Electronic fence intelligent monitoring method based on stress perception and variational uncertainty

Abstract

This invention provides an intelligent monitoring method for electronic fences based on stress perception and variational uncertainty. The method includes: deploying a multimodal sensor array to collect multimodal data and preprocessing it; performing multimodal data fusion optimization based on the preprocessed multimodal data to obtain global fusion features; determining the approximate range of parameters based on the global fusion features, locking local extrema, and generating an optimal parameter configuration through a reinforcement learning algorithm; and calculating the intrusion stress index and false alarm rate in real time based on the optimal parameter configuration, and optimizing and adjusting the physical threshold and calculation frequency of the electronic fence. This invention employs multi-fusion strategies such as cross-modal attention and dynamic projection fusion, adaptive ISI calculation and feature fusion, and conflict-aware gating fusion to deeply mine the complementary information of each modality's data, effectively resolving the heterogeneity and conflict problems between modalities. The generated global fusion features are more comprehensive and representative, improving the ability to characterize the features of the monitored targets.

Description

Technical Field

[0001] This invention belongs to the field of electronic fence technology, and in particular relates to an intelligent monitoring method for electronic fences based on stress perception and variational uncertainty. Background Technology

[0002] Electronic fence technology, as a core component of perimeter security systems, has been widely applied in scenarios such as military bases, government agencies, industrial sites, and wind farms. However, traditional electronic fence systems still have significant limitations in multi-sensor fusion, optimization algorithms, and intelligent decision-making.

[0003] Existing technologies primarily employ fixed weights or static rules to fuse multimodal data, which cannot adapt to environmental changes. For example, WO2024251941A1 uses location-tracking virtual fences but lacks integrated thermal infrared temperature anomaly detection, resulting in a false negative rate >20% in foggy conditions; US11218689B2 selectively fuses imagery equipment but lacks reinforcement learning dynamic optimization, leading to a response latency >0.3 seconds. Traditional fusion methods lack quantitative processing of intermodal conflicts; for instance, when different sensors produce different detection results for the same target, the system struggles to make accurate judgments.

[0004] Existing optimization methods are mostly based on single algorithms or centralized training, lacking dynamic collaborative mechanisms in distributed multi-agent reinforcement learning. For example, CN113704370A only uses NSGA-II for path planning, without combining local optimization (BAS) and dynamic reinforcement learning; while the PPO algorithm of PMC12041770 is robust, its single agent ignores modal collaboration. These algorithms struggle to quickly adapt and optimize parameters under different environmental conditions. Furthermore, existing technologies do not dynamically correlate thermal infrared point clouds with behavioral features, lacking multi-dimensional quantitative assessment of intrusion behavior; for example, the hotspot cloud fusion process of CN112686859A is not extended to thermal-distance stress quantification in the security domain. Existing systems lack mechanisms for quantifying and processing the uncertainty of sensor data, resulting in insufficient robustness in complex environments. Moreover, they often employ fixed thresholds or manual adjustment strategies, making it difficult to adapt to dynamic environmental changes. For example, while selective sensor fusion in US11218689B2 can reduce false alarm rates, it lacks a real-time feedback adjustment mechanism based on dynamic thresholds.

[0005] The limitations of existing technologies lie primarily in their failure to quantify intermodal conflicts and uncertainties; their singular optimization paths and lack of multi-level collaboration; and their failure to dynamically correlate thermal infrared point clouds with behavioral features, leading to high false alarm rates and slow response times in complex power system scenarios (such as strong winds and foggy weather in wind farms). Search results show that no patents combine ISI-quantified intrusion coercion, multimodal Transformer fusion, and variational uncertainty perception U-MAPPO distributed learning for electronic fences, and existing fusion formulas do not incorporate dynamic projection, behavioral attenuation, or V-GPR-based uncertainty perception. MAPPO is used for multi-agent applications in security systems, but no uncertainty perception variants for perimeter monitoring have been found, and there is a lack of ISI integration for power system scenarios. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent monitoring method for electronic fences based on stress perception and variational uncertainty.

[0007] Firstly, the method includes, A multimodal sensor array is deployed to collect multimodal data, and the collected multimodal data is preprocessed. Based on the preprocessed multimodal data, multimodal data fusion optimization is performed to obtain global fusion features; Based on global fusion features, the approximate range of parameters is determined, local extrema are locked, and the optimal parameter configuration is generated through reinforcement learning algorithms. Based on optimal parameter configuration, the intrusion threat index and false alarm rate are calculated in real time, and the physical threshold and calculation frequency of the electronic fence are adjusted in reverse to achieve a dynamic balance between resources and security of the electronic fence.

[0008] Furthermore, the arrangement of the multimodal sensor array specifically includes arranging the multimodal sensors using a hierarchical redundant coverage strategy.

[0009] Furthermore, the preprocessing of the collected multimodal data specifically includes optimizing the design of preprocessing parameters based on the noise characteristics and target size distribution of the wind farm scenario to achieve noise filtering, target enhancement, and geometric alignment of the original data; wherein, the preprocessing parameters include Gaussian denoising parameters, voxel clustering radius, dynamic sensitivity adjustment strategy, and unified coordinate transformation formula.

[0010] Furthermore, the preprocessing of the collected multimodal data also includes designing a timestamp-based synchronous acquisition mechanism to ensure the temporal consistency of the multimodal data; wherein, the timestamp-based synchronous acquisition mechanism includes a timestamp synchronization strategy, a data alignment algorithm, and a data integrity check.

[0011] Furthermore, the process of performing multimodal data fusion optimization to obtain global fusion features specifically includes: An improved 3D-2D dynamic projection formula is introduced, and thermal anomalies and velocity are used as dynamic parameters; Design a conflict-aware gating fusion mechanism, which dynamically adjusts the gating weights of each modality sensor by quantifying the cooperative similarity and conflict difference between sensors of each modality. By utilizing the adaptive intrusion stress index, a comprehensive stress quantification model of intrusion behavior is constructed by integrating multi-dimensional features such as heat, distance, speed, behavior, and environmental noise, enabling accurate assessment of intrusion risk levels. Furthermore, a dynamic feature fusion formula is designed based on the ISI index to achieve adaptive weight allocation of multi-modal features.

[0012] Furthermore, the process of determining the approximate range of parameters based on global fusion features, locking in local extrema, and generating the optimal parameter configuration through a reinforcement learning algorithm specifically includes: Offline training was performed using a non-dominated sorting genetic algorithm, with detection accuracy, response time, and resource consumption as mutually exclusive objectives, to generate a Pareto front solution set; Leveraging the lightweight nature of the BAS algorithm, a fast gradient search is performed near the Pareto front solution to lock in a local optimum working point, serving as a "hot start" state for reinforcement learning. Real-time parameter adjustment is achieved using variational uncertainty-aware distributed multi-agent reinforcement learning. The feature parameters, adjusted based on real-time parameters, are aggregated into the optimal parameter configuration.

[0013] Secondly, the system includes a data acquisition layer, a data fusion layer, a dynamic adjustment layer, and an iterative optimization layer; The data acquisition layer is used to deploy a multimodal sensor array, acquire multimodal data, and preprocess the acquired multimodal data; The data fusion layer is used to perform multimodal data fusion optimization based on the preprocessed multimodal data to obtain global fusion features; The dynamic adjustment layer is used to determine the approximate range of parameters based on global fusion features, lock local extrema, and generate the optimal parameter configuration through reinforcement learning algorithms. The iterative optimization layer is used to calculate the intrusion threat index and false alarm rate in real time based on the optimal parameter configuration, and to reversely optimize and adjust the physical threshold and calculation frequency of the electronic fence to achieve a dynamic balance between the resources and security of the electronic fence.

[0014] Furthermore, the data acquisition layer is specifically used to deploy multimodal sensors using a layered redundant coverage strategy.

