Low-altitude unmanned aerial vehicle reconnaissance and disposal integrated protection system

Abstract

The application provides a low-altitude unmanned aerial vehicle reconnaissance and disposal integrated protection system, and belongs to the technical field of low-altitude defense, airspace safety and unmanned aerial vehicle management, and comprises a low-altitude intrusion detection and tracking module, which obtains target area information through a radar and a camera, respectively performs 4DFFT pretreatment on radar data and image pretreatment on an RGB image; the pretreated data is input into respective input adaptation modules to be unified into a format required by a target detection and tracking algorithm; a multi-source information fusion decision mechanism is used to realize real-time and high-precision unmanned aerial vehicle detection and tracking functions. An unmanned aerial vehicle identity identification module obtains unmanned aerial vehicle identity information through an RFID label, and uses a joint optimization power control and time slot allocation algorithm to verify the identity legality of all unmanned aerial vehicles in a region. An unmanned aerial vehicle intelligent scheduling module optimizes path planning, task allocation and resource management through reinforcement learning, dynamically allocates tasks and optimizes route planning.

Description

Technical Field

[0001] The present invention relates to the fields of low-altitude defense, airspace security and unmanned aerial vehicle (UAV) management technology, and in particular to a low-altitude UAV reconnaissance and disposal integrated protection system. Background Art

[0002] With the rapid development of drone technology and 5G-A / 6G communication technologies, drones are increasingly being used in urban environments, encompassing a wide range of applications, including surveillance, logistics, and disaster relief. They demonstrate significant potential for improving production efficiency, optimizing resource allocation, and promoting the development of the low-altitude economy. However, unauthorized drone activity poses a potential security threat, posing risks to critical infrastructure and privacy. To combat illegal drone activity, counter-drone systems have garnered significant attention. Key functions include drone positioning, tracking, decision-making, and response strategies. Among these functions, drone detection plays a crucial role in enabling real-time tracking, assessment, and decision-making. Furthermore, behavioral analysis based on precise positioning is crucial for rapid response, effective combat against illegal activities, and protecting surveillance areas. Therefore, improving drone positioning and identification accuracy is key to enhancing the overall performance of counter-drone systems. However, unlike the positioning of other targets (such as pedestrians or vehicles), drone detection in complex backgrounds presents significant challenges. First, due to the small size and low altitude of civilian drones, their signal characteristics are not conspicuous to many sensors, making them difficult to distinguish from background and other flying targets (such as birds or aircraft). Secondly, drones have omnidirectional, high-speed and unpredictable movement characteristics in three-dimensional space, which requires the positioning algorithm to have powerful positioning capabilities with high precision, fast response and low false alarm rate.

[0003] Despite the availability of a variety of drone detection methods, significant limitations exist in practical applications, impacting the overall effectiveness of anti-drone systems. Traditional radar detection methods offer all-weather detection capabilities under varying lighting conditions and offer high positioning accuracy. However, for drones with low radar cross-sections (RCS), slow speeds, and low altitudes, radar systems are prone to blind spots, limiting their detection performance in low-altitude environments. While microphone array-based technologies can utilize the distinctive sound of rotor blades to detect drones, their positioning effectiveness degrades significantly in noisy environments and their detection range is limited. Vision-based drone detection offers unique advantages, enabling intuitive identification and tracking by capturing and analyzing drones' visual features. However, in practical applications, these approaches often face challenges such as large scale variations, low target resolution, insufficient detail, and susceptibility to background occlusion, all of which hinder the overall detection performance of the system. Therefore, research on how to effectively detect and track drones is not only crucial for urban airspace management and security monitoring but also crucial for promoting the rapid development of the low-altitude economy.

[0004] The information disclosed in this background technology section is only intended to enhance understanding of the overall background of the invention and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the Invention

[0005] The purpose of the present invention is to provide a low-altitude UAV reconnaissance and disposal integrated protection system that can improve overall detection performance.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A low-altitude UAV reconnaissance and disposal integrated protection system, comprising:

[0008] A low-altitude intrusion detection and tracking module, which includes at least one integrated radar and vision device;

[0009] The radar and vision integrated device is composed of a millimeter-wave radar and a long-range camera; the radar and vision integrated device obtains information about the surveyed area in real time through a wired transmission channel, a 5G-A or 6G wireless transmission channel, and the information is respectively encapsulated in a frame format;

[0010] The data analysis module is used to analyze radar information and visual information, and synchronize the time and space of radar information and visual information by downsampling at the lowest frequency;

[0011] The data preprocessing module is used to preprocess the radar-visual integrated information encapsulated by the calibrated radar information and visual information.

[0012] Optionally, the preprocessing of the radar-visual integrated information encapsulated from the calibrated radar information and visual information includes:

[0013] The radar information is preprocessed in the order of range fast Fourier transform, Doppler fast Fourier transform, elevation fast Fourier transform and azimuth fast Fourier transform to complete the 4DFFT cube data preprocessing;

[0014] Sending the 4DFFT cube data and visual information to an adaptive gated convolutional target detection and tracking network; the adaptive gated convolutional target detection and tracking network includes a radar adaptation module, a visual adaptation module, and an adaptive gated convolutional neural network;

[0015] The adaptive gated convolutional target detection and tracking network is used to receive the 4DFFT cube data and visual data sent by the data preprocessing module; and send the 4DFFT cube data to the radar adaptation module to ensure the uniformity of the input data; then, the radar adaptation module and the adaptive gated convolutional neural network perform feature extraction to determine the azimuth and pitch angle of the target, and send the results to the multi-source information fusion decision module.

