Trapping equipment cooperative control system based on multi-sensor fusion

Abstract

The invention relates to the technical field of trapping equipment control, and discloses a trapping equipment cooperative control system based on multi-sensor fusion, which comprises a multi-source data acquisition module, a target identification module, a risk assessment module, a cooperative relationship modeling module, a path optimization module, a decision fusion module and an execution control module. The multi-source data acquisition module generates an environment perception map containing temperature, humidity and illumination; the target recognition module analyzes the signals and generates data containing the target size and the motion category in a classified mode; the risk assessment module divides risk areas and inverts threats to generate a risk distribution map; the collaborative relation modeling module extracts collaborative parameters such as communication delay; the path optimization module dynamically plans a path; the decision fusion module performs weighted fusion on the data generation indexes; and the execution control module matches actions and adjusts parameters. The system realizes multi-dimensional perception, accurate identification and cooperative control, improves the trapping efficiency, and is suitable for pest control and other scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of trapping device control, in particular to a trapping device cooperative control system based on multi-sensor fusion. BACKGROUND

[0002] In the fields of pest control and wild animal management, the application of trapping devices is widespread and important. Traditional trapping devices mostly rely on a single sensor for environmental perception and target recognition, which has problems such as poor environmental adaptability, low target recognition accuracy, and insufficient device cooperation capability, making it difficult to cope with complex and variable trapping scenarios.

[0003] From the perspective of environmental perception, a single sensor can only obtain limited-dimensional environmental information. For example, although traditional infrared sensors can sense temperature changes, they cannot simultaneously obtain multi-dimensional environmental parameters such as humidity and light intensity, resulting in incomplete environmental perception graphs and difficulty in fully reflecting the actual environmental conditions of the trapping area. This makes it difficult for trapping devices to accurately regulate and control based on complete environmental information, such as in areas with high humidity or large changes in light intensity, where traditional devices may not be able to adjust trapping strategies in a timely manner, affecting trapping effectiveness.

[0004] In terms of target recognition, traditional methods often rely on simple signal features for target judgment, lacking in-depth spectral analysis of target reflection signals. For example, for different types of targets such as insects, rodents, and birds, traditional devices have difficulty accurately distinguishing their dynamic behavior characteristics and spectral differences, leading to misjudgment or missed judgment. In addition, the collection and analysis of key data such as target size parameters, motion categories, and occurrence frequency are not comprehensive enough in traditional target recognition processes, making it difficult to provide reliable basis for risk assessment and trapping strategy development.

[0005] In the risk assessment stage, traditional trapping devices often lack systematic collection and analysis of activity frequency data in the trapping area, making it difficult to scientifically divide risk level areas. At the same time, the ability to predict the evolution trend of potential threats is insufficient, and it is difficult to effectively invert combined with probability models, resulting in low accuracy and practicality of risk distribution graphs. This makes it difficult for trapping devices to focus on high-risk areas and prevent them in advance, reducing the efficiency and safety of trapping work.

[0006] In terms of device cooperation, traditional control systems lack real-time collection and in-depth analysis of trapping device state signals, making it difficult to effectively separate the intrinsic mode components of device interaction and extract the spatial correlation characteristics of components, resulting in inaccurate generation of cooperative characteristic parameters. This makes the modeling of the cooperative relationship between trapping devices inaccurate, making it difficult to achieve dynamic path optimization and resource allocation, leading to conflicts and insufficient coordination between devices, affecting overall trapping effectiveness.

[0007] In path planning, the traditional method often adopts a fixed path or a simple heuristic algorithm, which cannot be dynamically adjusted according to real-time coordination characteristic parameters and historical trapping data. For example, when facing a complex trapping area topology or a variable target distribution, the traditional path planning method is difficult to balance the constraints of path length, energy consumption, coverage range and other aspects, resulting in that the mobile path of the trapping device is not optimized and the trapping task cannot be efficiently completed.

[0008] In the decision fusion process, the traditional system has limited processing capacity for multi-source data and cannot effectively analyze the confidence weighting and the interaction relationship between parameters. Multi-source data is often in a scattered state, lacking an effective fusion mechanism, resulting in that the generated decision indicators are not scientific and comprehensive, and accurate guidance cannot be provided for execution control.

[0009] In the execution control link, the action sequence matching and adjustment of the traditional trapping device are not flexible, lacking real-time feedback mechanism and optimization learning ability. For example, during the execution of the trapping task, it is difficult to adjust the trapping intensity and the action range in time according to the real-time coverage effect detection data, and it is also difficult to optimize the execution efficiency through model predictive control, resulting in unstable trapping effect and inability to adapt to different trapping requirements in different scenes. SUMMARY

[0010] The purpose of the present application is to provide a trapping device cooperative control system based on multi-sensor fusion to solve the problems raised in the background art.

