deployed on a mooring platform and comprises a multi-modal fusion detection unit, a threat evaluation and intention recognition unit, a collaborative countering strategy generation unit and a dynamic resource scheduling and execution unit. Through multi-sensor data fusion, intelligent threat evaluation, collaborative strategy generation based on reinforcement learning and accurate resource scheduling, integrated intelligent processing from detection and identification to collaborative countering of an unmanned aerial vehicle target is realized, and countering efficiency and system reliability are improved.

Description

Technical Field

[0001] This invention belongs to the field of drone defense technology, specifically relating to a multi-mode detection and collaborative countermeasure system for drones based on a tethered platform. Background Technology

[0002] As an important branch of modern aviation, unmanned aerial vehicle (UAV) technology has been widely applied in various fields such as military reconnaissance, civilian aerial photography, and logistics transportation. With the popularization of UAV technology and the expansion of its application scenarios, the detection and countermeasures against unauthorized or potentially threatening UAVs within specific airspaces have become a technological requirement in the field of airspace security management. Building an efficient and reliable UAV detection and countermeasures system is a crucial link in ensuring the safety of low-altitude airspace.

[0003] In existing technologies, tethered drone detection systems typically rely on a single or limited range of sensor modalities for target identification. For example, visible light or infrared cameras are used for visual detection, or radar is used for target search. However, in long-range detection scenarios, drone targets appear as point-like targets with very few pixels and extremely weak features in image sensors. Traditional detection algorithms based on image feature extraction and classification struggle to distinguish targets from background noise, leading to a higher false negative rate.

[0004] Single-modal detection methods exhibit significant shortcomings in perception capability and reliability when facing complex weather conditions, strong background interference, or drones with low detectability designs. Existing countermeasure systems often lack deep information fusion and collaborative decision-making mechanisms with detection units, making it difficult to dynamically generate optimal countermeasure strategies based on real-time, multi-dimensional detection data. This results in delayed response times, insufficient effectiveness, or wasted resources. Therefore, achieving high-precision, high-reliability detection of long-range, low-feature drone targets, and realizing intelligent coordination between detection information and countermeasure actions, is a pressing technical problem to be solved in the field of drone defense. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles (UAVs) based on a tethered platform, in order to solve the problems of insufficient accuracy and reliability in the detection of long-range, low-feature UAV targets in the existing technology, as well as the low countermeasure effectiveness caused by the lack of deep collaboration between the detection and countermeasure links.

[0006] This invention provides a tethered platform-based multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles (UAVs), comprising:

[0007] The multimodal fusion detection unit is used to perceive the target airspace in multiple dimensions. The multimodal fusion detection unit integrates a high-resolution photoelectric imaging module, a long-wave infrared thermal imaging module, a phased array radar module, and a passive radio frequency detection module.

[0008] The threat assessment and intent recognition unit is used to receive and process multi-dimensional target situation vectors from the multimodal fusion detection unit to perform dynamic threat level determination and flight intent prediction.

[0009] The collaborative countermeasure strategy generation unit is used to generate a collaborative countermeasure instruction sequence based on the output of the threat assessment and intent recognition unit.

[0010] The dynamic resource scheduling and execution unit is used to receive and parse the strategy matrix issued by the collaborative countermeasure strategy generation unit, and drive specific countermeasure equipment to perform countermeasure actions.

[0011] Preferably, the high-resolution photoelectric imaging module and the long-wave infrared thermal imaging module achieve physical axis consistency through a common optical path design, and share a servo turntable with three-axis stabilization and high-precision pointing function;

[0012] The phased array radar module and the passive radio frequency detection module's antenna arrays are integrated on the same carrier in a spatial multiplexing manner. This same carrier is mechanically coupled to the servo turntable through a rigid connection. The multimodal fusion detection unit internally runs a deep learning-based cross-modal feature-level fusion algorithm. The processing flow of this cross-modal feature-level fusion algorithm is as follows:

[0013] Each detection module preprocesses the raw sensing data into feature maps under a unified spatiotemporal reference. Furthermore, the photoelectric image feature maps and infrared thermal image feature maps are extracted in parallel via a dual-branch convolutional neural network and then weighted and fused through a cross-channel attention mechanism to generate enhanced visual fusion features.

[0014] Radar point cloud features and radio frequency signal spectral features are encoded into high-dimensional vectors;

[0015] Visual fusion features and radar radio frequency encoded vectors are input into a multimodal Transformer encoder for deep correlation and information complementarity, and the output is a multidimensional target situation vector containing the target presence probability, spatial location, motion state and preliminary attribute labels.

[0016] Preferably, the threat assessment and intent recognition unit has a built-in prior knowledge base containing the characteristics and behavioral patterns of various typical UAV models;

[0017] The threat assessment process is performed based on a multidimensional weighted scoring model. The input parameters of this multidimensional weighted scoring model include the target relative distance, approach speed, flight altitude, aircraft type matching degree, signal feature anomaly degree, and whether it is in a no-fly zone.