[0016] Optionally, the multi-source information fusion decision module is used to receive the target azimuth and pitch angle information, analyze the regional position of the target in the visual information; crop the visual information according to the regional position, and send the cropped visual information to the visual adaptation module for further processing.

[0017] Optionally, the visual adaptation module is used to process the cropped visual information, extract target features, and generate a target tracking frame; map the target tracking frame back to the visual information before cropping, and confirm whether the target is a drone; if it is a drone, send the visual information to the drone identification module for identity verification; if it is other flying objects, send the visual information to the drone intelligent scheduling module and the alarm and disposal module respectively; if it is background interference, send abnormal information to the alarm and disposal module.

[0018] Optionally, also include:

[0019] The drone identity recognition module is used to provide energy for the drone's passive RFID tag through the signal sent by the ground base station reader; the drone provides identity and status information to the ground base station by modulating the backscatter signal; determines the legitimacy of the drone's identity; if the drone's identity is legal, continues to determine the legitimacy of the drone's flight path; if the flight path is legal, maintains the mission process; if the flight path is illegal, sends the information and drone status information to the drone intelligent scheduling module; if the drone's identity is illegal, sends visual information and drone status information to the drone intelligent scheduling module and alarm and disposal module respectively.

[0020] Optionally, it also includes: a joint optimization power control and time slot allocation algorithm, wherein:

[0021] Multiple ground base station readers use power modulation to avoid overlapping recognition ranges;

[0022] Reduce signal interference when multiple drones compete for channels through time slot allocation strategies;

[0023] The time slot allocation strategy includes:

[0024] Multiple polling cycles, with random selection of polling, operating channel, and transmit power at the beginning of each cycle;

[0025] The ground base station initiates the frame operation by broadcasting a Query command. The frame contains multiple time slots, and each drone randomly selects a time slot to respond.

[0026] After the frame ends, the ground base station sends the drone’s RFID tag information to the registration server to confirm its legitimacy;

[0027] Time slot response events include:

[0028] There are three types: idle, success and collision;

[0029] The number of drone RFID tags in the ground base station area is calculated based on maximum likelihood estimation, and the polling cycle and time are optimized.

[0030] Optionally, it also includes: a drone intelligent scheduling module for state perception, where the drone sends its current environmental information and local information to the scheduling center to determine whether intelligent scheduling is required;

[0031] If scheduling is required, the intelligent scheduling center adjusts the drone's path planning, cruising speed, and subsequent tasks and resource allocation based on the drone's current status information and global events;

[0032] If no scheduling is required, maintain the current cruise or mission status.

[0033] Optionally, it also includes: a UAV intelligent scheduling strategy optimization module, which is used to optimize the preset initial strategy using a genetic algorithm based on the UAV status information;

[0034] Based on the initial strategy, Double Q-learning is used to dynamically update the strategy according to the local area environment information;

[0035] In Double Q-learning reinforcement learning, the intelligent dispatch center balances exploring new strategies and utilizing existing strategies based on the drone status information.

[0036] Receive global regional data feedback in real time to analyze and optimize strategies;

[0037] Combine the dispatch center's decision-making and Double Q-learning optimization strategy to re-plan paths and assign tasks for each drone;

[0038] For long-term emergencies and abnormal events, adaptive optimization strategies are used to maximize long-term benefits, thereby gradually improving scheduling efficiency and task completion.

[0039] Optionally, also include:

[0040] The alarm and handling module is characterized by including:

[0041] Large models assist analysis and decision-making. Based on the needs of intelligent drone scheduling, a large model is built to analyze drone status information in real time and provide optimized handling decisions;

[0042] Large models assist in the training of analysis and decision-making models. They use crawler technology to obtain the required training data and combine format conversion and data cleaning to generate high-quality data sets that meet training requirements.

[0043] The large-scale model-assisted analysis and decision-making model combines pre-training, supervised fine-tuning and human preference optimization methods to improve the analysis and decision-making capabilities of the large-scale model-assisted analysis and decision-making model for complex scenarios;

[0044] The large model-assisted analysis and decision-making model adopts the LLaMA series as a benchmark and expands its capabilities through knowledge enhancement and domain adaptation technologies;

[0045] The large-scale model-assisted analysis and decision-making model prioritizes emergencies based on global and local environmental information, and formulates hierarchical handling strategies to achieve dynamic alarm response efficiency.