[0015] Furthermore, the data acquisition layer is specifically used to optimize the design of preprocessing parameters based on the noise characteristics and target size distribution of the wind farm scenario, so as to achieve noise filtering, target enhancement and geometric alignment of the original data; wherein, the preprocessing parameters include Gaussian denoising parameters, voxel clustering radius, dynamic sensitivity adjustment strategy and unified coordinate transformation formula.

[0016] Furthermore, the data acquisition layer is also used to design a timestamp-based synchronous acquisition mechanism to ensure the temporal consistency of multimodal data; wherein, the timestamp-based synchronous acquisition mechanism includes a timestamp synchronization strategy, a data alignment algorithm, and a data integrity check.

[0017] Furthermore, the data fusion layer is specifically used for, An improved 3D-2D dynamic projection formula is introduced, and thermal anomalies and velocity are used as dynamic parameters; Design a conflict-aware gating fusion mechanism, which dynamically adjusts the gating weights of each modality sensor by quantifying the cooperative similarity and conflict difference between sensors of each modality. By utilizing the adaptive intrusion stress index, a comprehensive stress quantification model of intrusion behavior is constructed by integrating multi-dimensional features such as heat, distance, speed, behavior, and environmental noise, enabling accurate assessment of intrusion risk levels. Furthermore, a dynamic feature fusion formula is designed based on the ISI index to achieve adaptive weight allocation of multi-modal features.

[0018] Furthermore, the dynamic adjustment layer is specifically used for, Offline training was performed using a non-dominated sorting genetic algorithm, with detection accuracy, response time, and resource consumption as mutually exclusive objectives, to generate a Pareto front solution set; Leveraging the lightweight nature of the BAS algorithm, a fast gradient search is performed near the Pareto front solution to lock in a local optimum working point, serving as a "hot start" state for reinforcement learning. Real-time parameter adjustment is achieved using variational uncertainty-aware distributed multi-agent reinforcement learning. The feature parameters, adjusted based on real-time parameters, are aggregated into the optimal parameter configuration.

[0019] Thirdly, a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the steps of the above-described intelligent monitoring method for electronic fences based on stress perception and variational uncertainty.

[0020] Fourthly, an electronic device, characterized in that it includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements any of the above-described steps for intelligent monitoring of electronic fences based on stress perception and variational uncertainty.

[0021] Compared with the prior art, the present invention has the following advantages: 1. This invention proposes a multi-modal fusion strategy, including cross-modal attention and dynamic projection fusion, adaptive ISI calculation and feature fusion, and conflict-aware gating fusion, to deeply mine complementary information from different modal data, effectively resolve heterogeneity and conflict issues between modalities, and generate more comprehensive and representative global fusion features. Compared with single-modal analysis or simple fusion methods, it significantly improves the ability to characterize the features of the monitored targets.

[0022] 2. Through a collaborative mechanism of NSGA-II global optimization, beetle whisker search for local tuning, and variational uncertainty-aware distributed multi-agent reinforcement learning for dynamic adjustment, comprehensive and refined optimization of feature parameters is achieved. Global adjustment ensures the overall rationality of parameter configuration, local tuning improves the accuracy of parameters in key areas, and dynamic adjustment adapts to changes in the scenario and data fluctuations. The final optimal parameter configuration can accurately match monitoring requirements.

[0023] 3. Based on optimal parameter configuration, the system adaptively adjusts thresholds and frequencies, and achieves iterative optimization throughout the entire process through a feedback mechanism, enabling the system to continuously improve itself. Faced with dynamic scenarios such as target movement and environmental changes in wind farms, the system can adjust operating parameters in real time, maintaining high monitoring performance and effectively avoiding the problem of decreased monitoring accuracy due to changes in the environment.

[0024] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 The diagram shows a flowchart of an intelligent monitoring method for electronic fences based on stress perception and variational uncertainty according to the present invention.

[0027] Figure 2A schematic diagram of an intelligent monitoring method system for electronic fences based on stress perception and variational uncertainty is shown. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] like Figure 1 As shown, this invention proposes an intelligent monitoring method for electronic fences based on stress perception and variational uncertainty. By introducing ISI to quantify the multidimensional stress level of intrusion behavior, and combining it with U-MAPPO to achieve self-learning collaboration between modalities and noise-robust decision-making, a closed-loop mechanism of "perceiving uncertainty - quantifying stress level - dynamically adjusting strategy" is established. Variational inference enables the electronic fence to possess "self-awareness," and a dynamic velocity correction term is introduced in the fusion stage to solve the spatiotemporal misalignment problem. The steps include... S1. Deploy multimodal sensors, collect multimodal data, and preprocess the collected multimodal data.

[0030] In this embodiment, considering the terrain features and monitoring needs of complex dynamic environments such as wind farms, the present invention employs a layered redundant coverage strategy to deploy a multimodal sensor array. The sensor types include millimeter-wave radar, lidar, visible light cameras, and thermal infrared imagers, ensuring comprehensive coverage and complementary acquisition of multimodal data in key areas. Among them, millimeter-wave radar has strong penetration capabilities and can achieve stable detection of target distance and speed in adverse weather conditions such as fog and rain; lidar has ultra-high ranging accuracy and point cloud density, and can provide three-dimensional contour information of the target; visible light cameras can collect the texture and color features of the target for target identification; and thermal infrared imagers can capture the temperature differences of the target and are not limited by lighting conditions.

[0031] In this embodiment, to ensure coverage integrity in critical areas, the present invention defines coverage overlap rate ρ as the core evaluation index for sensor deployment; wherein, the formula for calculating coverage overlap rate is expressed as follows:

[0032] in, This refers to the coverage overlap rate; Let be the sensing area (m²) of sensor i. Let n be the sensing area of ​​sensor j; n is the number of sensors. This formula ensures that multimodal coverage of key areas (such as a perimeter radius of 50m) is ≥80%, which is different from the blind spot problem of existing grid layouts. Through this optimization, the electronic fence can adapt to changes in wind farm terrain and improve data integrity.

[0033] In this embodiment, to improve the quality and consistency of multimodal data, the present invention optimizes the design of preprocessing parameters based on the noise characteristics of wind farm scenarios (such as vibration noise caused by wind speed and electromagnetic interference) and target size distribution (such as the size differences of different intrusion targets such as personnel and vehicles). These parameters include Gaussian denoising parameters, voxel clustering radius, dynamic sensitivity adjustment strategy, and unified coordinate transformation formula. This achieves noise filtering, target enhancement, and geometric alignment of the original data, including: 1. Gaussian denoising parameters: In wind farm environments, sensor vibrations caused by wind speed can lead to random noise in the raw data, affecting the accuracy of target detection. Therefore, this invention has verified through a large number of experiments that when the standard deviation of Gaussian denoising σ=1.0, it can effectively filter vibration noise while preserving the detailed features of the target to the greatest extent, thus achieving a balance between noise filtering and texture preservation.

[0034] 2. Voxel Clustering Radar: The point cloud density of the LiDAR at a distance of 50m is ≥5000 points / m². The original point cloud data contains a large amount of background noise (such as ground, vegetation, etc.). This invention sets the voxel clustering radius r=0.01m. The voxel grid clustering algorithm groups point clouds with a distance less than the clustering radius into the same target, effectively removing background noise while retaining the point cloud details of small targets such as people. The core logic of voxel clustering is: dividing the three-dimensional space into cubic voxels with a side length of r, counting the number of point clouds in each voxel, and removing voxels with a number of point clouds less than a set threshold (set to 5 in this invention) to filter background noise.

[0035] 3. Dynamic Sensitivity Adjustment: The lighting conditions in wind farm environments change drastically (such as sunrise and sunset, cloud cover), which can affect the detection performance of visible light cameras and thermal infrared imagers. This invention designs an adaptive gain adjustment strategy based on the ambient light intensity L (unit: lux): when L < 100 (low light environment, such as nighttime or cloudy days), the sensor gain is increased by 20% to enhance the target signal strength; when L > 1000 (strong light environment, such as direct midday sunlight), the sensor gain is reduced by 0% to avoid image overexposure and loss of target features; when 100 ≤ L ≤ 1000 (normal lighting environment), the gain remains unchanged.