[0011] To achieve the above purpose, the present application provides the following technical scheme: a trapping device cooperative control system based on multi-sensor fusion, the system comprising: a multi-source data acquisition module, which performs multi-dimensional scanning on the trapping area based on environmental perception signals, extracts spatio-temporal features of the scanned data, and generates an environmental perception map; a target identification module, which performs frequency spectrum analysis on the target reflection signal based on the environmental perception map, identifies the target dynamic behavior characteristics, classifies the target types, and generates target identification data; a risk assessment module, which acquires activity frequency data of the trapping area based on the target identification data, divides the risk level area, combines the probability model to inverse the potential threat evolution trend, and generates a risk distribution map; a cooperative relationship modeling module, which acquires trapping device state signals based on the risk distribution map, separates the intrinsic mode components of device interaction, extracts the spatial correlation characteristics of the components, and generates coordination characteristic parameters; a path optimization module, which dynamically plans the mobile path of the trapping device based on the coordination characteristic parameters, combines historical trapping data to construct the cost and coverage target function, adopts a multi-constraint optimization strategy to balance the conflict conditions in path planning, and generates optimized path instructions; A decision fusion module, based on the environment perception atlas, target recognition data, risk distribution atlas, coordination characteristic parameters, and optimized path instructions, applies an evidence reasoning system to weight the confidence of multi-source data, determines the interaction between parameters through Pearson correlation analysis, and generates a fusion decision index; An execution control module, based on the fusion decision index, matches the action sequence of the trapping device, adjusts the trapping intensity and the action range in combination with a real-time feedback mechanism, optimizes the execution efficiency through model predictive control, and generates a control execution instruction.

[0012] Preferably, the environment perception atlas specifically includes regional temperature distribution, humidity gradient, and light intensity, the target recognition data includes target size parameters, motion category, and occurrence frequency, the risk distribution atlas specifically refers to activity hotspot density and threat propagation direction, the coordination characteristic parameters include inter-device communication delay and resource allocation weight, the optimized path instructions include device movement speed adjustment amount and trajectory complexity parameters, the fusion decision index includes risk response weight and coordination efficiency coefficient, and the control execution instruction specifically is a trapping intensity adaptive parameter and an action range dynamic threshold.

[0013] Preferably, the multi-source data acquisition module includes a signal acquisition sub-module, a data preprocessing sub-module, and a feature generation sub-module. The signal acquisition sub-module synchronously acquires environment perception signals based on multiple infrared sensors, generates a time-frequency feature map using a wavelet transform algorithm, eliminates signal drift errors through baseline correction, extracts time series data of key environmental parameters, and generates an initial perception matrix. The data preprocessing sub-module, based on the initial perception matrix, applies a principal component analysis algorithm to denoise and fuse different sensor data, calculates the covariance matrix between data, eliminates redundant interference in the acquisition process, and generates refined perception data. The feature generation sub-module, based on the refined perception data, calculates spatial feature distribution using a gradient boosting algorithm, divides abnormal perception areas through clustering analysis, fits a benchmark model using a polynomial regression and calculates feature deviation amounts, and generates an environment perception atlas.

[0014] Preferably, the target recognition module includes a frequency spectrum analysis sub-module, a type classification sub-module, and a feature quantization sub-module. The frequency spectrum analysis sub-module acquires target reflection signals using a millimeter wave radar, reconstructs signal frequency domain features through a Fourier transform algorithm, extracts frequency spectrum differences between the target and the background using a band-pass filter, and generates target contour information. The type classification submodule applies a K-means algorithm to segment the overlapping target region based on the target contour information, extracts the motion trajectory, speed and direction features of each independent region, constructs a multi-class classification model using a random forest, distinguishes insect, rodent and bird targets, and generates a target class label; The feature quantization submodule calculates the quantity proportion of each type of target based on the target class label using a region summation algorithm, combines a simulated annealing optimization weight distribution strategy, comprehensively evaluates the influence degree of the target on the trapping effect, and generates target identification data.

[0015] Preferably, the risk assessment module includes an activity monitoring submodule, a region division submodule, and a threat inversion submodule. The activity monitoring submodule scans the trapping region using an ultrasonic sensor array, converts the original sound wave data into an activity frequency value through a signal amplification algorithm, and generates a high-resolution activity distribution map. The region division submodule divides the high-risk core area and the diffusion area based on the activity distribution map using a dynamic threshold method, removes noise interference in the segmented region using a morphological opening operation, and generates risk level region boundary data. The threat inversion submodule constructs a Markov chain probability transition model based on the risk level region boundary data, iteratively solves the coupling relationship between activity frequency and threat using the Monte Carlo method, calculates the equivalent threat distribution through a risk propagation equation, and generates a risk distribution map.

[0016] Preferably, the collaborative relationship modeling module includes a state acquisition submodule, a mode decomposition submodule, and a correlation analysis submodule. The state acquisition submodule synchronously acquires node state signals of the trapping device based on a wireless sensor network, eliminates transmission delay interference through Kalman filtering, and generates multi-channel state time domain signals. The mode decomposition submodule applies singular value decomposition technology to decompose the signal into multiple intrinsic mode components based on the state time domain signal, removes false mode components through energy threshold screening, and retains effective components reflecting the real interaction of the device. The correlation analysis submodule calculates the spatial correlation degree and phase difference of each component based on the effective mode components using cross-correlation analysis, identifies the collaborative dependence characteristics between devices through coherence analysis, and generates collaborative characteristic parameters.