[0018] Each parameter is linearly weighted and summed according to a preset threat contribution weight, and a risk premium based on the target's historical trajectory is added to predict whether it will intrude into the protected area in the next 2 minutes. Finally, a quantitative comprehensive threat index is calculated. The intent recognition process is implemented based on a long short-term memory network model, which takes the target's motion state sequence of 10 consecutive moments as input and outputs the classification probability of the target's possible intent.

[0019] Preferably, the collaborative countermeasure strategy generation unit includes a policy solver based on multi-agent reinforcement learning;

[0020] The strategy solver models multiple targets to be countered in the current airspace and various countermeasure resources of the system as a partial Markov decision process.

[0021] Countermeasure resources include a directional radio frequency jamming module, a navigation signal deception module, and a net capture and transmission module; the state space of the strategy solver is composed of the threat index, intent classification, spatial distribution of all targets, and the state and position of each countermeasure resource.

[0022] Action space is defined as the period in which each countermeasure resource is assigned a countermeasure action against the target or remains in standby mode in the next decision cycle.

[0023] The design of the reward function takes into account the success rate of counter-attack, resource consumption costs, collateral impact on friendly equipment, and the improvement of overall airspace security.

[0024] The cooperative countermeasure strategy generation unit combines offline training with online fine-tuning, enabling the strategy solver to solve the cooperative countermeasure strategy matrix that maximizes long-term cumulative rewards in real time.

[0025] Preferably, the dynamic resource scheduling and execution unit includes a real-time resource scheduler and multiple countermeasures equipment controllers;

[0026] The real-time resource scheduler calculates the scheduling timeline and spatial path planning based on the strategy matrix and the physical location, pointing capability, range of action, and readiness status of each countermeasure equipment.

[0027] The directional radio frequency jamming module controller, according to instructions, controls its high-gain directional antenna beam to aim at the target UAV and injects jamming signals;

[0028] The navigation signal deception module controller generates a deception signal that is spatiotemporally synchronized with the real satellite navigation signal;

[0029] The net-capturing launch module controller calculates launch parameters and controls the launching device to launch a kinetic projectile carrying a capture net.

[0030] Preferably, the multimodal Transformer encoder in the cross-modal feature-level fusion algorithm has the following specific workflow:

[0031] The visual fusion feature map is converted into a visual feature sequence through spatial location encoding and flattening operations; at the same time, the radar radio frequency encoded vector is used as an independent modal feature vector.

[0032] The visual feature sequence and the radar radio frequency feature vector are concatenated into a multimodal input sequence, and a modality type embedding vector is added; this multimodal input sequence is then fed into a network consisting of 12 stacked Transformer coding layers for processing.

[0033] Within each coding layer, the association weights between all elements within the sequence are calculated using a self-attention mechanism;

[0034] The feature vector corresponding to the target region in the encoder output sequence is taken, and then classified and regressed through a fully connected layer to output the final multidimensional target situation vector.

[0035] Preferably, the threat contribution weight of the multidimensional weighted scoring model in the threat assessment and intent recognition unit is dynamically optimized through an offline-trained adaptive weight allocation network;

[0036] The adaptive weight allocation network uses historical interception case data as a training set, inputs the parameter values ​​when the case occurred, and outputs a qualitative threat level label for the historical interception case.

[0037] The network continuously adjusts the weight coefficients of each input parameter through the backpropagation algorithm to maximize the correlation coefficient between the model's score output and the real threat level.

[0038] The system triggers a fine-tuning update of the weight allocation network every 100 new complete disposal cases accumulated.

[0039] Preferably, the multi-agent reinforcement learning policy solver in the collaborative countermeasure strategy generation unit is trained in a high-fidelity digital twin simulation system.

[0040] The simulation system accurately simulates the dynamic characteristics of the tethered platform, the noise model of each detection sensor, the propagation attenuation under different meteorological conditions, and the flight dynamics and anti-interference model of various UAVs.

[0041] Each counter-resource in the policy solver is modeled as an independent agent, which shares a centralized evaluation network but has its own independent behavioral policy network.

[0042] The training employs a near-end policy optimization algorithm and conducts millions of rounds of adversarial drills in a simulated environment.

[0043] The trained policy network parameters are deployed to the policy solver of the actual system, and transfer learning and rapid fine-tuning are performed based on a small amount of online data collected during actual operation.

[0044] Preferably, the real-time resource scheduler in the dynamic resource scheduling and execution unit uses a mixed integer programming model based on time window constraints for scheduling calculation;

[0045] The mixed-integer programming model takes minimizing the overall countermeasure mission completion time as the primary optimization objective, while using the physical motion constraints, the range of action constraints, and the temporal dependencies between tasks of each countermeasure equipment as hard constraints.

[0046] For the net capture and launch module, its scheduling calculation also needs to take into account the ballistic dynamics equations to ensure that the launch parameters are within the feasible region; the scheduler runs a fast solution algorithm every 0.5 seconds, and updates the fine-grained scheduling instruction sequence for the next 5 seconds based on the latest target situation and resource status, and sends it to the corresponding countermeasure equipment controller for execution.