[0036] 4. Unified Coordinate Transformation: Due to differences in the installation positions and orientations of different sensors, the data they collect reside in different local coordinate systems. Coordinate transformation is necessary to achieve geometric consistency. This invention employs an affine transformation formula to uniformly transform the raw data from each sensor to the world coordinate system, expressed as follows:

[0037] in, The coordinates of the transformed point; for Rotation matrix (to correct installation angle); for Translation vector (corrects for positional deviation); The original point coordinates are given; this formula ensures that data from different sensors are fused in a unified coordinate system, providing geometric consistency for subsequent steps.

[0038] 5. Environmental noise baseline calculation: In order to measure the signal-to-noise ratio of the current environment, it is necessary to calculate the standard deviation of environmental noise within the sliding window in real time.

[0039]

[0040] in, The sampling window size (e.g., 50 frames); For the first The sensor background amplitude of the frame (such as radar clutter intensity or thermal imaging background average temperature). The mean within the window; This will be used as the denominator in subsequent ISI calculations; when the ambient noise is high, this value increases, and the ISI value decreases, thereby suppressing false alarms in high-noise environments.

[0041] In this embodiment, the data acquisition of multimodal sensors has timing differences. If synchronization processing is not performed, it will lead to timing misalignment of multimodal data, affecting fusion accuracy and decision-making timeliness. The present invention also designs a high-precision synchronous acquisition mechanism based on timestamps to ensure the timing consistency of multimodal data, specifically including: timestamp synchronization strategy, data alignment algorithm and data integrity check.

[0042] 1. Timestamp Synchronization Strategy: This invention adopts a globally unified timestamp benchmark. Each sensor synchronously records the current timestamp when collecting data, and the timestamp accuracy reaches the microsecond level. For example, a timestamp synchronization threshold Δt=0.05s is set, that is, the timestamp difference of the data collected by each sensor shall not exceed 0.05 seconds. If the threshold is exceeded, a data re-sampling mechanism is triggered to ensure the synchronization of the acquisition time of multimodal data.

[0043] 2. Data Alignment Algorithm: Since different sensors have different sampling frequencies (e.g., the sampling frequency of LiDAR is 10Hz, and the sampling frequency of visible light camera is 30Hz), a data alignment algorithm is needed to achieve time sequence matching. This invention adopts a timestamp-based linear interpolation algorithm, taking the sensor with the highest sampling frequency (e.g., visible light camera) as the benchmark, and interpolating the low-frequency data of other sensors to generate a data sequence that is consistent with the timing of the benchmark sensor.

[0044] 3. Data Integrity Check: This invention monitors the data transmission status of each sensor in real time and establishes a data integrity verification mechanism. Each frame of collected data is checked and calculated. If the checksum does not match or the data is missing, it is determined to be a data transmission anomaly, and a compensation mechanism is immediately triggered: For occasionally missing single frame data, the average of the previous frame and the next frame data is used to fill it; for more than 3 consecutive missing frames of data, a sensor fault warning is issued, and the weights of other sensors are automatically adjusted.

[0045] S2. Based on the preprocessed multimodal data, perform multimodal data fusion optimization to obtain global fusion features.

[0046] In this embodiment, the present invention achieves deep association, stress quantization and conflict resolution of multimodal features through three-level processing: cross-modal attention and dynamic projection fusion, adaptive ISI calculation and feature fusion, and conflict-aware gating fusion, thereby generating robust and highly recognizable global fusion features.

[0047] S2.1 Introducing 3D-2D dynamic lag projection.

[0048] In this embodiment, traditional projection only considers geometric relationships and ignores the positional lag caused by the integration time of the thermal imaging sensor. This invention introduces an improved 3D-2D dynamic projection formula and uses thermal anomaly ΔT and velocity v as dynamic parameters to achieve dynamic projection and weight allocation of multimodal features. This allows high-speed moving or thermally anomalous intruding targets to obtain higher fusion weights, solving the matching error between thermal infrared trailing and radar point clouds caused by high-speed movement. Specifically, this invention introduces a projection formula for velocity and thermal anomaly correction terms, expressed as follows:

[0049]

[0050]

[0051] in, : Projected coordinates of the image plane; : The three-dimensional spatial coordinates detected by radar; Camera internal references; : The velocity component of the target along the x-axis; : The inherent integration delay time of the thermal imaging sensor (constant, e.g., 0.04s); Thermo-dynamic coupling coefficient, when the target temperature difference... When the target is large and moves quickly, the thermal trailing effect is significant. This term is used to nonlinearly compensate for the center position of the predicted frame. This formula expresses the physical phenomenon that "the hotter and faster the target, the more obvious the positional lag in thermal imaging," and is further demonstrated through... The correction enables precise alignment of the thermal signatures of fast-moving targets with radar point clouds.

[0052] S2.2 Design a conflict-aware gating mechanism In this embodiment, during the acquisition and transmission of multimodal data, inconsistencies in detection results between modalities may occur due to factors such as sensor accuracy and environmental interference (e.g., lidar detects a target, but a visible light camera does not). Traditional fusion methods lack quantification processing for modal conflicts, resulting in insufficient robustness of the fusion results. This invention designs a conflict-aware gating fusion mechanism. By quantifying the cooperative similarity and conflict differences between sensors of different modalities, the gating weights of each modal sensor are dynamically adjusted to achieve adaptive conflict resolution, ensuring the accuracy and robustness of the fusion results. The steps include... a. Quantify the similarity of cooperation and the difference in conflict among different modalities. During the fusion process, it is necessary to quantize the sensors. and The "disagreement" between them; this invention proposes a dual constraint mechanism, in which the sensor and Conflict level The calculation formula is expressed as follows:

[0053] in, , These are the eigenvectors of mode i and mode j, respectively; , respectively, feature vectors , L2 norm; conflict degree The value range is [-1, 1]. The closer the value is to 1, the higher the consistency of the two modal features and the stronger the cooperation.

[0054] Optionally, this formula employs dual constraints to ensure that only features that "reach consensus" in both direction and intensity are fused with high weights, including: 1. Directional Conflict: Using Measurement. Scenario description: Radar detects a target moving to the left (vector). Optical flow method detected that the target moved to the right (vector). At this point, the included angle is close to 180 degrees, the cosine value is -1, and the conflict term reaches its maximum value of 2, which means that the sensor's perception of the motion state is completely opposite; 2. Intensity Conflict: Measured using normalized difference in modulus; for example, when radar detects a large metallic target (truck), the characteristic modulus... The area is large; however, thermal imaging shows that the temperature in this area is consistent with the background (possibly due to a cold vehicle or camouflage), and the feature magnitude is [missing information]. It is close to 0. At this point, although the direction may be randomly consistent due to noise, there is an essential contradiction in the signal strength.

[0055] b. Calculate the gating parameters.

[0056] In this embodiment, the present invention is based on the aforementioned conflict degree. and their respective signal-to-noise ratios Generate dynamic gating weights This is used to control the contribution of each modal feature in the fusion process, and its calculation formula is:

[0057] in, σ represents the gating weights for mode i and mode j, with values ​​ranging from [0,1]; σ is the sigmoid activation function, used to map the input values ​​to the interval [0,1]. The feature collaboration similarity between mode i and mode j is used to quantify the consistency of features between the two modes. is the feature conflict difference degree between mode i and mode j, used to quantify the inconsistency of features between the two modes.

[0058] Optionally, when its own signal-to-noise ratio The higher the value, the greater the dynamic gating weight. The larger the value, the greater the conflict with other modalities, and the weight decays exponentially.

[0059] S2.3 Calculation of Adaptive Intrusion Threat Index (ISI).