[0046] Optionally, it also includes: an intrusion threat processing and response module for real-time monitoring of controlled airspace and collecting data such as target flight trajectory and spectrum characteristics;

[0047] Conduct preliminary threat identification of potential targets based on flight behavior, speed, and position parameters;

[0048] When abnormal flight behavior or possible low-altitude intrusion is detected, the intrusion threat processing and response module immediately generates an alarm signal and notifies the dispatch center;

[0049] After receiving the alarm signal, the dispatch center generates an auxiliary decision-making plan;

[0050] For targets determined to be non-threatening, the system continuously tracks and dynamically updates their threat registration;

[0051] For targets identified as high-threat, the system automatically initiates the emergency response process;

[0052] The emergency response process specifies the optimal response plan based on target characteristics, image information, and current environment, combined with existing dispatchable resources;

[0053] The intrusion threat processing and response module executes the disposal plan in real time and transmits key information to relevant management personnel;

[0054] The relevant managers can view the target status, resource allocation and disposal progress in real time through the intelligent human-computer interaction interface;

[0055] The emergency response process supports manual intervention and dynamic command, which can further optimize the corresponding process;

[0056] It also includes a registration server for storing the RFID tag of each legally registered drone.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The integrated low-altitude drone reconnaissance and disposal protection system provided by this invention addresses current technical bottlenecks in drone detection and protection by proposing a series of innovative methods and improvements, effectively enhancing drone detection, tracking, identification, and disposal capabilities. The system integrates radar and visual data, effectively combining the characteristics of 4DFFT cube data and RGB images to propose an adaptive gated convolutional neural network. This system achieves precise detection and tracking of slow-moving, hovering, and drone targets in complex backgrounds. This significantly reduces the blind spots of traditional radars when detecting low-RCS targets, significantly improving the accuracy and robustness of low-altitude drone detection. Regarding drone authentication, the system utilizes RFID tag technology, combined with power control and time slot allocation optimization algorithms, to efficiently verify the legitimacy of drones within a region, providing technical support for the accurate identification of illegal targets. The system optimizes drone scheduling through reinforcement learning, enabling dynamic adjustments in path planning, task allocation, and resource management to ensure maximum coverage and optimal task execution efficiency when multiple drones work together. This system can flexibly respond to emergencies in dynamic environments and enhance the monitoring capabilities and response speed of low-altitude protection zones. The warning and disposal module, combined with artificial intelligence algorithms, can assess the threat level in real time and trigger various disposal strategies including warning, isolation, interference, and interception according to specific circumstances, ensuring the rapid control and effective strike of illegal drones.

[0059] This invention is not only technologically innovative but also demonstrates immense value in terms of economic and social benefits. By enhancing the intelligence level of low-altitude drone management and dispatch, it significantly reduces the cost of low-altitude safety control and management, improves the safety of critical infrastructure, contributes to the healthy development of the low-altitude economy, and promotes the sustainable development of smart security. Therefore, this invention provides a comprehensive and efficient technical solution for low-altitude drone reconnaissance and disposal, with broad application prospects and promotional value. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1 A schematic diagram of the overall system flow provided by an embodiment of the present invention.

[0062] Figure 2 A schematic diagram of timestamp synchronization provided by an embodiment of the present invention.

[0063] Figure 3 4DFFT cube data processing flow chart provided by an embodiment of the present invention.

[0064] Figure 4 Schematic diagram of a target detection and tracking network based on adaptive gated convolution provided by an embodiment of the present invention.

[0065] Figure 5 This is a flow chart of feature extraction using an adaptive gated convolution block according to an embodiment of the present invention.

[0066] Figure 6 This is a preprocessing flow chart of the input adaptation module provided in an embodiment of the present invention.

[0067] Figure 7 This is a multi-source information fusion decision flow chart provided by an embodiment of the present invention.

[0068] Figure 8 This is a processing flow chart of the intrusion detection and tracking module provided by an embodiment of the present invention.

[0069] Figure 9 A schematic flow chart of a joint optimization power control and time slot allocation algorithm provided in an embodiment of the present invention.

[0070] Figure 10 Flowchart of the intelligent scheduling strategy for drones provided by an embodiment of the present invention.

[0071] Figure 11 This is a flowchart of the intelligent large model training process of the dispatching center platform provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0072] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0073] The purpose of the present invention is to provide a low-altitude UAV reconnaissance and disposal integrated protection system that can improve overall detection performance.

[0074] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] Example 1:

[0076] This embodiment provides an integrated protection system for low-altitude drone reconnaissance and disposal, which includes the following modules: a low-altitude intrusion detection and tracking module obtains target area information through radar and camera, performs 4DFFT preprocessing on radar data and image preprocessing on RGB images respectively; the preprocessed data is unified into the format required by the target detection and tracking algorithm through respective input adaptation modules to ensure consistency and compatibility of algorithm processing; finally, the proposed multi-source information fusion decision-making mechanism is used to achieve real-time and high-precision drone detection and tracking functions. The drone identification module obtains drone identity information through RFID tags and uses a joint optimization power control and time slot allocation algorithm to verify the legitimacy of the identities of all drones in the area. The drone intelligent scheduling module optimizes path planning, task allocation and resource management through reinforcement learning, dynamically allocates tasks and optimizes route planning to ensure that drones can flexibly respond to and complete reconnaissance missions, and ensure continuous monitoring and timely response of low-altitude protection areas. The alarm and disposal module combines large models to process data in real time and automatically assess the threat level, triggering corresponding disposal measures. This application can effectively realize the reconnaissance and disposal of low-altitude drones and ensure safety management within the security control area. The overall flow chart of the system is as follows: Figure 1 The specific process includes:

[0077] Step 1: Low-altitude intrusion detection and tracking

[0078] (1) Radar and vision fusion preprocessing

[0079] The present invention combines a high-resolution camera and an ultra-long-range millimeter-wave radar to achieve the detection and tracking of low-altitude drones. Since radar and camera are different types of sensors and have inconsistent sampling frequencies, this leads to differences in time and space between the two types of data. Therefore, before fusing radar data and visual data, time synchronization and spatial calibration are required. Time synchronization is relatively simple. By unifying all sensors to the reference frequency of the lowest-frequency sensor, frame synchronization is performed using downsampling. Figure 2 shown.