[0017] Preferably, the path optimization module includes a trajectory planning submodule, a constraint definition submodule, and a multi-objective solving submodule. The trajectory planning submodule generates an initial device movement path using an artificial potential field method based on the topological structure of the trapping region, dynamically adjusts the attractive force weight of the path nodes through a potential field gradient update mechanism, and generates a candidate path set. The constraint definition submodule builds an energy consumption model and a coverage prediction model based on historical trapping data, defines path length, turn frequency, and communication load as optimization variables, and generates multidimensional constraint conditions; The multi-objective solution submodule applies the particle swarm optimization algorithm to solve the Pareto frontier solution set based on the candidate path set and the constraint conditions, selects the path solution with the best comprehensive performance through the entropy weight method, and generates the optimized path instructions.

[0018] Preferably, the decision fusion module includes a data compression submodule, an interactive analysis submodule, and an indicator calculation submodule; The data compression submodule uses independent component analysis to extract key feature vectors based on multi-source heterogeneous data, screens the main independent components by feature contribution rate to reduce data dimensions, and generates a reduced-dimensional feature matrix; The interaction analysis submodule calculates the interaction strength between parameters based on the dimensionality reduction feature matrix and applies the Spearman rank correlation algorithm to construct a parameter dependency network diagram to identify the core parameter group that affects the trapping effect; The indicator calculation submodule is based on the parameter dependency network diagram, uses a logistic regression model to quantify the decision probability of each parameter, generates a comprehensive evaluation index through probability weighted fusion, and generates a fusion decision index.

[0019] Preferably, the execution control module includes a sequence generation submodule, a feedback adjustment submodule, and an optimization learning submodule; The sequence generation submodule uses a tabu search algorithm to search for the optimal action combination of the trapping device based on the fusion decision index, and balances the global exploration and local development capabilities through the tabu length strategy to generate the initial action sequence; The feedback adjustment submodule estimates the execution deviation based on the coverage effect detection data during the execution process using a proportional-integral-derivative controller, and dynamically adjusts the correction coefficients of the trapping intensity and range parameters; The optimization learning submodule constructs a deep Q network model based on historical control data, iteratively updates the action value function through the policy gradient algorithm, optimizes the adaptive adjustment rules of the execution parameters, and generates control execution instructions.

[0020] Preferably, the K-means algorithm adopts a multi-scale distance measurement strategy, calculates the similarity characteristics of the target region through Mahalanobis distance, combines density clustering to enhance regional separability, and generates independent target region boundaries.

[0021] Compared with the prior art, the present invention has the following beneficial effects: In terms of environmental perception, the multi-source data acquisition module synchronously acquires environmental perception signals based on multiple infrared sensors, and combines wavelet transform, principal component analysis, and gradient boosting algorithm, etc. to perform multi-dimensional scanning on the trapping area, extract spatio-temporal features, and generate an environmental perception spectrum containing information such as regional temperature distribution, humidity gradient, and illumination intensity. This enables the system to comprehensively and accurately perceive the environmental conditions of the trapping area, providing rich and reliable basic data for subsequent target identification, risk assessment, etc., effectively solving the problem of incomplete environmental perception by traditional single sensors.

[0022] The target identification module uses millimeter wave radar to collect target reflection signals, and through Fourier transform, K-means algorithm, and random forest technology, it can analyze the frequency spectrum of the target reflection signals, identify the dynamic behavior characteristics of the targets, accurately distinguish between insect, rodent, and bird target types, and generate target identification data containing target size parameters, motion categories, and occurrence frequency. This greatly improves the accuracy and efficiency of target identification, reduces misjudgment and omission, and provides strong support for scientifically formulating trapping strategies.

[0023] The risk assessment module uses an ultrasonic sensor array to scan the trapping area, and combines dynamic threshold method, Markov chain probability transfer model, and Monte Carlo method, etc. to collect activity frequency data of the trapping area, divide risk level areas, invert potential threat evolution trends, and generate a risk distribution spectrum containing activity hotspot density and threat propagation direction. This module enables the system to timely discover high-risk areas, deploy and prevent in advance, effectively improving the targeting and safety of trapping work, and reducing the impact of potential threats.

[0024] The collaborative relationship modeling module synchronously collects trapping device state signals based on a wireless sensor network, and through singular value decomposition, cross-correlation analysis, and coherence analysis, etc. it can separate the intrinsic mode components of device interaction, extract spatial correlation features of the components, and generate collaborative characteristic parameters containing communication delay and resource allocation weight between devices. This makes it possible to accurately model the collaborative relationship between trapping devices, enabling devices to adjust collaboratively based on real-time states, improving the collaborative efficiency and coordination between devices, and avoiding the problem of insufficient device collaboration in traditional systems.

[0025] The path optimization module is based on collaborative characteristic parameters, and combines artificial potential field method, particle swarm optimization algorithm, and entropy weight method, etc. to dynamically plan the moving path of the trapping device, balance the conflict conditions in path planning, and generate optimization path instructions containing device moving speed adjustment and trajectory complexity parameters. This module enables the trapping device to select the optimal moving path in a complex trapping area, effectively improving the moving efficiency and coverage range of the device, reducing energy consumption, and improving the overall trapping effect.

[0026] The decision fusion module applies evidence reasoning system, Pearson correlation analysis, independent component analysis and Spearman rank correlation algorithm, etc., can perform confidence weighting and interaction relationship analysis on multi-source data such as environmental perception atlas and target recognition data, and generate fusion decision indicators including risk response weight and synergy efficiency coefficient, etc. This enables the system to make scientific decisions by integrating various information, improves the scientificity and comprehensiveness of the decision, and provides accurate guidance for execution control.