[0047] Preferably, the system also includes a system health management and reconfiguration unit;

[0048] The system health management and reconstruction unit continuously monitors the health status of each sensor module in the multimodal fusion detection unit, the computing node in the collaborative countermeasure strategy generation unit, and each countermeasure equipment in the dynamic resource scheduling and execution unit.

[0049] When any sensor is detected to have malfunctioned or degraded in performance, the system's health management and reconstruction unit immediately activates the modal reconstruction algorithm, adjusts the weight allocation of the feature-level fusion algorithm in the multimodal fusion detection unit, reduces the contribution of the faulty mode, and enhances the confidence of the remaining normal modes.

[0050] When a countermeasure fails, the multimodal fusion detection unit sends a resource status update to the cooperative countermeasure strategy generation unit, triggering a re-solution of the strategy.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] 1. This invention constructs a multimodal fusion detection unit integrating photoelectric, infrared, radar, and radio frequency sensors, and employs a deep learning-based cross-modal feature-level fusion algorithm to achieve multi-dimensional information complementarity and enhanced perception of long-range, low-feature UAV targets. The design overcomes the problem of insufficient perception capability of a single sensor in complex environments. By using a Transformer encoder to achieve deep feature association, it improves the target detection probability and recognition accuracy in noisy, low-contrast scenarios, and reduces the overall false alarm rate and missed detection rate of the system.

[0053] 2. This invention establishes a complete intelligent chain from perception to cognition to decision-making by setting up a threat assessment and intent recognition unit and a collaborative countermeasure strategy generation unit. The system can not only quantitatively assess the immediate threat of a target, but also predict its behavioral intent and generate a globally optimal collaborative countermeasure strategy based on a multi-agent reinforcement learning model. This design transforms countermeasure actions from passive, isolated responses to proactive, collaborative, and precise interventions, improving the utilization efficiency of countermeasure resources and the overall success rate of countermeasures, while avoiding resource conflicts and waste.

[0054] 3. This invention ensures high reliability and robustness of the system under complex dynamic environments and partial equipment failures by coordinating dynamic resource scheduling and execution units with system health management and reconfiguration units. Real-time scheduling based on optimization algorithms guarantees precise spatiotemporal execution of countermeasure commands, forming closed-loop control; while the system's self-sensing and self-reconfiguration capabilities enable the system to dynamically adjust its operating mode in the face of internal losses or external interference, maintaining continuous and effective functionality, and improving the system's practical survivability and long-term task adaptability. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0056] Figure 2 This is a schematic diagram of the core principle framework of the deep learning-based cross-modal feature-level fusion algorithm in this invention;

[0057] Figure 3 This is a logical flowchart of the threat assessment and intent recognition unit in this invention;

[0058] Figure 4 This is a schematic diagram of the principle framework of the cooperative countermeasure strategy generation unit based on multi-agent reinforcement learning in this invention.

[0059] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the dynamic resource scheduling and execution unit in this invention. Detailed Implementation

[0060] Example 1: The overall technical architecture of the tethered platform-based UAV multi-mode detection and cooperative countermeasure system described in this invention is shown in the attached figure. Figure 1As shown, the system is deployed on a tethered aerial platform with continuous power supply and high-speed data link capabilities. This tethered aerial platform is connected to the ground station via a high-strength tether cable, enabling bidirectional transmission of power, communication, and control signals, thus ensuring the system's ability to operate at high altitudes for extended periods. The entire system comprises five functional modules: a multimodal fusion detection unit, a threat assessment and intent recognition unit, a collaborative countermeasure strategy generation unit, a dynamic resource scheduling and execution unit, and a system health management and reconfiguration unit. These modules are tightly coupled through a high-speed internal bus and a unified spatiotemporal reference, forming a closed-loop intelligent system of perception, cognition, decision-making, execution, and feedback.

[0061] The multimodal fusion detection unit's structural design follows the principles of physical axis consistency and spatial field-of-view overlap. This multimodal fusion detection unit integrates a high-resolution photoelectric imaging module, a long-wave infrared thermal imaging module, a phased array radar module, and a passive radio frequency detection module.

[0062] The high-resolution optoelectronic imaging module and the long-wave infrared thermal imaging module adopt a common optical path design, sharing the same three-axis stabilized servo turntable. This three-axis stabilized servo turntable has omnidirectional coverage capability with an azimuth angle of ±180 degrees and a pitch angle of -90 degrees to +60 degrees, a pointing accuracy better than 0.05 degrees, and a stabilization bandwidth of no less than 20 Hz, ensuring accurate target tracking even under strong wind disturbances.

[0063] The antenna arrays of the phased array radar module and the passive radio frequency detection module are integrated on the same rigid carrier in a spatially multiplexed manner. This rigid carrier is rigidly connected to the aforementioned optoelectronic servo turntable via a mechanical flange, ensuring that the instantaneous main beam direction of the radar and radio frequency antennas is highly aligned with the optoelectronic / infrared line of sight in space, with a typical deviation controlled within 0.1 degrees. This integrated structure eliminates registration deviations caused by installation errors between multiple sensors, providing a physical basis for subsequent feature-level fusion.