[0060] In this embodiment, traditional electronic fences often use single-dimensional intrusion judgment indicators (such as distance thresholds and speed thresholds), which are difficult to comprehensively quantify the degree of intrusion threat, leading to inaccurate judgment of low-speed, close-range but high-risk intrusion behaviors. This invention utilizes an Adaptive Intrusion Threat Index (ISI) to construct a comprehensive threat quantification model of intrusion behavior by integrating multi-dimensional features such as heat, distance, speed, behavior, and environmental noise. This enables accurate assessment of intrusion risk levels. Based on the ISI index, a dynamic feature fusion formula is designed to achieve adaptive weight allocation of multi-modal features. The Adaptive Intrusion Threat Index (ISI) calculation formula is expressed as follows:

[0061] in, The adaptive intrusion threat index represents the overall severity or risk level of an intrusion event; The target core temperature; This is the dry season reference temperature, representing the environmental reference temperature during the dry season, used for seasonal adjustments; This is the reference temperature for the warm season, representing the reference temperature for the wet season, used for seasonal adjustments; The intrusion distance (m) represents the distance between the intrusion target and the fence or detection point; The radius of the fence (m) represents the effective radius of the fence or detection area. This is the velocity coefficient (s / m), used to adjust the amplifying effect of velocity on threats. The velocity (m / s) represents the speed at which the intruding object moves. The noise figure (1 / stress unit) is used to adjust the effect of noise on ISI suppression. for

[0062] Alternatively, based on the Adaptive Intrusion Threat Index (ISI), the aforementioned "conflict-aware gating mechanism" can be used to calculate the conflict dissimilarity. With gate weights The aligned features are weighted and concatenated to generate a globally fused feature that includes heat, distance, velocity, and behavioral information. Its formula is expressed as follows:

[0063] in, The final fusion decision power represents the overall intrusion judgment score after multimodal feature fusion; For thermal features (normalized values), represent feature scores based on temperature anomalies; For distance features, represent feature scores based on distance; For velocity features, represent feature scores based on velocity; For behavioral features, the feature score represents the feature score based on the intrusion behavior pattern (such as intent matching); , The hot feature weights (unitless, ranging from 0 to 1) are the dominant weights, dynamically adjusted by ISI; ensuring that high-stress targets (high ISI) dominate feature fusion. The distance feature weights (range 0 to 1); For velocity feature weights (range 0 to 1), this With ISI The difference lies in the fact that they are independent weight symbols. The weights for behavioral features (ranging from 0 to 1).

[0064] The formula dynamically prioritizes hot and high-speed targets, and enhances cross-modal weight allocation through Transformer attention.

[0065] Optionally, in high-stress (high ISI) scenarios, when γ→1, α→0, β→0, δ→0, and thermal infrared features dominate the fusion result (which conforms to "high-stress target thermal infrared priority"). In low-stress (low ISI) scenarios, when γ→0, β=0.5×(1 γ)→0.5, with speed having the highest weight, and distance and behavior each accounting for 0.25 (which conforms to "prioritizing high-speed targets under low stress").

[0066] S3, Multi-objective collaborative optimization and uncertainty-aware decision-making In this embodiment, the present invention adopts a three-level cascaded optimization strategy of "offline global optimization - online local fine-tuning - real-time dynamic adaptation". Based on the global fusion features generated in step S2, the initial search space and local extreme points of the parameters are determined by a multi-objective optimization algorithm. The local extreme points are used as input to the state space, and the optimal parameter configuration is generated by a variational uncertainty-aware distributed multi-agent reinforcement learning algorithm.

[0067] S3.1 Global parameter space initialization based on NSGA-II In this embodiment, the present invention employs a non-dominated sorting genetic algorithm (NSGA-II) for offline training, using detection accuracy, response time, and resource consumption as mutually exclusive objectives to generate a Pareto front solution set, thereby defining the effective search boundary for parameters in subsequent steps. This includes... a. Initialize the population ( Includes sensitivity s, weight , wait).

[0068] b. Define the objective function as follows:

[0069] in, To improve detection accuracy (minimize the negative value to maximize accuracy); For electronic fence response time; To calculate resource utilization.

[0070] c. Calculate the crowding distance. In this embodiment, in order to maintain the diversity of solutions, the present invention calculates individual... Crowding distance , which is represented as,

[0071] in, The total number of objective functions (3 in this case); For individuals In the Neighboring individual values ​​in each target dimension; The first in the current population The maximum and minimum values ​​of each objective.

[0072] Optionally, this invention prioritizes selecting individuals with large crowding distances to prevent the algorithm from prematurely converging to a single solution, and outputs a Pareto front set. This provides a high-quality initial search point for the BAS algorithm.

[0073] S3.2. Based on longhorn beetle whisker search (BAS), perform local fine-tuning.

[0074] In this embodiment, the present invention utilizes the lightweight nature of the BAS algorithm to perform fast gradient search near the Pareto front solution, locking in a locally optimal operating point as a "warm start" state for reinforcement learning, which includes... a. from Selecting an initial solution .

[0075] b. Generate left and right coordinates And normalize the random direction vector, its representation is, ;

[0076] in, For the first Centroid of the next iteration; Interantennae spacing; A normalized random direction vector; It is a random vector.

[0077] c. Based on the selected initial solution, generate the first set of left and right coordinates. , which serves as the starting point for gradient search.

[0078] d. Based on the step size decay rule, perform position updates, where the position update formula is expressed as follows:

[0079] in: The current search step size, by attenuation( ); This is the overall fitness function; This is the sign function, indicating the direction of gradient descent.

[0080] e. Initialize the solution of NSGA-II Refine to As the initial weight configuration for U-MAPPO .

[0081] S3.3, U-MAPPO Dynamic Decision Based on Variational Uncertainty Awareness In this embodiment, to address sudden environmental changes (such as sudden heavy rain or strong interference), the present invention employs variational uncertainty-aware distributed multi-agent reinforcement learning (U-MAPPO) for real-time parameter adjustment, which includes: a. Variational uncertainty perception and dynamic adjustment of exploration rate.

[0082] Variational Gaussian process regression (V-GPR) is introduced to predict the "cognitive uncertainty" of the current state, and the trust region pruning coefficient of the PPO algorithm is adjusted accordingly. :

[0083] in: This is the base pruning threshold for the PPO algorithm; The variance (i.e. uncertainty) of the reward distribution predicted by the V-GPR network; Smoothing constant; Optionally, when When the environment is large (extremely unfamiliar), Automatically reduce and forcefully shrink the update amplitude of the strategy to prevent the model from catastrophic forgetting or parameter oscillation in unknown environments.

[0084] b. Max-Min pessimistic reward function.

[0085] Set state space (ISI index, speed, illumination, load), motion space To adjust the weights, construct a reward function, which includes: Let the state space be S: Action Space A: Weight adjustment, updating the formula for joint rewards;

[0086] in, The reward value is used to train the V-GPR and PPO networks; For intelligent agents Assigned sensor weights; This is the penalty coefficient; The rate of change of the invasion threat index (the reward increases when the target is suppressed).

[0087] In this embodiment, the present invention uses a Max-Min mechanism to ensure that, under the worst-case sensor collision scenario ( Even at its maximum, it can still obtain a guaranteed minimum return, enhancing robustness; it ensures that even in the worst-case scenario of sensor failure, a guaranteed minimum return can still be obtained.

[0088] c. Calculate the U-MAPPO loss function Based on dynamic adjustment The loss function of the policy network is calculated, and its formula is expressed as follows:

[0089] in, For strategy ratio; Generalized advantage estimation (GAE); Entropy term (coefficient) ), used to maintain a moderate level of exploration.

[0090] S3.4 Generating the optimal parameter configuration The feature parameters adjusted in steps S3.1-S3.3 are combined into an optimal parameter configuration.