[0027] The execution control module can match the action sequence of the trapping device, adjust the trapping intensity and action range in real time, optimize the execution efficiency, and generate control execution instructions such as trapping intensity adaptive parameters and action range dynamic threshold based on the fusion decision indicators, combined with tabu search algorithm, proportional-integral-derivative controller and deep Q network model, etc. This module enables the trapping device to flexibly adjust the action according to the actual situation, improves the accuracy and adaptability of the execution, and ensures the efficient completion of the trapping task. BRIEF DESCRIPTION OF DRAWINGS

[0028] Fig. 1 The working principle diagram of the multi-sensor fusion-based trapping device cooperative control system described in the application; Fig. 2 The working principle diagram of the multi-source data acquisition module; Fig. 3 The working principle diagram of the target recognition module. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0030] Please refer to Figs. 1-3 The application relates to a multi-sensor fusion-based trapping device cooperative control system, which comprises a multi-source data acquisition module, a target recognition module, a risk assessment module, a synergy relationship modeling module, a path optimization module, a decision fusion module and an execution control module, and each module is cooperatively operated according to the following steps: The multi-source data acquisition module performs multi-dimensional scanning on the trapping area through multiple groups of sensors based on environmental perception signals, extracts the space-time features of the scanning data, and generates an environmental perception atlas containing regional temperature distribution, humidity gradient and illumination intensity.

[0031] The target recognition module performs spectral analysis on the target reflection signal based on the environmental perception map, identifies the target dynamic behavior characteristics and classifies the target type, and generates target recognition data including target size parameters, motion category and occurrence frequency.

[0032] The risk assessment module collects activity frequency data in the trapping area based on target identification data, divides the risk level areas, and combines the probability model to invert the evolution trend of potential threats to generate a risk distribution map reflecting the density of activity hotspots and the direction of threat propagation.

[0033] The collaborative relationship modeling module collects and captures device status signals based on the risk distribution map, separates the intrinsic mode components of device interactions, extracts the spatial correlation features of the components, and generates collaborative characteristic parameters including communication delays and resource allocation weights between devices.

[0034] The path optimization module dynamically plans the moving path of the trapping device based on the collaborative characteristic parameters, constructs the cost and coverage objective functions based on the historical trapping data, adopts a multi-constraint optimization strategy to balance the conflicting conditions in the path planning, and generates the optimized path instructions including the device movement speed adjustment and trajectory complexity parameters.

[0035] The decision fusion module is based on environmental perception maps, target recognition data, risk distribution maps, collaborative characteristic parameters, and optimized path instructions. It uses the evidence reasoning system to confidence-weight multi-source data, determines the interaction between parameters through Pearson correlation analysis, and generates a fusion decision indicator that includes risk response weights and collaborative efficiency coefficients.

[0036] The execution control module matches the action sequence of the trapping device based on the fusion decision indicators, adjusts the trapping intensity and range of action in combination with the real-time feedback mechanism, optimizes the execution efficiency through model predictive control, and generates control execution instructions that include adaptive parameters of trapping intensity and dynamic thresholds of range of action.

[0037] The present invention will be further described below in conjunction with Examples 1 to 5: Example

[0038] The output data of each module in the system has a clear technical definition: the environmental perception map specifically includes regional temperature distribution, humidity gradient, and illumination intensity, representing the basic environmental state of the trapping area through the spatial distribution of multi-dimensional environmental parameters; target recognition data includes target size parameters, motion categories, and occurrence frequency, used to describe the physical properties and behavior characteristics of the target; the risk distribution map specifically refers to the activity hotspot density and threat propagation direction, reflecting the spatial aggregation degree and potential threat diffusion trend of the target activity; the coordination characteristic parameters include communication delay and resource allocation weight between devices, used to describe the interaction state and resource allocation relationship between the trapping device nodes; the optimization path instruction includes device movement speed adjustment and trajectory complexity parameters, guiding the dynamic movement strategy of the device; the fusion decision index includes risk response weight and coordination efficiency coefficient, generating decision basis by integrating multi-source data; the control execution instruction specifically includes trapping intensity adaptive parameter and action range dynamic threshold, directly driving the execution action of the trapping device.

[0039] The multi-source data acquisition module, as the data foundation layer of the system, is composed of three sub-modules: signal acquisition sub-module, data preprocessing sub-module, and feature generation sub-module. Each sub-module realizes the acquisition, optimization, and feature extraction of environmental perception signals through specific algorithms and processing procedures.

[0040] The signal acquisition sub-module constructs a distributed perception network through multiple infrared sensors, realizing multi-dimensional synchronous scanning of the trapping area. The infrared sensor is based on the principle of thermal radiation and can collect temperature field distribution, humidity gradient change, and illumination intensity data in the area in real time. Each group of sensors is uniformly deployed according to the preset spatial interval, forming a three-dimensional coverage of the target area. During the acquisition process, the sensor converts the physical environment signal into an electrical signal and sends it to the processing unit through a wired or wireless transmission link. At this stage, the wavelet transform algorithm is used for time-frequency analysis of the original signal, converting the time-domain signal into a time-frequency feature map. Through the decomposition of different scale wavelet basis functions, the separation of high-frequency noise and low-frequency trend in the signal is realized. At the same time, the baseline correction technique is applied to eliminate the signal drift error caused by long-term work or environmental temperature and humidity changes of the sensor. Specifically, the signal baseline is estimated and corrected through polynomial fitting or moving average method, ensuring that the time series data of the extracted environmental key parameters (such as temperature, humidity, and illumination intensity) have high stability and accuracy. After the above processing, the initial perception matrix containing multi-sensor spatio-temporal sampling data is generated. The rows in the matrix represent different sensor nodes, the columns represent sampling time points, and the element values are the environmental parameter measurement values of the corresponding nodes at that time.