[0064] At the data processing level, the multimodal fusion detection unit internally runs a deep learning-based cross-modal feature-level fusion algorithm, the process of which is shown in the attached figure. Figure 2 As shown. First, each detection module preprocesses the raw sensing data into feature maps under a unified spatiotemporal reference. The high-resolution photoelectric imaging module outputs a visible light image with a resolution of 4096×2160, which, after denoising, distortion correction, and color space conversion, is input into the first branch of the dual-branch convolutional neural network; the long-wave infrared thermal imaging module outputs a thermal image with a resolution of 1280×1024, which, after non-uniformity correction and temperature calibration, is input into the second branch.

[0065] Both branches use a ResNet-50 backbone network for feature extraction, each outputting a feature map with 256 channels. Then, a cross-channel attention mechanism is used to weightedly fuse the two feature maps: the similarity matrix between the photoelectric feature map and the infrared feature map in the channel dimension is calculated to generate attention weights; the infrared features are then reweighted by channel and added element-wise to the photoelectric features to generate an enhanced visual fusion feature map with a spatial resolution of 64×64 and 512 channels.

[0066] Meanwhile, the point cloud data output by the phased array radar module, after clustering and motion compensation, extracts the echo intensity, Doppler velocity, range gate information, etc., of each potential target and encodes them into a 128-dimensional vector; the radio frequency signal captured by the passive radio frequency detection module, after fast Fourier transform and time-frequency analysis, extracts parameters such as center frequency, bandwidth, modulation type, and signal strength, and encodes them into another 128-dimensional vector. The two are concatenated to form a 256-dimensional radar radio frequency encoded vector. Finally, the visual fusion feature map and the radar radio frequency encoded vector are fed into a multimodal Transformer encoder for deep association. The specific workflow of the multimodal Transformer encoder is as follows: First, the 64×64×512 visual fusion feature map is flattened into 4096 feature vectors according to spatial position, each vector having a dimension of 512, and a two-dimensional sinusoidal position code is added; the radar radio frequency encoded vector is then added to the sequence as the 4097th independent feature vector. All feature vectors are superimposed with learnable modality type embedding vectors (photoelectric / infrared is one type, radar / RF is another type) to form a multimodal input sequence of length 4097.

[0067] The sequence is fed into a network consisting of 12 stacked Transformer encoding layers. Each layer includes a multi-head self-attention mechanism (8 heads) and a feedforward neural network, achieving full interaction of cross-modal information by calculating the correlation weights between all elements in the sequence. Finally, the spatial feature vectors corresponding to the target candidate regions in the output sequence (determined through a region proposal mechanism) are extracted and classified and regressed through a three-layer fully connected network, outputting the probability of target presence and three-dimensional spatial coordinates. Velocity vector A multi-dimensional target situation vector containing size estimation and preliminary attribute labels (such as rotor / fixed-wing, commercial / domestic). This multi-dimensional target situation vector is updated at a frequency of 10 frames per second and transmitted in real time to the threat assessment and intent recognition unit via a high-speed data bus.

[0068] The threat assessment and intent recognition unit receives a multi-dimensional target situation vector from the multimodal fusion detection unit, performs dynamic threat level determination and flight intent prediction, and its logical flow is shown in the appendix. Figure 3As shown, the threat assessment and intent recognition unit has a built-in prior knowledge base storing the geometric features, dynamic parameters, radio frequency fingerprints, and typical behavioral patterns of over 500 typical drone models. The threat assessment process is performed based on a multi-dimensional weighted scoring model. The input parameters of this multi-dimensional weighted scoring model include: target relative distance (in meters), approach speed (in meters per second), flight altitude (in meters), model matching degree (similarity score between 0 and 1), signal feature anomaly degree (calculated based on the KL divergence between the radio frequency fingerprint and the known database; the larger the value, the more abnormal), and whether it is within a no-fly zone defined by a geofence (Boolean value, 1 for yes, 0 for no). Each parameter is linearly weighted and summed according to a preset threat contribution weight.

[0069] For example, the relative distance weight is -0.002 (closer distance means higher threat), the approach speed weight is 0.05, the flight altitude weight is -0.001 (lower altitudes are more dangerous), the aircraft type matching weight is 0.8, the signal anomaly weight is 0.6, and the no-fly zone sign weight is 2.0. In addition, the multi-dimensional weighted scoring model also incorporates a risk premium, which is based on the target's historical trajectory and predicts whether it will intrude into the protected area within the next 120 seconds. The trajectory prediction uses a Kalman filter combined with a kinematic constraint model; if the predicted intrusion probability exceeds 70%, the risk premium is 1.5; otherwise, it is 0. (Comprehensive Threat Index) The calculation formula is as follows:

[0070] ;

[0071] This represents the relative distance (taking negative values). To approach the speed, This represents the flight altitude (negative values). For model compatibility, For signal anomaly degree, This is a no-fly zone sign. For risk premium, , , , , , Corresponding weights. Overall Threat Index The value ranges from 0 to 10, with higher values ​​indicating a greater threat. When this occurs, the system determines that the target must be countered immediately.