[0091] The static benchmark output by BAS Dynamic adjustment of U-MAPPO output Combined, the final optimal parameter configuration is generated. This will be directly applied to the feature concatenation weights in step S2:

[0092] S4. Based on optimal parameter configuration, calculate the "Intrusion Threat Index (ISI)" and "False Alarm Rate" in real time, and reversely optimize and adjust the physical threshold and calculation frequency of the electronic fence to achieve a dynamic balance between resources and security. In this embodiment, step S4 is a "physical layer external closed loop" that is independent of the internal parameter update in step S3. While the electronic fence generates and distributes the optimal fusion weight configuration at the bottom layer using U-MAPPO, at the top physical control end, a dynamic threshold adaptive adjustment strategy is executed based on the real-time calculated Intrusion Threat Index (ISI) and historical false alarm rate.

[0093] S4.1 Dynamic threshold adaptive adjustment strategy: The alarm threshold is dynamically adjusted based on the current threat level (ISI) and historical false alarm data. This includes: a. Sliding window statistics: Update length is... Calculate the mean false alarm using a sliding window of frames. , which is represented as,

[0094] in, For the first The number of false positive targets in a frame; Calculate the standard deviation Threshold update formula:

[0095] in: The minimum safety threshold set; The intrusion threat index is calculated in real time for step S2.

[0096] Optionally, when When the threat increases, the denominator increases. Lowering the sensitivity of alarms makes them more sensitive; it's better to have a false alarm than a missed alarm. When historical false alarms ( As the concentration increases, the molecular size increases. If the alarm level is raised, the alarm becomes more robust, suppressing false alarms.

[0097] S4.2, Resource frequency and parameter update are adaptive.

[0098] In this embodiment, to balance real-time performance with limited computing power, the present invention adjusts the calculation frequency based on load and threat level; wherein, the update frequency formula is expressed as follows:

[0099] in, This represents the maximum hardware resource capacity. This represents the current resource utilization rate. The baseline update cycle is used.

[0100] Optionally, the higher the threat level (higher ISI) or the more idle the resources, the higher the update frequency, thus ensuring a faster response time during high-risk situations.

[0101] In this embodiment, when updating the model through backpropagation, the present invention employs gradient balancing to prevent gradient explosion and adjusts the learning rate weights of the agent according to ISI, which is expressed as follows:

[0102]

[0103] in: To prevent small amounts from being divided by zero; Base learning rate; This is the attenuation coefficient.

[0104] Alternatively, in a high-threat (high ISI) state, the learning rate weights can be reduced. To maintain relatively stable model parameters and avoid drastic model changes in critical moments.

[0105] like Figure 2 As shown, the present invention also proposes an intelligent monitoring and optimization system for electronic fences, which includes a data acquisition layer, a data fusion layer, a dynamic adjustment layer, and an iterative optimization layer.

[0106] 1. The data acquisition layer is used to deploy multimodal sensors, collect multimodal data, and preprocess the collected multimodal data.

[0107] In this embodiment, considering the terrain features and monitoring needs of complex dynamic environments such as wind farms, the data acquisition layer employs a layered redundant coverage strategy to deploy a multimodal sensor array. Sensor types include millimeter-wave radar, lidar, visible light cameras, and thermal infrared imagers, ensuring comprehensive coverage and complementary acquisition of multimodal data in key areas. Millimeter-wave radar possesses strong penetration capabilities, enabling stable detection of target distance and speed even in adverse weather conditions such as fog and rain. Lidar offers ultra-high ranging accuracy and point cloud density, providing three-dimensional contour information of the target. Visible light cameras can capture the texture and color features of the target for target identification. Thermal infrared imagers can capture temperature differences in the target, unaffected by lighting conditions.

[0108] In this embodiment, to ensure coverage integrity in critical areas, the data acquisition layer defines coverage overlap rate ρ as the core evaluation index for sensor deployment; the formula for calculating coverage overlap rate is expressed as follows:

[0109] in, This refers to the coverage overlap rate; Let be the sensing area (m²) of sensor i. Let n be the sensing area of ​​sensor j; n is the number of sensors. This formula ensures that multimodal coverage of key areas (such as a perimeter radius of 50m) is ≥80%, which is different from the blind spot problem of existing grid layouts. Through this optimization, the electronic fence can adapt to changes in wind farm terrain and improve data integrity.

[0110] In this embodiment, to improve the quality and consistency of multimodal data, the data acquisition layer optimizes preprocessing parameters based on the noise characteristics of the wind farm scenario (such as vibration noise caused by wind speed and electromagnetic interference) and the target size distribution (such as the size differences of different intrusion targets such as personnel and vehicles). These parameters include Gaussian denoising parameters, voxel clustering radius, dynamic sensitivity adjustment strategy, and unified coordinate transformation formula. This achieves noise filtering, target enhancement, and geometric alignment of the original data, including: (1) Gaussian noise reduction parameters: In the wind farm environment, the sensor vibration caused by wind speed will cause random noise in the original data, affecting the target detection accuracy. Therefore, through a large number of experiments, this invention has verified that when the standard deviation of Gaussian noise reduction σ=1.0, it can effectively filter vibration noise while retaining the detailed features of the target to the greatest extent, thus achieving a balance between noise filtering and texture preservation.

[0111] (2) Voxel clustering radius: The point cloud density of the lidar at a distance of 50m is ≥5000 / m², and the original point cloud data contains a lot of background noise (such as ground, vegetation, etc.). This invention sets the voxel clustering radius r=0.01m, and uses the voxel grid clustering algorithm to group point clouds with a distance less than the clustering radius into the same target, effectively removing background noise, while retaining the point cloud details of small targets such as people. The core logic of voxel clustering is: dividing the three-dimensional space into cubic voxels with a side length of r, counting the number of point clouds in each voxel, and removing voxels with a number of point clouds less than the set threshold (set to 5 in this invention) to achieve background noise filtering.

[0112] (3) Dynamic sensitivity adjustment: The lighting conditions in the wind farm environment change drastically (such as sunrise and sunset, cloud cover), which will affect the detection performance of the visible light camera and the thermal infrared imager. The present invention designs an adaptive gain adjustment strategy based on the ambient light intensity L (unit: lux): when L<100 (low light environment, such as night or cloudy day), the sensor gain is increased by 20% to enhance the target signal intensity; when L>1000 (strong light environment, such as direct sunlight at noon), the sensor gain is reduced by 0% to avoid image overexposure and loss of target features; when 100≤L≤1000 (normal light environment), the gain remains unchanged.

[0113] (4) Unified Coordinate Transformation: Due to differences in the installation positions and orientations of different sensors, the data they collect reside in different local coordinate systems, requiring coordinate transformation to achieve geometric consistency. This invention employs an affine transformation formula to uniformly transform the raw data from each sensor to the world coordinate system, expressed as follows:

[0114] in, The coordinates of the transformed point; for Rotation matrix (to correct installation angle); for Translation vector (corrects for positional deviation); The original point coordinates are given; this formula ensures that data from different sensors are fused in a unified coordinate system, providing geometric consistency for subsequent steps.

[0115] (5) Environmental noise baseline calculation: In order to measure the signal-to-noise ratio of the current environment, it is necessary to calculate the standard deviation of environmental noise within the sliding window in real time.

[0116]

[0117] in, The sampling window size (e.g., 50 frames); For the first The sensor background amplitude of the frame (such as radar clutter intensity or thermal imaging background average temperature). The mean within the window; This will be used as the denominator in subsequent ISI calculations; when the ambient noise is high, this value increases, and the ISI value decreases, thereby suppressing false alarms in high-noise environments.

[0118] In this embodiment, the data acquisition from the multimodal sensors exhibits temporal differences. Without synchronization, this can lead to temporal misalignment of the multimodal data, affecting fusion accuracy and decision-making timeliness. This invention also designs a high-precision synchronization acquisition mechanism based on timestamps to ensure the temporal consistency of the multimodal data. Specifically, this includes: a timestamp synchronization strategy, a data alignment algorithm, and data integrity checks. Timestamp synchronization strategy: This invention adopts a globally unified timestamp benchmark. Each sensor synchronously records the current timestamp when collecting data, and the timestamp accuracy reaches the microsecond level. For example, a timestamp synchronization threshold Δt=0.05s is set, that is, the timestamp difference of the data collected by each sensor shall not exceed 0.05 seconds. If the threshold is exceeded, a data re-sampling mechanism is triggered to ensure the synchronization of the acquisition time of multimodal data.