[0041] The data preprocessing submodule adopts principal component analysis (PCA) algorithm to reduce noise and fuse data in view of noise interference and data redundancy in the initial perception matrix. First, the covariance matrix between different sensor data is calculated to reflect the linear correlation degree between variables. Through eigenvalue decomposition, the principal components of the covariance matrix, i.e. orthogonal eigenvectors with larger variance contribution, are extracted. These principal components can retain the main information of the original data in fewer dimensions. According to the preset cumulative variance contribution rate threshold (such as 95%), the key principal components are selected, and the original high-dimensional data is projected into the low-dimensional principal component space to realize data dimension reduction. In the dimension reduction process, the measurement deviation and redundant information between different sensors are eliminated by reconstructing the data to generate refined perception data. For example, for three highly correlated parameters of temperature, humidity and light intensity, principal component analysis can combine them into one or two principal components, which not only retains the main trend of environmental characteristics, but also reduces the complexity of subsequent calculation.

[0042] The feature generation submodule generates an environmental perception map based on the refined perception data by combining gradient boosting algorithm, clustering analysis and polynomial regression. Gradient boosting algorithm constructs multiple weak learners (such as decision trees) through iteration to gradually fit the spatial feature distribution of the data, which can effectively capture the nonlinear variation trend of temperature, humidity and other parameters in space. For example, in a forest environment, gradient boosting algorithm can identify the boundary features of the low-temperature and high-humidity area under the tree shade and the high-temperature and low-humidity area in the open land. Clustering analysis uses algorithms such as density-based spatial clustering of applications with noise (DBSCAN) to divide the perception anomaly area (such as temperature sudden change point, humidity mutation zone) and normal area according to the spatial density distribution of data points, forming different clustering clusters. At the same time, the baseline distribution of environmental parameters is fitted by combining the polynomial regression model, for example, the variation trend of light intensity with altitude in the region is fitted by a quadratic polynomial, and the deviation between the actual measurement value and the predicted value of the baseline model is calculated. The size of the deviation reflects the abnormality degree of the environmental parameter at the point. Finally, the spatial feature distribution obtained by gradient boosting algorithm, the abnormal area division result obtained by clustering analysis and the deviation information obtained by polynomial regression are fused to generate an environmental perception map containing contour lines, abnormal area labels and deviation color scales. The map presents the environmental state of the trapping area in a visual form, providing basic data support for subsequent target recognition and risk assessment.

[0043] In the implementation process, the infrared sensor of the signal acquisition submodule needs to meet the requirements of wide temperature working range, high sensitivity, and strong anti-interference ability, such as using a composite sensor unit integrating a pyroelectric infrared sensor (PIR) and a humidity sensor. The principal component analysis calculation of the data preprocessing submodule can be implemented on an embedded processor or edge computing node through a matrix operation library (such as LAPACK), ensuring real-time processing requirements. The gradient boosting algorithm of the feature generation submodule can use the XGBoost or LightGBM framework, accelerating the model training process through parallel computing, and clustering analysis and polynomial regression are implemented through a numerical calculation library (such as NumPy). The processing flow of the entire multi-source data acquisition module forms a closed loop, and standardized algorithms and parameter configurations are used in each link from signal acquisition to atlas generation to ensure data consistency and repeatability. Embodiments

[0044] This embodiment relates to the specific structure and workflow of the target recognition module, including the technical implementation of the spectrum analysis, type classification, and feature quantization submodules, as well as the improved application of the K-means algorithm in target region segmentation.

[0045] As the core processing layer of the system, the target recognition module undertakes the key task of analyzing target information from the environmental perception atlas, and its output target recognition data includes target size parameters, motion categories, and occurrence frequencies, which are realized through the following three-level submodules: The spectrum analysis submodule uses a millimeter wave radar as the core perception device to achieve high-precision detection of targets using the short wavelength characteristics of millimeter waves (30GHz-300GHz frequency band, corresponding to a wavelength of 10mm-1mm). The millimeter wave radar transmits a wideband signal and receives the target reflection echo, and through the Fourier transform algorithm, the time-domain echo signal is converted into a frequency-domain feature atlas. This process is implemented through a fast Fourier transform (FFT) processor, which can accurately extract the frequency component, amplitude, and phase information of the signal. Due to the differences in material and structure between the target and the background (such as vegetation and terrain), the frequency spectrum characteristics of their reflection signals are significantly different. By setting a specific frequency window through a bandpass filter, environmental clutter can be effectively removed, and the effective frequency band of the target reflection signal can be retained. For example, the micro-Doppler effect caused by the wing vibration of an insect target will form a specific frequency component in the frequency spectrum, which can be separated through bandpass filtering to generate target contour information, including target radial size, shape factor, and other parameters. The processing flow of the spectrum analysis submodule needs to ensure that the signal sampling rate meets the Nyquist criterion, which is usually achieved through a hardware trigger synchronization mechanism to match the timing of the radar transmission and reception signals, ensuring the accuracy of the frequency spectrum characteristics.