[0080] Because drones move randomly in three-dimensional space, traditional two-dimensional range-Doppler spectrum analysis is ineffective for locating them, as the signal lacks pitch angle information. Therefore, to more accurately extract targets and suppress clutter, it is necessary to observe the target's time-varying motion, including range, velocity, azimuth, and pitch angle. To this end, millimeter-wave radar can be used to acquire time-varying four-dimensional observation data of slow-moving targets, thereby determining the target's three-dimensional trajectory.

[0081] According to the Time Division Multiplexing Multiple-Input Multiple-Output (TDM-MIMO) strategy, T Transmitting antenna and receiving antenna N R The MIMO array is combined to contain N T N R A two-dimensional virtual array of array elements. Taking the frequency modulated continuous wave (FMCM) millimeter wave radar as an example, assuming that the signal When transmitted to the monitoring area, the backscattered signal from the target can be approximately expressed as:

[0082]

[0083] Where A is the amplitude, τ = 2[R0+v0(t f +t p )] / c0 represents the round trip time, t f ∈[0,T c ] represents fast time, T c represents the pulse duration, t c =n c T c (n c =0,1,2,...,N c -1) is the number of pulses, R0 is the number of transmitted chirps, R0 is the initial tilt distance, v0 is the radial velocity of the target, c0 is the speed of light, and k = c0 / f c is the wavelength, f c is the carrier frequency, θ0 and are the initial azimuth and elevation angles, n a =0,1,2,...,N a -1 and n e =0,1,2,...,N e -1 are the column and row numbers of the two-dimensional virtual array, The rectangular function is shown as follows:

[0084]

[0085] The first antenna is used as the reference element. Applying the 4D Fast Fourier Transform (FFT) to the inputs in equation (1) yields:

[0086]

[0087] where f r , f v , f a and f e are the frequency variables of range, Doppler, azimuth and elevation respectively, sinc(·)=sin(·) / · represents the function, and A0 is a constant. The detailed 4DFFT cube data processing flow is as follows Figure 3 shown.

[0088] (2) Adaptive Gated Convolutional Object Detection and Tracking Network

[0089] After preprocessing, the radar 4D cube data and RGB images have significant differences in characteristics. In order to achieve compatible processing of these two heterogeneous modal data, this paper proposes a target detection and tracking network based on adaptive gated convolution, such as Figure 4 shown.

[0090] The adaptive gated convolutional target detection and tracking network consists primarily of two components: an input adaptation module and a feature extraction module. The present invention proposes an adaptive gated convolutional neural network in the feature extraction module, capable of extracting deep semantic information from RGB images and temporal features from 4D cube radar data for different types of input data. Furthermore, an adaptive computation mechanism is implemented in the input adaptation module to process different types of input data, ensuring consistency and uniformly passing it to the feature extraction module, ultimately achieving efficient perception, detection, and tracking.

[0091] The adaptive gated convolutional network is composed of a stack of adaptive gated convolutional blocks. The adaptive gated convolutional blocks include multi-scale kernel convolution and adaptive gating, which aims to extract information with spatiotemporal multi-scale features from complex input data. Figure 5 shown.

[0092] Specifically, the input data X inputThe size of is B×C×H×W. For RGB image input, C contains deep semantic information from different color channels; for radar 4D cube data, C contains the real and imaginary part information of signals and frequencies from different chirps. First, the input data is biased by group normalization, and then the number of channels is expanded to twice the original through 1×1 convolution. Subsequently, after the data is split into channels, it is passed to the multi-scale kernel convolution branch and the adaptive gating branch for feature extraction. In the multi-scale kernel convolution branch, the data is passed to multiple parallel convolution branches, each branch uses a convolution kernel of a different size for feature extraction; here, 3×3 and 7×7 depthwise separable convolutions are used for feature extraction to capture local and global spatiotemporal information of each dimension under different receptive fields; the extracted features are standardized by group normalization to ensure the stability and consistency of the features. In the adaptive gating branch, the input data first undergoes different 1×1 convolutions to reduce the spatial and channel dimensions to 1, respectively, to prepare for subsequent attention calculations. Subsequently, a 1×7 convolution combined with a sigmoid activation function is used to refine the channel attention, while another sub-branch uses a 7×7 convolution combined with a sigmoid activation function to refine the spatial attention. Next, a dot product operation is performed on the channel attention map and the spatial attention map to generate an adaptive attention gating for each unit, thereby enhancing the representation's expressiveness and task adaptability. Finally, the outputs of the two branches are fused element-wise to achieve a joint expression of the channel and spatial attention features. The fused result is then further fused using a 1×1 convolution as a channel confuser to restore the dimensionality to C. Residual connections are introduced to preserve the original feature information and promote efficient gradient propagation.

[0093] The input adaptation module is divided into two types: radar input adaptation module and image input adaptation module. It mainly preprocesses the input data of different modes to provide input optimization for feature extraction of feature encoder. Its detailed structure is as follows: Figure 6 shown.