[0072] The intent recognition process is implemented based on a Long Short-Term Memory (LSTM) network model. This LSM model takes a sequence of the target's motion states at 10 consecutive moments (0.1-second intervals) as input. Each moment's state includes position, velocity, heading angle, and relative azimuth angle to a preset point of interest. The network structure consists of two LSTM layers with 256 hidden units. The output layer is a Softmax classifier, outputting probability distributions for four types of intents: reconnaissance (circling around the point of interest), approach (approaching at high speed in a straight line), loitering (moving erratically in a local area), and retreat (moving away from the point of interest). This LSM network model was trained offline using a real flight dataset, achieving an accuracy exceeding 92%. The intent recognition results, together with the threat index, constitute a complete cognitive profile of the target and serve as the decision-making basis for the collaborative countermeasure strategy generation unit.

[0073] Based on the above findings, the collaborative countermeasure strategy generation unit generates the optimal collaborative countermeasure instruction sequence, the principle framework of which is shown in the appendix. Figure 4 As shown, the collaborative countermeasure strategy generation unit includes a policy solver based on multi-agent reinforcement learning. The system models the N targets to be countered and the M available countermeasure resources in the current airspace as a partially observable Markov decision process. The countermeasure resources include a directional radio frequency jamming module, a navigation signal deception module, and a net-capture transmission module, each of which is considered an independent agent.

[0074] The state space S of the strategy solver consists of the following elements: the threat index vector of all targets (N-dimensional), the intent classification probability matrix (N×4), the three-dimensional spatial position matrix (N×3), and the state (ready / cooled / faulty) and current position (M×3) of each countermeasure resource. The action space A is defined as follows: for each countermeasure resource i, in the next decision cycle, one of four actions is selected to be executed: "interfering with target j", "deceiving target k", "launching a net to capture target l", or "standing still". The design of the reward function R comprehensively considers four indicators: countermeasure success rate (+5 points for successfully causing the target to lose control or return to base), resource consumption cost (-1, -1.5, and -3 points are deducted for each use of interference / deception / net capture, respectively), collateral impact on friendly equipment (-2 points are deducted if the interference beam covers the friendly frequency band), and overall airspace security improvement (an additional 10 points are added after all high-threat targets are eliminated).

[0075] The policy solver employs a centralized training and decentralized execution architecture. The training phase takes place in a high-fidelity digital twin simulation system, which accurately simulates the six-degree-of-freedom motion of a tethered platform, the impact of atmospheric turbulence on photoelectric imaging, the loss of radio frequency propagation due to rain attenuation, and the flight control response characteristics and anti-interference thresholds of various UAVs (such as DJI Mavic and Parrot Anafi). The training algorithm uses near-end policy optimization, and the policy network converges after 5 million rounds of adversarial exercises. During online deployment, the policy solver receives updated state information every second and generates an M×(N+1) policy matrix through forward inference. Each row corresponds to a countermeasure resource, and each column corresponds to a target or standby state. The matrix elements represent the probability of executing the action. The system selects the action with the highest probability as the command for that countermeasure resource and applies a confidence threshold (actions below 0.6 are not executed) to ensure the reliability of the decision.

[0076] The dynamic resource scheduling and execution unit is responsible for translating the policy matrix into precise physical execution instructions. Its multi-level interaction relationships are shown in the appendix. Figure 5 As shown. This dynamic resource scheduling and execution unit includes a real-time resource scheduler and three countermeasure equipment controllers. The real-time resource scheduler uses a mixed-integer programming model based on time window constraints for scheduling calculation. The optimization objective is to minimize the sum of the countermeasure completion times for all high-threat targets. Hard constraints include: the antenna rotation angular velocity of the directional radio frequency jamming module should not exceed 30 degrees / second, and the effective range should not exceed 3000 meters; the navigation signal deception module needs to start signal synchronization 2 seconds in advance, and the deception signal power should be at least 6 dB higher than the real signal; the launch elevation angle range of the net capture launch module is 10 degrees to 80 degrees, the initial velocity is adjustable from 20 to 50 meters / second, and the trajectory must avoid friendly airspace. For the net capture mission, the scheduler also needs to solve the ballistic dynamics equations:

[0077] ;

[0078] and The flight times are respectively At that time, the horizontal and vertical positions of the net capture and launch module are... Initial velocity, Angle of elevation The acceleration due to gravity is 9.8 m / s². The flight time is denoted by θ. The scheduler uses numerical methods to solve for the (v0, θ) combination that satisfies the target position and verifies that it is within the feasible region. The scheduler runs a fast branch and bound algorithm every 0.5 seconds, continuously updating the fine-grained scheduling instruction sequence for the next 5 seconds, including the action start time, pointing angle, power parameters, etc. of each piece of equipment, and sends it to the corresponding controller.