[0046] The type classification submodule first performs overlapping target region segmentation using the K-means algorithm based on the target contour information output by the spectrum analysis submodule. The traditional K-means algorithm uses Euclidean distance to measure sample similarity, but in complex environments, target feature vectors (such as motion trajectory, speed, and direction) can have strong correlations, resulting in insufficient segmentation accuracy. To this end, the present embodiment introduces a multi-scale distance measurement strategy and uses Mahalanobis distance to calculate the similarity features of the target region. Mahalanobis distance takes into account the covariance structure of the feature vector, which can eliminate the influence of correlation between features and achieve scale normalization. Specifically, by calculating the weighted distance of the inverse matrix of the covariance matrix of the sample point and the cluster center, the clustering effect on high-dimensional correlated data is improved. At the same time, combining the density clustering idea, a density threshold is introduced in the K-means clustering process to count the density of sample points in each cluster, eliminate noise points in low-density regions, and enhance the separability of overlapping regions. For example, in a multi-target (such as coexistence of insect groups and rodents) scene, Mahalanobis distance can effectively distinguish the feature distribution of different target types, and density clustering can further eliminate the fuzzy boundaries between targets to generate clear independent target region boundaries.

[0047] After completing the region segmentation, the submodule extracts the motion trajectory, speed, and direction features of each independent region. The motion trajectory is represented by a sequence of target position coordinates at consecutive time points, the speed is the ratio of the position change amount and the time interval between adjacent time points, and the direction is determined by the tangent direction of the trajectory. A multi-class classification model is constructed using the random forest algorithm, which consists of multiple decision trees. Each decision tree randomly selects a sample subset from the training data based on the.bootstrap sampling method and randomly selects a feature subset for optimal split attribute calculation when splitting nodes. By voting on the classification results of multiple decision trees, accurate classification of insect, rodent, and bird targets is achieved, and a target class label is generated. During training, the cross-validation method is used to evaluate the model's generalization ability, and the classification performance is optimized by adjusting hyperparameters such as the number of decision trees and the maximum depth.

[0048] The feature quantification sub-module quantifies the spatial distribution of each target class based on the target class label using a region summation algorithm. This algorithm iterates through the grid cells of the trapping region, counts the number of different target classes within each cell, and calculates the proportion of each target class and the density distribution. For example, in a farmland scene, the number of insects and rodents per square meter can be counted to generate a target density heat map. To comprehensively evaluate the impact of targets on trapping effectiveness, a simulated annealing optimization strategy is introduced for weight distribution. The simulated annealing algorithm simulates the physical annealing process to find the optimal weight combination in the solution space. In the initial stage, a higher temperature (control parameter) allows for larger weight adjustments, and as the temperature decreases, it gradually converges to a local optimal solution. Weight distribution needs to consider factors such as the ecological harm level, activity frequency, and trapping difficulty of targets. For example, insect targets that spread diseases are given higher weights, while occasional bird targets are given lower weights. A comprehensive impact index is generated by weighted summation, which, together with target size parameters and frequency information, forms the target recognition data to provide quantitative basis for subsequent risk assessment and decision fusion.

[0049] In terms of hardware implementation, the millimeter wave radar of the spectrum analysis sub-module can use an integrated chip solution (such as the TII WR series), which integrates the transmitting and receiving antennas and signal processing units, supporting multi-channel synchronous acquisition. The K-means algorithm and random forest model of the type classification sub-module can run on edge computing nodes (such as the NVIDIA Jetson series), utilizing GPU acceleration to improve computational efficiency. The region summation algorithm of the feature quantification sub-module is implemented through a parallel computing framework (such as OpenMP) to ensure real-time processing capability under large-scale data.

[0050] The processing flow of the entire target recognition module embodies a multi-level information processing architecture from the signal level to the feature level and then to the decision level: the spectrum analysis sub-module completes feature extraction of physical signals, the type classification sub-module realizes semantic understanding of target classes, and the feature quantification sub-module provides quantitative indicators required for decision-making. Each sub-module transmits information through standardized data interfaces, for example, the target contour information output by the spectrum analysis sub-module is stored in the form of a multi-dimensional feature vector, the type classification sub-module reads this vector and outputs the class label, and the feature quantification sub-module generates the final recognition result by combining the label and the original perception data. This hierarchical design not only improves the scalability of the module, but also enhances the adaptability of the system to complex target scenarios through algorithm optimization (such as Mahalanobis distance metric, simulated annealing weight distribution), ensuring stable target recognition and feature quantification under conditions of multiple species coexistence and strong environmental interference. Embodiments

[0051] The embodiment relates to the specific structure and working process of a risk assessment module, and includes a technical implementation of an activity monitoring, region division and threat inversion sub-module, and a threat evolution trend analysis method based on a probability model.

[0052] The risk assessment module is a threat analysis layer of the system, and a core function thereof is to generate a graph reflecting a risk distribution of a trapping region based on target identification data through multi-stage processing, specifically including a three-stage processing flow of an activity monitoring sub-module, a region division sub-module and a threat inversion sub-module, and each sub-module realizes quantitative evaluation of a risk level through specific sensor configuration and algorithm combination.