[0119] Data alignment algorithm: Since different sensors have different sampling frequencies (e.g., the sampling frequency of LiDAR is 10Hz and the sampling frequency of visible light camera is 30Hz), a data alignment algorithm is needed to achieve time sequence matching. This invention adopts a timestamp-based linear interpolation algorithm, which takes the sensor with the highest sampling frequency (e.g., visible light camera) as the benchmark, and interpolates and supplements the low-frequency data of other sensors to generate a data sequence that is consistent with the timing of the benchmark sensor.

[0120] Data integrity check: This invention monitors the data transmission status of each sensor in real time and establishes a data integrity verification mechanism; it performs verification and calculation on each frame of collected data. If the verification and calculation do not match or the data is missing, it is determined that the data transmission is abnormal and a compensation mechanism is immediately triggered: for occasionally missing single frame data, the average of the previous frame and the next frame data is used to fill the gap; for more than 3 consecutive missing frames of data, a sensor fault warning is issued and the weights of other sensors are automatically adjusted.

[0121] 2. The data fusion layer is used to perform multimodal data fusion optimization based on the preprocessed multimodal data to obtain global fusion features.

[0122] In this embodiment, the data fusion layer achieves deep correlation, stress quantization, and conflict resolution of multimodal features through three levels of processing: cross-modal attention and dynamic projection fusion, adaptive ISI calculation and feature fusion, and conflict-aware gating fusion. This generates robust and highly distinctive global fusion features, specifically including: (1) Introduce 3D-2D dynamic lag projection.

[0123] In this embodiment, traditional projection only considers geometric relationships and ignores the positional lag caused by the integration time of the thermal imaging sensor. The data fusion layer introduces an improved 3D-2D dynamic projection formula and uses thermal anomaly ΔT and velocity v as dynamic parameters to achieve dynamic projection and weight allocation of multimodal features. This allows high-speed moving or thermally anomalous intrusive targets to obtain higher fusion weights, solving the matching error between thermal infrared trailing and radar point clouds caused by high-speed movement. The projection formula introducing velocity and thermal anomaly correction terms is expressed as follows:

[0124]

[0125]

[0126] in, : Projected coordinates of the image plane; : The three-dimensional spatial coordinates detected by radar; Camera internal references; : The velocity component of the target along the x-axis; : The inherent integration delay time of the thermal imaging sensor (constant, e.g., 0.04s); Thermo-dynamic coupling coefficient, when the target temperature difference... When the target is large and moves quickly, the thermal trailing effect is significant. This term is used to nonlinearly compensate for the center position of the predicted frame. This formula expresses the physical phenomenon that "the hotter and faster the target, the more obvious the positional lag in thermal imaging," and is further demonstrated through... The correction enables precise alignment of the thermal signatures of fast-moving targets with radar point clouds.

[0127] (2) Design a conflict-aware gating mechanism In this embodiment, during the acquisition and transmission of multimodal data, inconsistencies in detection results between modalities may arise due to factors such as sensor accuracy and environmental interference (e.g., the lidar detects a target, but the visible light camera does not). Traditional fusion methods lack quantification processing for modal conflicts, resulting in insufficient robustness of the fusion results. The data fusion layer designs a conflict-aware gating fusion mechanism. By quantifying the cooperative similarity and conflict differences between sensors of different modalities, it dynamically adjusts the gating weights of each modal sensor to achieve adaptive conflict resolution, ensuring the accuracy and robustness of the fusion results. The steps include… a. Quantify the similarity of cooperation and the difference in conflict among different modalities. During the fusion process, it is necessary to quantize the sensors. and The "disagreement" between them; the data fusion layer proposes a dual constraint mechanism, in which sensors and Conflict level The calculation formula is expressed as follows:

[0128] in, , These are the eigenvectors of mode i and mode j, respectively; , respectively, feature vectors , L2 norm; conflict degree The value range is [-1, 1]. The closer the value is to 1, the higher the consistency of the two modal features and the stronger the cooperation.

[0129] Optionally, this formula employs dual constraints to ensure that only features that "reach consensus" in both direction and intensity are fused with high weights, including: Directional Conflict: Using Measurement. Scenario description: Radar detects a target moving to the left (vector). Optical flow method detected that the target moved to the right (vector). At this point, the included angle is close to 180 degrees, the cosine value is -1, and the conflict term reaches its maximum value of 2, which means that the sensor's perception of the motion state is completely opposite; Intensity conflict: Measured using normalized difference in modulus; for example, when radar detects a large metallic target (truck), the characteristic modulus... The area is large; however, thermal imaging shows that the temperature in this area is consistent with the background (possibly due to a cold vehicle or camouflage), and the feature magnitude is [missing information]. It is close to 0. At this point, although the direction may be randomly consistent due to noise, there is an essential contradiction in the signal strength.

[0130] b. Calculate the gating parameters.

[0131] In this embodiment, the data fusion layer is based on the aforementioned conflict level. and their respective signal-to-noise ratios Generate dynamic gating weights This is used to control the contribution of each modal feature in the fusion process, and its calculation formula is:

[0132] in, σ represents the gating weights for mode i and mode j, with values ​​ranging from [0,1]; σ is the sigmoid activation function, used to map the input values ​​to the interval [0,1]. The feature collaboration similarity between mode i and mode j is used to quantify the consistency of features between the two modes. is the feature conflict difference degree between mode i and mode j, used to quantify the inconsistency of features between the two modes.

[0133] Optionally, when its own signal-to-noise ratio The higher the value, the greater the dynamic gating weight. The larger the value, the greater the conflict with other modalities, and the weight decays exponentially.

[0134] (3) Calculation of Adaptive Intrusion Threat Index (ISI).

[0135] In this embodiment, traditional electronic fences often use single-dimensional intrusion judgment indicators (such as distance thresholds and speed thresholds), which are difficult to comprehensively quantify the degree of intrusion threat, leading to inaccurate judgment of low-speed, close-range but high-risk intrusion behaviors. The data fusion layer utilizes the Adaptive Intrusion Threat Index (ISI) to construct a comprehensive threat quantification model of intrusion behavior by integrating multi-dimensional features such as heat, distance, speed, behavior, and environmental noise, thereby achieving accurate assessment of intrusion risk levels. Based on the ISI index, a dynamic feature fusion formula is designed to achieve adaptive weight allocation of multi-modal features. The Adaptive Intrusion Threat Index (ISI) calculation formula is expressed as follows:

[0136] in, The adaptive intrusion threat index represents the overall severity or risk level of an intrusion event; The target core temperature; This is the dry season reference temperature, representing the environmental reference temperature during the dry season, used for seasonal adjustments; This is the reference temperature for the warm season, representing the reference temperature for the wet season, used for seasonal adjustments; The intrusion distance (m) represents the distance between the intrusion target and the fence or detection point; The radius of the fence (m) represents the effective radius of the fence or detection area. This is the velocity coefficient (s / m), used to adjust the amplifying effect of velocity on threats. The velocity (m / s) represents the speed at which the intruding object moves. The noise figure (1 / stress unit) is used to adjust the effect of noise on ISI suppression. for

[0137] Alternatively, based on the Adaptive Intrusion Threat Index (ISI), the aforementioned "conflict-aware gating mechanism" can be used to calculate the conflict dissimilarity. With gate weights The aligned features are weighted and concatenated to generate a globally fused feature that includes heat, distance, velocity, and behavioral information. Its formula is expressed as follows:

[0138] in, The final fusion decision power represents the overall intrusion judgment score after multimodal feature fusion; For thermal features (normalized values), represent feature scores based on temperature anomalies; For distance features, represent feature scores based on distance; For velocity features, represent feature scores based on velocity; For behavioral features, the feature score represents the feature score based on the intrusion behavior pattern (such as intent matching); , The hot feature weights (unitless, ranging from 0 to 1) are the dominant weights, dynamically adjusted by ISI; ensuring that high-stress targets (high ISI) dominate feature fusion. The distance feature weights (range 0 to 1); For velocity feature weights (range 0 to 1), this With ISI The difference lies in the fact that they are independent weight symbols. The weights for behavioral features (ranging from 0 to 1).