[0079] Upon receiving the command, the directional radio frequency jamming module controller drives the high-gain parabolic antenna (30 dB gain) to complete the pointing adjustment within 0.5 seconds and injects broadband noise interference or specific protocol-breaking interference targeting the target's remote control link or image transmission link. The navigation signal spoofing module controller simultaneously generates GPS...

[0080] The deceptive signals in the L1 band (1575.42 MHz) and the Beidou B1 band (1561.098 MHz) maintain sub-meter-level synchronization with the real satellite signals in terms of pseudorange and carrier phase, inducing the target's flight control system to misjudge its position. The net-capture launch module controller, based on the calculated launch parameters, controls the compressed air drive to launch a 2-meter-diameter Kevlar capture net, which carries a miniature deceleration parachute to expand the interception cross-section. During execution, all countermeasures equipment transmits real-time status information (such as the actual antenna pointing angle, jamming power, and net-projectile flight trajectory) via built-in sensors and uploads it to the multi-modal fusion detection unit through the tethered platform's data link. This information is used to evaluate the countermeasures' effectiveness and trigger a new detection-decision cycle.

[0081] The system health management and reconstruction unit operates throughout the entire system operation process. This unit continuously monitors the health status of each component through an embedded diagnostic agent. For the multimodal fusion detection unit, monitoring indicators include the signal-to-noise ratio of the photoelectric camera, the non-uniformity residual of the infrared detector, radar transmit power fluctuations, and the noise floor level of the RF receiver. If the performance degradation of a sensor exceeds a threshold, the modal reconstruction algorithm is immediately activated: in the cross-modal fusion algorithm, the weight coefficient of the cross-modal feature vector is reduced, while the weights of the remaining normal modes are increased, and the missing information is compensated by online fine-tuning of the last two layers of the Transformer encoder.

[0082] For countermeasures equipment, if the air pressure of the net-capture launch module is insufficient or the power amplifier of the jamming module overheats, the multi-modal fusion detection unit immediately marks its status as "unavailable" and sends a resource update message to the cooperative countermeasure strategy generation unit, triggering a re-solution of the strategy, with the remaining resources dynamically taking over the task. For example, if the net-capture fails, the strategy solver may switch to using high-intensity deception to induce the target to return autonomously. This adaptive reconfiguration mechanism ensures that the system can maintain more than 80% of its mission capability even under a single point of failure.

[0083] In summary, this embodiment utilizes the persistent hovering capability provided by the tethered platform to construct an anti-drone system integrating multi-modal perception, intelligent cognition, collaborative decision-making, precise execution, and self-healing reconfiguration. The various units collaborate efficiently under a unified architecture, achieving high-precision detection, high-confidence threat assessment, and highly efficient collaborative countermeasures against low-speed, small targets, outperforming existing discrete and reactive countermeasure solutions.

[0084] Example 2: Based on Example 1, this example further optimizes the cross-modal alignment mechanism in the multimodal fusion detection unit and the online learning capability in the cooperative countermeasure strategy generation unit to address the problems of extreme weather conditions and unknown new types of UAVs.

[0085] In the multimodal fusion detection unit, to address the degradation of photoelectric and infrared images caused by rain and fog, a cross-modal alignment enhancement module guided by a physical model is introduced. This cross-modal alignment enhancement module first utilizes the precise range information acquired by the phased array radar module to construct a sparse depth map of the target area.

[0086] A lightweight depth completion network expands sparse depth maps into dense depth maps, which are then used as geometric constraints to dehaze and enhance the contrast of degraded photoelectric and infrared images. Specifically, depth information is encoded as a spatial attention mask, guiding a bi-branch convolutional neural network to focus more on edges and structural regions less affected by weather conditions during the feature extraction stage.

[0087] A new "meteorological interference factor" embedding vector is added to the input sequence of the multimodal Transformer encoder. This vector is generated by normalizing visibility, humidity, and precipitation data uploaded from ground weather stations, and has a dimension of 32. This embedding vector, along with visual and radar radio frequency features, participates in self-attention calculation, enabling the model to adaptively suppress feature channels severely affected by meteorological interference. Experiments show that under dense fog conditions with visibility below 500 meters, the enhancement mechanism improves target detection recall.

[0088] In the collaborative countermeasure strategy generation unit, for newly encountered unknown UAV models (whose flight control protocols and anti-jamming characteristics are not recorded in the prior knowledge base), the system enables an online rapid adaptation mechanism based on meta-learning. When the threat assessment unit determines that the target is "high threat but the model matching degree is less than 0.3", the collaborative countermeasure strategy generation unit automatically switches to exploration mode. In this mode, the strategy solver no longer relies on a pre-trained fixed policy network, but instead activates the meta-controller.