[0053] The activity monitoring sub-module adopts an ultrasonic sensor array to construct a space scanning network, so as to realize high-resolution monitoring of target activities in the trapping region. The ultrasonic sensor emits high-frequency sound waves (20 kHz-200 kHz) and receives reflected echoes, calculates the target distance based on the time of flight (ToF) principle, and forms a fan-shaped or omnidirectional scanning coverage through the cooperative work of multiple sensors in the array. The original sound wave data includes environmental noise (such as wind noise, plant swing) and effective signals generated by target activities, and the signal-to-noise ratio of the effective signal is improved through a signal amplification algorithm. Specifically, the echo signal is amplified in amplitude through a multi-stage amplifier circuit, and low-frequency environmental noise is removed through a high-pass filter. The processed signal is converted into a digital signal through an analog-to-digital conversion (ADC), and target activity events (such as echo amplitude mutations caused by target movement) are extracted through a threshold detection algorithm. The number of activity events in a unit of time is counted to generate an activity frequency value, which reflects the frequency of target activities in a unit area. Through grid sampling (such as dividing the trapping region into 1m x 1m grid cells), an activity frequency value is assigned to each grid cell to generate a high-resolution activity distribution map containing spatial coordinates and activity intensity. The pixel gray value or color depth in the map corresponds to the high and low of the activity frequency.

[0054] The region division sub-module divides the risk level region based on the activity distribution map by combining a dynamic threshold method with morphological processing. The dynamic threshold method adaptively determines the risk level boundary value according to the statistical characteristics of the activity frequency, specifically by calculating the mean value and the standard deviation of the activity frequency, setting the threshold value of the high-risk core area as (K is an empirical coefficient, usually 1-3), the medium-risk diffusion area as to , and the low-risk area as lower than The method avoids the problem of fixed threshold being not adaptive to different scenes, for example, in the high and low peak periods of farmland pest, the risk level division standard can be automatically adjusted. The divided regions may have isolated noise points (such as single high activity frequency outliers) or narrow spike regions, which are optimized by morphological opening operation: first, the erosion operation (remove the pixel points of the region boundary) is used to eliminate isolated noise points, and then the dilation operation (fill the small holes in the region) is used to smooth the boundary, to generate continuous and regular risk level region boundary data, including the high-risk core area polygon contour and the buffer range of the medium-risk diffusion area.

[0055] The threat inversion submodule is based on the risk level region boundary data, and through the construction of Markov chain probability transfer model and Monte Carlo simulation, the evolution trend of potential threat is inverted. The Markov chain model assumes that the target activity state only depends on the current state and has nothing to do with the historical state, and the state transfer relationship between different risk level regions is described by a transition probability matrix. Assuming that the trapping region is divided into n risk level regions (such as high-risk core area, medium-risk diffusion area, and low-risk area), the elements in the transition probability matrix p represent the probability of the target transferring from region i to region j, and the matrix is obtained by statistical analysis of historical activity frequency data. For example, if it is found by statistics that 60% of the targets in the high-risk core area stay in the original region, 30% spread to the medium-risk area, and 10% enter the low-risk area, then the elements of the corresponding row in the transition probability matrix are . .

[0056] The Monte Carlo method solves the coupling relationship between activity frequency and threat by simulating a large number of random paths. For each target individual, the next risk region is randomly selected according to the Markov chain transition probability, and the simulation is repeated N times (N is a large sample number, such as 100,000 times), and the threat cumulative value of each region is counted. The calculation of the threat cumulative value uses the risk propagation equation:

[0057] where S is the threat cumulative value of the target in the time period T, is the activity frequency value of the region at time t, ​​The threat weight coefficient of the corresponding area (high-risk area weight is higher than that of medium-risk area and low-risk area). By statistically averaging the threat cumulative value of all simulation paths, the equivalent threat distribution of each area is obtained, which reflects the spatial diffusion pattern of potential threats that may be triggered by target activities. Combined with the risk level area boundary data and the equivalent threat distribution, a risk distribution map containing activity hotspot density (represented by activity frequency value) and threat propagation direction (represented by Markov chain transition probability dominant direction) is generated. In the map, the main direction of threat propagation is marked by arrow symbols, and the spatial distribution of threat level is represented by color gradient. In terms of hardware implementation, the ultrasonic sensor array of the activity monitoring submodule adopts time division multiplexing technology to avoid channel interference. Each sensor transmits and receives sound wave signals in a predetermined time sequence, and the time sequence control and data acquisition are realized through a microcontroller (such as STM32 series). The dynamic threshold calculation and morphological processing of the area division submodule are realized on an embedded processor through an image processing library (such as OpenCV), which supports real-time updating of risk area division results. The Markov chain modeling and Monte Carlo simulation of the threat inversion submodule are realized through a numerical calculation library (such as NumPy), which utilizes parallel computing technology to accelerate the simulation process and ensure the completion of threat trend prediction within a reasonable time.

[0058] The processing flow of the entire risk assessment module embodies the closed-loop logic from data acquisition to model inversion: the activity monitoring submodule provides real-time activity data, the area division submodule realizes spatial discretization of risk level, and the threat inversion submodule reveals the statistical law of threat evolution through probability model. The technical selection of each submodule is closely related to the actual needs of the trapping scene, for example, the non-visual perception characteristics of ultrasonic sensors are suitable for complex lighting conditions (such as night or vegetation sheltered environment), dynamic threshold method and morphological processing adapt to changes in target activity mode at different times, and Markov chain and Monte Carlo simulation provide theoretical support for threat prediction.