[0139] The formula dynamically prioritizes hot and high-speed targets, and enhances cross-modal weight allocation through Transformer attention.

[0140] Optionally, in high-stress (high ISI) scenarios, when γ→1, α→0, β→0, δ→0, and thermal infrared features dominate the fusion result (which conforms to "high-stress target thermal infrared priority"). In low-stress (low ISI) scenarios, when γ→0, β=0.5×(1 γ)→0.5, with speed having the highest weight, and distance and behavior each accounting for 0.25 (which conforms to "prioritizing high-speed targets under low stress").

[0141] 3. The dynamic adjustment layer is used for multi-objective collaborative optimization and uncertainty-aware decision-making. In this embodiment, the dynamic adjustment layer adopts a three-level cascaded optimization strategy of "offline global optimization - online local fine-tuning - real-time dynamic adaptation". Based on the global fusion features generated in step S2, the initial search space and local extreme points of the parameters are determined by a multi-objective optimization algorithm. The local extreme points are used as input to the state space, and the optimal parameter configuration is generated by a variational uncertainty-aware distributed multi-agent reinforcement learning algorithm.

[0142] (1) Global parameter space initialization based on NSGA-II In this embodiment, the present invention employs a non-dominated sorting genetic algorithm (NSGA-II) for offline training, using detection accuracy, response time, and resource consumption as mutually exclusive objectives to generate a Pareto front solution set, thereby defining the effective search boundary for parameters in subsequent steps. This includes... a. Initialize the population ( Includes sensitivity s, weight , wait).

[0143] b. Define the objective function as follows:

[0144] in, To improve detection accuracy (minimize the negative value to maximize accuracy); For electronic fence response time; To calculate resource utilization.

[0145] c. Calculate the crowding distance. In this embodiment, in order to maintain the diversity of solutions, the present invention calculates individual... Crowding distance , which is represented as,

[0146] in, The total number of objective functions (3 in this case); For individuals In the Neighboring individual values ​​in each target dimension; The first in the current population The maximum and minimum values ​​of each objective.

[0147] Optionally, this invention prioritizes selecting individuals with large crowding distances to prevent the algorithm from prematurely converging to a single solution, and outputs a Pareto front set. This provides a high-quality initial search point for the BAS algorithm.

[0148] (2) Based on the longhorn beetle whisker search (BAS), local fine-tuning is performed.

[0149] In this embodiment, the present invention utilizes the lightweight nature of the BAS algorithm to perform fast gradient search near the Pareto front solution, locking in a locally optimal operating point as a "warm start" state for reinforcement learning, which includes... a. from Selecting an initial solution .

[0150] b. Generate left and right coordinates And normalize the random direction vector, its representation is, ;

[0151] in, For the first Centroid of the next iteration; Interantennae spacing; A normalized random direction vector; It is a random vector.

[0152] c. Based on the selected initial solution, generate the first set of left and right coordinates. , which serves as the starting point for gradient search.

[0153] d. Based on the step size decay rule, perform position updates, where the position update formula is expressed as follows:

[0154] in: The current search step size, by attenuation( ); This is the overall fitness function; This is the sign function, indicating the direction of gradient descent.

[0155] e. Initialize the solution of NSGA-II Refine to As the initial weight configuration for U-MAPPO .

[0156] (3) U-MAPPO dynamic decision-making based on variational uncertainty perception In this embodiment, to address sudden environmental changes (such as sudden heavy rain or strong interference), the present invention employs variational uncertainty-aware distributed multi-agent reinforcement learning (U-MAPPO) for real-time parameter adjustment, which includes: a. Variational uncertainty perception and dynamic adjustment of exploration rate.

[0157] Variational Gaussian process regression (V-GPR) is introduced to predict the "cognitive uncertainty" of the current state, and the trust region pruning coefficient of the PPO algorithm is adjusted accordingly. :

[0158] in: This is the base pruning threshold for the PPO algorithm; The variance (i.e. uncertainty) of the reward distribution predicted by the V-GPR network; Smoothing constant; Optionally, when When the environment is large (extremely unfamiliar), Automatically reduce and forcefully shrink the update amplitude of the strategy to prevent the model from catastrophic forgetting or parameter oscillation in unknown environments.

[0159] b. Max-Min pessimistic reward function.

[0160] Set state space (ISI index, speed, illumination, load), motion space To adjust the weights, construct a reward function, which includes: Let the state space be S: Action Space A: Weight adjustment, updating the formula for joint rewards;

[0161] in, The reward value is used to train the V-GPR and PPO networks; For intelligent agents Assigned sensor weights; This is the penalty coefficient; The rate of change of the invasion threat index (the reward increases when the target is suppressed).

[0162] In this embodiment, the present invention uses a Max-Min mechanism to ensure that, under the worst-case sensor collision scenario ( Even at its maximum, it can still obtain a guaranteed minimum return, enhancing robustness; it ensures that even in the worst-case scenario of sensor failure, a guaranteed minimum return can still be obtained.

[0163] c. Calculate the U-MAPPO loss function Based on dynamic adjustment The loss function of the policy network is calculated, and its formula is expressed as follows:

[0164] in, For strategy ratio; Generalized advantage estimation (GAE); Entropy term (coefficient) ), used to maintain a moderate level of exploration.

[0165] (4) Generate the optimal parameter configuration The feature parameters adjusted in steps S3.1-S3.3 are combined into an optimal parameter configuration.

[0166] The static benchmark output by BAS Dynamic adjustment of U-MAPPO output Combined, the final optimal parameter configuration is generated. This will be directly applied to the feature concatenation weights in step S2:

[0167] 4. The iterative optimization layer calculates the "Intrusion Threat Index (ISI)" and "False Alarm Rate" in real time based on the optimal parameter configuration, and reversely optimizes and adjusts the physical threshold and calculation frequency of the electronic fence to achieve a dynamic balance between resources and security.

[0168] (1) Dynamic threshold adaptive adjustment strategy: The alarm threshold is dynamically adjusted based on the current threat level (ISI) and historical false alarm data. This includes: a. Sliding window statistics: Update length is... Calculate the mean false alarm using a sliding window of frames. , which is represented as,

[0169] in, For the first The number of false positive targets in a frame; Calculate the standard deviation Threshold update formula:

[0170] in: The minimum safety threshold set; The intrusion threat index is calculated in real time for step S2.

[0171] Optionally, when When the threat increases, the denominator increases. Lowering the sensitivity of alarms makes them more sensitive; it's better to have a false alarm than a missed alarm. When historical false alarms ( As the concentration increases, the molecular size increases. If the alarm level is raised, the alarm becomes more robust, suppressing false alarms.

[0172] (2) Resource frequency and parameter update are adaptive.

[0173] In this embodiment, to balance real-time performance with limited computing power, the present invention adjusts the calculation frequency based on load and threat level; wherein, the update frequency formula is expressed as follows:

[0174] in, This represents the maximum hardware resource capacity. This represents the current resource utilization rate. The baseline update cycle is used.

[0175] Optionally, the higher the threat level (higher ISI) or the more idle the resources, the higher the update frequency, thus ensuring a faster response time during high-risk situations.

[0176] In this embodiment, when updating the model through backpropagation, the present invention employs gradient balancing to prevent gradient explosion and adjusts the learning rate weights of the agent according to ISI, which is expressed as follows:

[0177]

[0178] in: To prevent small amounts from being divided by zero; Base learning rate; This is the attenuation coefficient.