[0089] The meta-controller maintains a small experience replay buffer, recording the five most recent probing countermeasures against the unknown target (such as low-power jamming and short-term deception) and their observed responses (such as whether the target hovers, changes course, or switches frequency bands). Based on these limited samples, the meta-controller employs a model-independent meta-learning algorithm to fine-tune the parameters of the last two layers of the policy network within 10 seconds, generating customized countermeasure strategies for specific targets. For example, if the target is observed to immediately switch from 2.4 GHz to 5.8 GHz under 2.4 GHz jamming, subsequent strategies will prioritize dual-band synchronous jamming. This online rapid adaptation mechanism enables the system to formulate countermeasures within 30 seconds when facing zero-day attack drones, significantly improving the system's generalization ability and combat adaptability.

[0090] Furthermore, the dynamic resource scheduling and execution unit in this embodiment adds priority arbitration logic for resolving multi-objective conflicts. When multiple high-threat targets appear simultaneously and countermeasure resources are insufficient, the scheduler no longer sorts targets solely by threat index but introduces a mission criticality factor. This mission criticality factor is dynamically issued by the ground command system; for example, a factor of 1.5 is assigned to targets heading towards nuclear power plants, while a factor of 1.0 is assigned to targets heading towards ordinary warehouses. In the objective function of the mixed-integer programming model, the scheduler multiplies the countermeasure completion time by the mission criticality factor and then sums them up, thereby ensuring that the most critical targets are dealt with first. This improvement enhances the system's ability to protect high-value assets in complex urban environments.

[0091] Through the above enhancements, this embodiment has achieved further breakthroughs in harsh environment adaptability, unknown target response capability, and mission priority management, enabling the system to maintain efficient and reliable countermeasure performance in a wider range of combat scenarios.

[0092] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0093] 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 multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles (UAVs) based on a tethered platform, characterized in that, include: The multimodal fusion detection unit is used to perceive the target airspace in multiple dimensions. The multimodal fusion detection unit integrates a high-resolution photoelectric imaging module, a long-wave infrared thermal imaging module, a phased array radar module, and a passive radio frequency detection module. The threat assessment and intent recognition unit is used to receive and process multi-dimensional target situation vectors from the multimodal fusion detection unit to perform dynamic threat level determination and flight intent prediction. The collaborative countermeasure strategy generation unit is used to generate a collaborative countermeasure instruction sequence based on the output of the threat assessment and intent recognition unit. The dynamic resource scheduling and execution unit is used to receive and parse the strategy matrix issued by the collaborative countermeasure strategy generation unit, and drive specific countermeasure equipment to perform countermeasure actions.

2. The UAV multi-mode detection and cooperative countermeasure system based on a tethered platform according to claim 1, characterized in that, The high-resolution photoelectric imaging module and the long-wave infrared thermal imaging module achieve physical axis consistency through a common optical path design and share a servo turntable with three-axis stabilization and high-precision pointing function. The phased array radar module and the passive radio frequency detection module's antenna arrays are integrated on the same carrier in a spatial multiplexing manner. This same carrier is mechanically coupled to the servo turntable through a rigid connection. The multimodal fusion detection unit internally runs a deep learning-based cross-modal feature-level fusion algorithm. The processing flow of this cross-modal feature-level fusion algorithm is as follows: Each detection module preprocesses the raw sensing data into feature maps under a unified spatiotemporal reference. Furthermore, the photoelectric image feature maps and infrared thermal image feature maps are extracted in parallel via a dual-branch convolutional neural network and then weighted and fused through a cross-channel attention mechanism to generate enhanced visual fusion features. Radar point cloud features and radio frequency signal spectral features are encoded into high-dimensional vectors; Visual fusion features and radar radio frequency encoded vectors are input into a multimodal Transformer encoder for deep correlation and information complementarity, and the output is a multidimensional target situation vector containing the target presence probability, spatial location, motion state and preliminary attribute labels.

3. The UAV multi-mode detection and cooperative countermeasure system based on a tethered platform according to claim 2, characterized in that, The threat assessment and intent recognition unit has a built-in prior knowledge base containing the characteristics and behavioral patterns of various typical UAV models; The threat assessment process is performed based on a multidimensional weighted scoring model. The input parameters of this multidimensional weighted scoring model include the target relative distance, approach speed, flight altitude, aircraft type matching degree, signal feature anomaly degree, and whether it is in a no-fly zone. Each parameter is linearly weighted and summed according to the preset threat contribution weight, and a risk premium based on the prediction of whether the target will invade the protected area in the next 2 minutes is added to calculate the final quantitative comprehensive threat index. The intent recognition process is based on a long short-term memory network model, which takes a sequence of the target's motion states over 10 consecutive time steps as input and outputs the classification probability of the target's possible intent.

4. A multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles based on a tethered platform according to claim 3, characterized in that, The collaborative countermeasure strategy generation unit includes a policy solver based on multi-agent reinforcement learning; The strategy solver models multiple targets to be countered in the current airspace and various countermeasure resources of the system as a partial Markov decision process. Countermeasure resources include a directional radio frequency jamming module, a navigation signal deception module, and a net capture and transmission module; the state space of the strategy solver is composed of the threat index, intent classification, spatial distribution of all targets, and the state and position of each countermeasure resource. Action space is defined as the period in which each countermeasure resource is assigned a countermeasure action against the target or remains in standby mode in the next decision cycle. The design of the reward function takes into account the success rate of counter-attack, resource consumption costs, collateral impact on friendly equipment, and the improvement of overall airspace security. The cooperative countermeasure strategy generation unit combines offline training with online fine-tuning, enabling the strategy solver to solve the cooperative countermeasure strategy matrix that maximizes long-term cumulative rewards in real time.