[0094] The image input adaptation module preprocesses RGB images, first applying a 7×7 convolution with a stride of 4 and padding of 3 to adjust them to the model input dimensions while compressing the spatial dimensions. A GELU activation function is then used to introduce nonlinear representation capabilities, and GN is used for normalization. The radar input adaptation module preprocesses radar 4D cube data, first splitting the input data by channel dimension and rearranging the channels in a cross-stacked manner. A 7×7 grouped convolution is then performed to achieve information obfuscation between different groups of features. Similar to the image input adaptation module, a 7×7 convolution is then used to adjust the features to the model input dimensions. GELU activation and GN are then used to further normalize and optimize the feature representation, completing the radar data preprocessing. Because radar data predictions are processed within a 4D cube, the stride of the 7×7 convolution is set to 1 to preserve higher spatial resolution.

[0095] (3) Multi-source information fusion decision-making

[0096] In complex low-altitude environments, relying solely on radar or vision alone is difficult to meet the needs of target detection and tracking: radar is insufficient in resolution and semantic information, while vision is less robust to the environment. Using either technology alone will lead to performance bottlenecks in the perception system. In order to fully utilize the complementary advantages of the two modalities, this paper proposes a multi-information fusion decision-making, as detailed in Figure 7 shown.

[0097] Specifically, by inspecting 4D cube data, the approximate location and motion state of candidate targets can be quickly and cost-effectively determined. The predicted azimuth and elevation angles provide prior information for visual inspection. Subsequently, visual inspection can finely perceive and classify the candidate areas provided by radar, compensating for radar's shortcomings of low resolution and insufficient detail. This avoids the high computational cost of global visual inspection while ensuring accurate target detection.

[0098] The processing flow chart of the entire low-altitude intrusion detection and tracking module is as follows: Figure 8 As shown:

[0099] Step 2: Drone Identification

[0100] Drone RFID tags typically use a passive structure, and their operation relies on obtaining energy from the received reader signal and communicating with the ground base station by modulating the backscattered signal. During the communication process between the ground base station and the drone, especially when the ground base station reader performs data storage operations, communication failure may occur due to signal interference. Specifically, when the ground base station reader signal received by the drone tag is interfered with by other readers, or when the drone tag signal received by the ground base station reader is interfered with by other drones, multiple signals with the same frequency may be received at the same time. In this case, the signals will interfere with each other, resulting in the ground base station being unable to correctly identify the drone's identity information, which in turn causes signal collision problems. This collision phenomenon significantly affects the reliability and efficiency of communication and is a challenge that needs to be addressed in drone RFID communications.

[0101] Aiming at the low-altitude safety control scenario, the present invention proposes a joint optimization power control and time slot allocation algorithm, as detailed in Figure 9 shown.

[0102] The algorithm optimizes the ground base station's recognition process for drone RFID tags by introducing a polling mechanism. Assume that the ground base station deploys n readers within a monitoring area with a radius of R. The transmission power of these readers can be adjusted to avoid overlapping recognition ranges. The reader's working channel can be randomly selected from m channels. When the ground base station is active, each of its readers sends polling instructions to multiple drone tags within its coverage area. The drone tags return the ID information to the reader through backscattered modulated signals according to a predetermined communication protocol. Within the monitoring range of the base station, due to mutual interference between the signals between readers, the interference phenomenon will become more serious, especially as the number of readers increases. In addition, when multiple drones sharing the same channel respond to the reader's instructions at the same time, signal collision problems will occur. In order to solve this problem, the present invention uses the widely used and standardized MAC layer protocol - EPC Global Class-1Generation-2 (C1G2). In this protocol, the operation cycle of each reader is T r At the beginning of each polling cycle, the reader randomly decides whether to perform polling, selects the working channel, and sets the transmission power. According to the EPC C1G2 protocol, the reader first issues a Query command to start the first frame. The rest of the frame contains multiple time slots. Each drone tag randomly selects a time slot to send a response message. After the frame ends, the base station uses maximum likelihood estimation to determine the number of drone tags N based on the drone response status of each time slot (including idle, successful, and collision types). r , and calculate the optimal true length based on this to start the next frame. Define the effective polling time of reader o as Q o, the polling period is T o This method optimizes the polling process, reduces channel interference and collision probability, and improves the efficiency and reliability of drone cluster identification in complex scenarios.

[0103] Step 3: Drone Intelligent Scheduling Strategy

[0104] With the rapid popularization of drone technology, its application scope continues to expand, especially in the fields of real-time monitoring and security patrols, but it also poses new challenges to low-altitude safety and protection. The lowering of the threshold for drone operation has led to the frequent occurrence of "illegal flying". These unauthorized drones may operate in no-fly zones or even be used maliciously, directly threatening regional security. In addition, the low-altitude environment itself is full of complexity and dynamic changes (such as buildings, vegetation, birds and other interference factors), which further increases the difficulty of protection. The unpredictability and diversity of emergencies (such as drone power exhaustion, hardware failure or communication terminal, etc.) make the low-altitude protection task more arduous. In the scheduling of multiple drones, the differences in task priority, flight status and resource distribution make university scheduling difficult, and the contradiction between the limited battery life of drones and the growing task requirements further aggravates the pressure of resource allocation. In response to the above problems, the present invention proposes a drone intelligent scheduling strategy, which can achieve efficient drone scheduling in complex dynamic environments by combining intelligent optimization and reinforcement learning. This algorithm is mainly used to optimize the task allocation, path planning and resource management of UAVs to ensure the efficient operation of the system in a changing environment, especially in low-altitude reconnaissance and protection missions. Figure 10 The specific steps are as follows:

[0105] (1) State perception: Each UAV perceives its local environment based on its current state (e.g., low-altitude intrusion trajectory, wind speed, obstacles, etc.). Perception information, including target location, distance, speed, etc., is transmitted as input to the dispatch center controller. This perception information helps the dispatch center optimize and adjust the UAV's mission.