[0179] Alternatively, in a high-threat (high ISI) state, the learning rate weights can be reduced. To maintain relatively stable model parameters and avoid drastic model changes in critical moments.

[0180] Based on the above disclosure, the present invention also provides an electronic device. The electronic device of this embodiment includes at least one processor and at least one storage medium electrically connected to the processor. The storage medium is electrically connected to the processor, wherein the storage medium stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0181] Based on the same inventive concept, the present invention also provides a storage medium storing instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described above.

[0182] The foregoing description and accompanying drawings fully illustrate embodiments of the invention to enable those skilled in the art to practice them. Other embodiments may include structural and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Some portions and features of some embodiments may be included or substituted for portions and features of other embodiments. Embodiments of the invention are not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from their scope. The scope of the invention is limited only by the appended claims.

Claims

1. An electronic fence intelligent monitoring method based on stress perception and variation uncertainty, characterized in that, The method includes, A multimodal sensor array is deployed to collect multimodal data, and the collected multimodal data is preprocessed. Based on the preprocessed multimodal data, multimodal data fusion optimization is performed to obtain global fusion features; Based on global fusion features, the approximate range of parameters is determined, local extrema are locked, and the optimal parameter configuration is generated through reinforcement learning algorithms. Based on optimal parameter configuration, the intrusion threat index and false alarm rate are calculated in real time, and the physical threshold and calculation frequency of the electronic fence are adjusted in reverse to achieve a dynamic balance between resources and security of the electronic fence. 2.The electronic fence intelligent monitoring method based on stress perception and variation uncertainty according to claim 1, characterized in that, The arrangement of the multimodal sensor array specifically includes arranging the multimodal sensors using a hierarchical redundancy coverage strategy. 3.The electronic fence intelligent monitoring method based on stress perception and variation uncertainty according to claim 1, characterized in that, The preprocessing of the collected multimodal data specifically includes optimizing the design of preprocessing parameters based on the noise characteristics and target size distribution of the wind farm scenario to achieve noise filtering, target enhancement, and geometric alignment of the original data; wherein the preprocessing parameters include Gaussian denoising parameters, voxel clustering radius, dynamic sensitivity adjustment strategy, and unified coordinate transformation formula.

4. The method of claim 1, wherein, The preprocessing of the collected multimodal data also includes designing a timestamp-based synchronous acquisition mechanism to ensure the temporal consistency of the multimodal data; wherein the timestamp-based synchronous acquisition mechanism includes a timestamp synchronization strategy, a data alignment algorithm, and a data integrity check.

5. The intelligent monitoring method for electronic fences based on stress perception and variational uncertainty according to claim 1, characterized in that, The process of performing multimodal data fusion optimization to obtain global fusion features specifically includes: An improved 3D-2D dynamic projection formula is introduced, and thermal anomalies and velocity are used as dynamic parameters; Design a conflict-aware gating fusion mechanism, which dynamically adjusts the gating weights of each modality sensor by quantifying the cooperative similarity and conflict difference between sensors of each modality. By utilizing the adaptive intrusion stress index, a comprehensive stress quantification model of intrusion behavior is constructed by integrating multi-dimensional features such as heat, distance, speed, behavior, and environmental noise, enabling accurate assessment of intrusion risk levels. Furthermore, a dynamic feature fusion formula is designed based on the ISI index to achieve adaptive weight allocation of multi-modal features.

6. The intelligent monitoring method for electronic fences based on stress perception and variational uncertainty according to claim 1, characterized in that, Based on global fusion features, the approximate range of parameters is determined, local extrema are identified, and an optimal parameter configuration is generated using a reinforcement learning algorithm. Specifically, this includes... Offline training was performed using a non-dominated sorting genetic algorithm, with detection accuracy, response time, and resource consumption as mutually exclusive objectives, to generate a Pareto front solution set; Leveraging the lightweight nature of the BAS algorithm, a fast gradient search is performed near the Pareto front solution to lock in a local optimum working point, serving as a "hot start" state for reinforcement learning. Real-time parameter adjustment is achieved using variational uncertainty-aware distributed multi-agent reinforcement learning. The feature parameters, adjusted based on real-time parameters, are aggregated into the optimal parameter configuration.

7. An intelligent monitoring system for electronic fences based on stress perception and variational uncertainty, characterized in that, The system includes a data acquisition layer, a data fusion layer, a dynamic adjustment layer, and an iterative optimization layer; The data acquisition layer is used to deploy a multimodal sensor array, acquire multimodal data, and preprocess the acquired multimodal data; The data fusion layer is used to perform multimodal data fusion optimization based on the preprocessed multimodal data to obtain global fusion features; The dynamic adjustment layer is used to determine the approximate range of parameters based on global fusion features, lock local extrema, and generate the optimal parameter configuration through reinforcement learning algorithms. The iterative optimization layer is used to calculate the intrusion threat index and false alarm rate in real time based on the optimal parameter configuration, and to reversely optimize and adjust the physical threshold and calculation frequency of the electronic fence to achieve a dynamic balance between the resources and security of the electronic fence.

8. The intelligent monitoring system for electronic fences based on stress perception and variational uncertainty according to claim 7, characterized in that, Specifically, the data acquisition layer is used to deploy multimodal sensors using a layered redundant coverage strategy.

9. The intelligent monitoring system for electronic fences based on stress perception and variational uncertainty according to claim 7, characterized in that, The data acquisition layer is specifically used to optimize the design of preprocessing parameters based on the noise characteristics and target size distribution of the wind farm scenario, so as to achieve noise filtering, target enhancement and geometric alignment of the original data; wherein, the preprocessing parameters include Gaussian denoising parameters, voxel clustering radius, dynamic sensitivity adjustment strategy and unified coordinate transformation formula.

10. The intelligent monitoring system for electronic fences based on stress perception and variational uncertainty according to claim 7, characterized in that, The data acquisition layer is also used to design a timestamp-based synchronous acquisition mechanism to ensure the temporal consistency of multimodal data; wherein the timestamp-based synchronous acquisition mechanism includes a timestamp synchronization strategy, a data alignment algorithm, and a data integrity check.

11. The intelligent monitoring system for electronic fences based on stress perception and variational uncertainty according to claim 7, characterized in that, The data fusion layer is specifically used for, An improved 3D-2D dynamic projection formula is introduced, and thermal anomalies and velocity are used as dynamic parameters; Design a conflict-aware gating fusion mechanism, which dynamically adjusts the gating weights of each modality sensor by quantifying the cooperative similarity and conflict difference between sensors of each modality. By utilizing the adaptive intrusion stress index, a comprehensive stress quantification model of intrusion behavior is constructed by integrating multi-dimensional features such as heat, distance, speed, behavior, and environmental noise, enabling accurate assessment of intrusion risk levels. Furthermore, a dynamic feature fusion formula is designed based on the ISI index to achieve adaptive weight allocation of multi-modal features.

12. The intelligent monitoring system for electronic fences based on stress perception and variational uncertainty according to claim 7, characterized in that, The dynamic adjustment layer is specifically used for, Offline training was performed using a non-dominated sorting genetic algorithm, with detection accuracy, response time, and resource consumption as mutually exclusive objectives, to generate a Pareto front solution set; Leveraging the lightweight nature of the BAS algorithm, a fast gradient search is performed near the Pareto front solution to lock in a local optimum working point, serving as a "hot start" state for reinforcement learning. Real-time parameter adjustment is achieved using variational uncertainty-aware distributed multi-agent reinforcement learning. The feature parameters, adjusted based on real-time parameters, are aggregated into the optimal parameter configuration.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the electronic fence intelligent monitoring method based on stress perception and variational uncertainty as described in any one of claims 1-6.

14. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the electronic fence intelligent monitoring steps based on stress perception and variational uncertainty as described in any one of claims 1-6.

Images