5. A multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles based on a tethered platform according to claim 4, characterized in that, The dynamic resource scheduling and execution unit includes a real-time resource scheduler and multiple countermeasures equipment controllers. The real-time resource scheduler calculates the scheduling timeline and spatial path planning based on the strategy matrix and the physical location, pointing capability, range of action, and readiness status of each countermeasure equipment. The directional radio frequency jamming module controller, according to instructions, controls its high-gain directional antenna beam to aim at the target UAV and injects jamming signals; The navigation signal deception module controller generates a deception signal that is spatiotemporally synchronized with the real satellite navigation signal; The net-capturing launch module controller calculates launch parameters and controls the launching device to launch a kinetic projectile carrying the capture net.

6. A multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles based on a tethered platform according to claim 5, characterized in that, The multimodal Transformer encoder in the cross-modal feature-level fusion algorithm has the following specific workflow: The visual fusion feature map is converted into a visual feature sequence through spatial location encoding and flattening operations; at the same time, the radar radio frequency encoded vector is used as an independent modal feature vector. The visual feature sequence and the radar radio frequency feature vector are concatenated into a multimodal input sequence, and a modality type embedding vector is added; this multimodal input sequence is then fed into a network consisting of 12 stacked Transformer coding layers for processing. Within each coding layer, the association weights between all elements within the sequence are calculated using a self-attention mechanism. The feature vector corresponding to the target region in the encoder output sequence is taken, and then classified and regressed through a fully connected layer to output the final multidimensional target situation vector.

7. A multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles based on a tethered platform according to claim 6, characterized in that, The threat contribution weights of the multidimensional weighted scoring model in the threat assessment and intent recognition unit are dynamically optimized through an offline-trained adaptive weight allocation network. The adaptive weight allocation network uses historical interception case data as a training set, inputs the parameter values ​​when the case occurred, and outputs a qualitative threat level label for the historical interception case. The network continuously adjusts the weight coefficients of each input parameter through the backpropagation algorithm to maximize the correlation coefficient between the model's score output and the real threat level. The system triggers a fine-tuning update of the weight allocation network every 100 new complete disposal cases accumulated.

8. A multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles based on a tethered platform according to claim 7, characterized in that, The multi-agent reinforcement learning policy solver in the collaborative countermeasure strategy generation unit is trained in a high-fidelity digital twin simulation system. The simulation system accurately simulates the dynamic characteristics of the tethered platform, the noise model of each detection sensor, the propagation attenuation under different meteorological conditions, and the flight dynamics and anti-interference model of various UAVs. Each counter-resource in the policy solver is modeled as an independent agent, which shares a centralized evaluation network but has its own independent behavioral policy network. The training employs a near-end policy optimization algorithm and conducts millions of rounds of adversarial drills in a simulated environment. The trained policy network parameters are deployed to the policy solver of the actual system, and transfer learning and rapid fine-tuning are performed based on a small amount of online data collected during actual operation.

9. A multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles based on a tethered platform according to claim 8, characterized in that, The real-time resource scheduler in the dynamic resource scheduling and execution unit uses a mixed integer programming model based on time window constraints for scheduling calculation; The mixed-integer programming model takes minimizing the overall countermeasure mission completion time as the primary optimization objective, while using the physical motion constraints, the range of action constraints, and the temporal dependencies between tasks of each countermeasure equipment as hard constraints. For the net capture and launch module, its scheduling calculation also needs to take into account the ballistic dynamics equations to ensure that the launch parameters are within the feasible region; the scheduler runs a fast solution algorithm every 0.5 seconds, and updates the fine-grained scheduling instruction sequence for the next 5 seconds based on the latest target situation and resource status, and sends it to the corresponding countermeasure equipment controller for execution.

10. A multi-mode detection and cooperative countermeasure system for unmanned aerial vehicles based on a tethered platform according to claim 9, characterized in that, The system also includes a system health management and refactoring unit; The system health management and reconstruction unit continuously monitors the health status of each sensor module in the multimodal fusion detection unit, the computing node in the collaborative countermeasure strategy generation unit, and each countermeasure equipment in the dynamic resource scheduling and execution unit. When any sensor is detected to have malfunctioned or degraded in performance, the system's health management and reconstruction unit immediately activates the modal reconstruction algorithm, adjusts the weight allocation of the feature-level fusion algorithm in the multimodal fusion detection unit, reduces the contribution of the faulty mode, and enhances the confidence of the remaining normal modes. When a countermeasure fails, the multimodal fusion detection unit sends a resource status update to the cooperative countermeasure strategy generation unit, triggering a re-solution of the strategy.

Images