[0106] (2) Strategy Generation (Genetic Algorithm Initialization): The genetic algorithm is used to optimize the initial strategy of the UAV. The genetic algorithm generates a set of initial scheduling strategies, which may include action selection in terms of path planning, task allocation, cruising speed, etc. The genetic algorithm simulates evolutionary operations such as selection, crossover, and mutation, giving the UAV the initial ability to perform tasks.

[0107] (3) Reinforcement learning update: Based on the initial strategy, the drone updates its strategy using Double Q-learning through interaction with the environment. Compared with traditional Q-learning methods, Double Q-learning mitigates the overestimation bias problem and improves learning stability by introducing two Q networks (Q1 and Q2). After each action, the drone adjusts its strategy based on feedback (such as rewards or penalties). The reward function is constrained by the drone's mission, coverage range, and confidence level of low-altitude intrusion targets.

[0108] (4) Balance between exploration and exploitation: In reinforcement learning, drones need to strike a balance between exploring new strategies and exploiting existing ones. Drones explore unknown strategies to discover potential better solutions, while simultaneously utilizing existing strategies to execute the current task. By dynamically adjusting the ratio of exploration to exploitation, drones can improve the efficiency of the overall reconnaissance mission.

[0109] (5) Real-time feedback and optimization: The dispatch center receives data feedback from each drone (such as target detection results, environmental changes, etc.) and performs real-time analysis and decision adjustments. The reinforcement learning algorithm continuously optimizes the drone's behavior strategy by analyzing the feedback, ensuring that it can flexibly respond to different low-altitude intrusion threats in complex and dynamic environments.

[0110] (6) Task Allocation and Path Planning: Based on a reinforcement learning optimization strategy, the intelligent dispatch center replans paths and assigns tasks to each drone. Path planning can dynamically avoid obstacles and ensure that there are no conflicts between drones. In this way, the system can achieve all-round, no-blind-angle low-altitude reconnaissance and protection.

[0111] (7) Long-term reward optimization: The goal of reinforcement learning is to maximize long-term cumulative rewards. Therefore, during mission execution, the drone continuously optimizes its strategy through trial and error, pursuing long-term benefits rather than short-term returns. Through this process, the system can gradually improve scheduling efficiency and mission completion in different mission scenarios.

[0112] Step 4: Alert and Handle

[0113] (1) Large-scale model-assisted analysis and decision-making model construction

[0114] Artificial Intelligence Generated Content (AIGC) is one of the latest advances in the field of artificial intelligence. It can generate original content, including text, images, audio, and other multimedia, by learning data patterns and patterns. AIGC can play a vital role in low-altitude drone reconnaissance and disposal applications. Low-altitude disposal methods are often documented in various regulations and books, while airspace records of the surveyed area and low-altitude flight paths are stored in information systems centered around inspection logs. Furthermore, the latest guidelines for low-altitude management are often presented in the form of articles or regulations. Low-altitude reconnaissance and disposal missions rely on a thorough understanding and judgment of the airspace, mission resources, and guidelines. In some cases, dispatchers may lack sufficient knowledge of currently available disposal resources, which can lead to erroneous disposal decisions. For example, when identifying an illegal flight, dispatchers may lack comprehensive resource information, resulting in suboptimal response decisions. In such cases, the use of auxiliary strategies generated by large-scale models can help dispatchers gain a comprehensive understanding of currently available resources, enabling them to develop more accurate and comprehensive disposal plans.

[0115] In order to make full use of advanced technology to realize the integrated protection of UAV low-altitude reconnaissance and disposal, the present invention constructs an intelligent large-scale model of the dispatching center platform. Figure 11 shown.

[0116] The training process mainly includes the following steps: 1) Corpus collection and preprocessing: Through crawler technology, format conversion and data cleaning, the corpus data required for experimental training is initially collected; 2) Construction of a low-altitude application and management question-answering dataset: Using knowledge guidance to generate question-answering data, and optimizing data quality, ultimately generating a high-quality question-answering dataset for the supervised fine-tuning stage. 3) Model training and evaluation: Based on the baseline model, through three stages of continued pre-training, supervised fine-tuning and direct preference optimization, a large intelligent model of the dispatching center platform is gradually constructed. Finally, the model is verified for performance through both automatic evaluation and manual evaluation. Considering the high cost of ChatGPT fine-tuning, this paper uses open source general models such as the LLaMA series as a benchmark. The training data mainly comes from open source datasets and automatically constructed small batch question-answering data. Although it has not undergone secondary cleaning and iterative optimization, it basically meets the requirements. The training method is mainly supervised fine-tuning or fine-tuning methods combined with continued pre-training, and there is currently little research on further optimization based on human feedback preferences. Therefore, the present invention proposes to achieve full-process training from continued pre-training to human preference optimization through massive and high-quality dialogue data.

[0117] (2) Intrusion threat processing and response

[0118] In the integrated low-altitude drone reconnaissance and disposal protection system, the alerting and intrusion response processes primarily rely on the threat identification and decision-making support capabilities of the intelligent large-scale model. First, the system continuously monitors the monitored airspace using all-weather detection equipment, collecting real-time data such as target flight paths and spectral characteristics. For each potential target, the system performs preliminary threat identification based on flight behavior, speed, and location. When an anomaly or potential low-altitude intrusion is detected, the system immediately issues an alert signal and notifies the dispatch center. Once the alert signal is triggered, the dispatch center platform, combined with the intelligent large-scale model, generates a supporting decision-making plan for further analysis and response to the intruder. For targets deemed non-threatening, the system continuously tracks and monitors changes in their threat level. For targets deemed high-threatening, the system immediately initiates a response process, automatically directing video surveillance to precisely track the target, capturing image data and transmitting it to the dispatch center for analysis. The dispatch center analyzes the image information and target characteristics, combined with currently available resources, to develop an optimal response plan. This plan is automatically communicated to relevant personnel, directing the dispatch of appropriate drones, patrol vehicles, or other embodied equipment for real-time response. Dispatchers can also conduct real-time command and dispatch through the intelligent human-computer interaction interface to ensure rapid and accurate handling of intrusion incidents and maximize the safety and stability of low-altitude airspace.

[0119] The integrated low-altitude drone reconnaissance and disposal protection system provided by this invention addresses current technical bottlenecks in drone detection and protection by proposing a series of innovative methods and improvements, effectively enhancing drone detection, tracking, identification, and disposal capabilities. The system integrates radar and visual data, effectively combining the characteristics of 4DFFT cube data and RGB images to propose an adaptive gated convolutional neural network. This system achieves precise detection and tracking of slow-moving, hovering, and drone targets in complex backgrounds. This significantly reduces the blind spots of traditional radars when detecting low-RCS targets, significantly improving the accuracy and robustness of low-altitude drone detection. Regarding drone authentication, the system utilizes RFID tag technology, combined with power control and time slot allocation optimization algorithms, to efficiently verify the legitimacy of drones within a region, providing technical support for the accurate identification of illegal targets. The system optimizes drone scheduling through reinforcement learning, enabling dynamic adjustments in path planning, task allocation, and resource management to ensure maximum coverage and optimal task execution efficiency when multiple drones work together. This system can flexibly respond to emergencies in dynamic environments and enhance the monitoring capabilities and response speed of low-altitude protection zones. The warning and disposal module, combined with artificial intelligence algorithms, can assess the threat level in real time and trigger various disposal strategies including warning, isolation, interference, and interception according to specific circumstances, ensuring the rapid control and effective strike of illegal drones.

[0120] This invention is not only technologically innovative but also demonstrates immense value in terms of economic and social benefits. By enhancing the intelligence level of low-altitude drone management and dispatch, it significantly reduces the cost of low-altitude safety control and management, improves the safety of critical infrastructure, contributes to the healthy development of the low-altitude economy, and promotes the sustainable development of smart security. Therefore, this invention provides a comprehensive and efficient technical solution for low-altitude drone reconnaissance and disposal, with broad application prospects and promotional value.

[0121] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0122] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A low-altitude UAV reconnaissance and disposal integrated protection system, characterized in that: include: A low-altitude intrusion detection and tracking module, which includes at least one integrated radar and vision device; The radar and vision integrated device is composed of a millimeter-wave radar and a long-range camera; the radar and vision integrated device obtains information about the surveyed area in real time through a wired transmission channel, a 5G-A or 6G wireless transmission channel, and the information is respectively encapsulated in a frame format; The data analysis module is used to analyze radar information and visual information, and synchronize the time and space of radar information and visual information by downsampling at the lowest frequency; The data preprocessing module is used to preprocess the radar-visual integrated information encapsulated by the calibrated radar information and visual information; The preprocessing of the radar-visual integrated information encapsulated from the calibrated radar information and visual information includes: The radar information is preprocessed in the order of range fast Fourier transform, Doppler fast Fourier transform, elevation fast Fourier transform and azimuth fast Fourier transform to complete the 4DFFT cube data preprocessing; Sending the 4DFFT cube data and visual information to an adaptive gated convolutional target detection and tracking network; the adaptive gated convolutional target detection and tracking network includes a radar adaptation module, a visual adaptation module, and an adaptive gated convolutional neural network; The adaptive gated convolutional target detection and tracking network is used to receive the 4DFFT cube data and visual data sent by the data preprocessing module; and send the 4DFFT cube data to the radar adaptation module to ensure the uniformity of the input data; then the radar adaptation module and the adaptive gated convolutional neural network perform feature extraction to determine the azimuth and pitch angle of the target, and send the results to the multi-source information fusion decision module; The multi-source information fusion decision module is used to receive the target azimuth and pitch angle information, analyze the regional position of the target in the visual information; crop the visual information according to the regional position, and send the cropped visual information to the visual adaptation module for further processing; The visual adaptation module is used to process the cropped visual information, extract target features, and generate a target tracking bounding box; map the target tracking bounding box back to the visual information before cropping, and confirm whether the target is a drone; if it is a drone, send the visual information to the drone identification module for identity verification; if it is other flying objects, send the visual information to the drone intelligent scheduling module and the alarm and disposal module respectively; if it is background interference, send an abnormal information to the alarm and disposal module; Also included: a joint optimization power control and time slot allocation algorithm, wherein: Multiple ground base station readers use power modulation to avoid overlapping recognition ranges; Reduce signal interference when multiple drones compete for channels through time slot allocation strategies; The time slot allocation strategy includes: Multiple polling cycles, with random selection of polling, operating channel, and transmit power at the beginning of each cycle; The ground base station initiates the frame operation by broadcasting a Query command. The frame contains multiple time slots, and each drone randomly selects a time slot to respond. After the frame ends, the ground base station sends the drone’s RFID tag information to the registration server to confirm its legitimacy; Time slot response events include: There are three types: idle, success and collision; The number of drone RFID tags in the ground base station area is calculated based on maximum likelihood estimation, and the polling cycle and time are optimized.

2. The low-altitude UAV reconnaissance and disposal integrated protection system according to claim 1 is characterized in that: Also includes: The drone identity recognition module is used to provide energy for the drone's passive RFID tag through the signal sent by the ground base station reader; the drone provides identity and status information to the ground base station by modulating the backscatter signal; determines the legitimacy of the drone's identity; if the drone's identity is legal, continues to determine the legitimacy of the drone's flight path; if the flight path is legal, maintains the mission process; if the flight path is illegal, sends the information and drone status information to the drone intelligent scheduling module; if the drone's identity is illegal, sends visual information and drone status information to the drone intelligent scheduling module and alarm and disposal module respectively.

3. The low-altitude UAV reconnaissance and disposal integrated protection system according to claim 2 is characterized in that: Also includes: The UAV intelligent dispatch module is used for state perception. The UAV sends its current environmental information and local information to the dispatch center to determine whether intelligent dispatch is needed. If scheduling is required, the intelligent scheduling center adjusts the drone's path planning, cruising speed, and subsequent tasks and resource allocation based on the drone's current status information and global events; If no scheduling is required, maintain the current cruise or mission status.

4. The low-altitude UAV reconnaissance and disposal integrated protection system according to claim 3 is characterized in that: Also includes: The UAV intelligent scheduling strategy optimization module is used to optimize the preset initial strategy using genetic algorithms based on the UAV status information; Based on the initial strategy, Double Q-learning is used to dynamically update the strategy according to the local area environment information; In Double Q-learning reinforcement learning, the intelligent dispatch center balances exploring new strategies and utilizing existing strategies based on the drone status information. Receive global regional data feedback in real time to analyze and optimize strategies; Combine the dispatch center's decision-making and Double Q-learning optimization strategy to re-plan paths and assign tasks for each drone; For long-term emergencies and abnormal events, adaptive optimization strategies are used to maximize long-term benefits, thereby gradually improving scheduling efficiency and task completion.

5. The low-altitude UAV reconnaissance and disposal integrated protection system according to claim 1 is characterized in that: Also includes: The alarm and handling module is characterized by including: Large models assist analysis and decision-making. Based on the needs of intelligent drone scheduling, a large model is built to analyze drone status information in real time and provide optimized handling decisions; Large models assist in the training of analysis and decision-making models. They use crawler technology to obtain the required training data and combine format conversion and data cleaning to generate high-quality data sets that meet training requirements. The large-scale model-assisted analysis and decision-making model combines pre-training, supervised fine-tuning and human preference optimization methods to improve the analysis and decision-making capabilities of the large-scale model-assisted analysis and decision-making model for complex scenarios; The large model-assisted analysis and decision-making model adopts the LLaMA series as a benchmark and expands its capabilities through knowledge enhancement and domain adaptation technologies; The large-scale model-assisted analysis and decision-making model prioritizes emergencies based on global and local environmental information, and formulates hierarchical handling strategies to achieve dynamic alarm response efficiency.

6. The low-altitude UAV reconnaissance and disposal integrated protection system according to claim 1 is characterized in that: Also includes: Intrusion threat processing and response module, used to monitor controlled airspace in real time and collect data such as target flight trajectory and spectrum characteristics; Conduct preliminary threat identification of potential targets based on flight behavior, speed, and position parameters; When abnormal flight behavior or possible low-altitude intrusion is detected, the intrusion threat processing and response module immediately generates an alarm signal and notifies the dispatch center; After receiving the alarm signal, the dispatch center generates an auxiliary decision-making plan; For targets determined to be non-threatening, the system continuously tracks and dynamically updates their threat registration; For targets identified as high-threat, the system automatically initiates the emergency response process; The emergency response process specifies the optimal response plan based on target characteristics, image information, and current environment, combined with existing dispatchable resources; The intrusion threat processing and response module executes the disposal plan in real time and transmits key information to relevant management personnel; The relevant managers can view the target status, resource allocation and disposal progress in real time through the intelligent human-computer interaction interface; The emergency response process supports manual intervention and dynamic command, which can further optimize the corresponding process; It also includes a registration server for storing the RFID tag of each legally registered drone.

Images