[0059] Embodiment 4: This embodiment relates to the specific structure and workflow of the coordination relationship modeling module and the path optimization module, and describes the technical implementation details in combination with typical application scenarios (such as trapping pests in farmland).

[0060] The synergy modeling module is used to analyze the interaction characteristics between the trapping devices, which is composed of a state acquisition submodule, a mode decomposition submodule, and a correlation analysis submodule. The state acquisition submodule constructs a device communication network based on a wireless sensor network (such as Zigbee or LoRa). Each trapping device (such as an insecticidal lamp or a trap) is equipped with a microcontroller and a sensor node to collect device state signals in real time, including power supply capacity, communication signal strength, actuator working state (such as the opening angle of the trapping net), etc. When the signals are transmitted to the central processing unit through the wireless link, they may be affected by multipath fading, co-frequency interference, etc., resulting in transmission delay. Therefore, Kalman filtering algorithm is used to reduce noise of the time-domain signals. Taking the device power signal as an example, Kalman filtering establishes a state space model, fuses the predicted value and the measured value, and estimates the power state closer to the true value to generate a smoothed multi-channel state time-domain signal.

[0061] The mode decomposition submodule uses singular value decomposition (SVD) technology to extract features from the state time-domain signal. Singular value decomposition can decompose a multi-dimensional signal matrix into the product of a left singular matrix, a singular value matrix, and a right singular matrix. The singular values are arranged in descending order, reflecting the energy proportion of different intrinsic mode components in the signal. By setting an energy threshold (such as retaining components with a cumulative energy proportion of 90%), effective components reflecting the real interaction of the device are selected, and false mode components caused by environmental interference or accidental fluctuations of the device are removed. For example, in the scenario of multiple insecticidal lamps working together, the communication delay signals between devices may contain periodic fluctuations (such as synchronization errors caused by power grid voltage fluctuations) and random noise. Singular value decomposition can separate the intrinsic mode components corresponding to periodic fluctuations (high-energy singular values) from random noise components (low-energy singular values), and retain the former for subsequent analysis.

[0062] The correlation analysis submodule calculates the spatial correlation and phase difference of each component based on the effective mode components using cross-correlation analysis. The cross-correlation function describes the similarity of two signals at different time delays. By calculating the cross-correlation coefficient of the state components of device A and device B, the spatial correlation between the two devices can be quantified: a coefficient close to 1 indicates strong positive correlation (e.g., device B adjusts the direction synchronously when device A moves), a coefficient close to -1 indicates strong negative correlation (e.g., device B reduces resource occupation when device A increases trapping intensity), and a coefficient close to 0 indicates no significant correlation. Phase difference analysis is used to identify the synergy dependence characteristics between devices. For example, if there is a fixed phase difference between the trapping period signals of two traps, it may indicate that they have achieved alternate work through some mechanism to avoid resource conflict. Through the above analysis, synergy characteristic parameters including communication delay (such as signal transmission time) and resource allocation weight (such as power consumption priority) between devices are generated, providing a basis for device interaction state for path optimization.

[0063] The path optimization module plans the moving path of the trapping device based on the synergy characteristic parameters, and is composed of a trajectory planning submodule, a constraint definition submodule, and a multi-objective solving submodule. The trajectory planning submodule first constructs a topological structure of the trapping area, abstracts a farmland scene as a two-dimensional grid graph containing obstacles (such as crop rows and irrigation facilities) and target areas (such as high-density distribution areas of pests), and generates an initial moving path by using an artificial potential field method: the target area is set as an "attractive force source", the obstacles are set as "repulsive force sources", and each trapping device is subjected to the combined force of the attractive force and the repulsive force in the virtual potential field and moves along the direction of the combined force. The attractive force weight of the path node is dynamically adjusted through a potential field gradient updating mechanism, for example, when multiple devices simultaneously tend to the same target area, the attractive force weight of the subsequent device is reduced to avoid clustering, and a candidate path set containing multiple feasible paths is generated.

[0064] The constraint definition submodule constructs an energy consumption model and a coverage range prediction model based on historical trapping data. The energy consumption model statistically analyzes the relationship between the moving speed, the turning frequency and the power consumption of the device, for example, high-speed movement or frequent turning can cause the power consumption to increase by 30%-50%; the coverage range prediction model estimates the trapping coverage area of the device at different positions by using an interpolation algorithm, for example, the effective coverage radius of a pesticide lamp is positively correlated with its installation height. The path length, the turning frequency and the communication load are defined as optimization variables: the path length affects the time for the device to reach the target, the turning frequency affects the wear of mechanical parts, and the communication load affects the network bandwidth occupation. By setting constraint conditions (such as the path length being not more than 50 meters, the turning frequency being not more than 2 times per 10 meters, and the communication load being less than 80% of the network bandwidth), the path optimization problem is converted into an optimal solution search problem under multiple constraint conditions.

[0065] The multi-objective solving submodule solves the Pareto front solution set by using a particle swarm optimization (PSO) algorithm. Each particle in the particle swarm represents a candidate path, and the position vector corresponds to the path node coordinates and the speed vector represents the path adjustment direction. The algorithm iteratively updates the particle position to search for non-dominated solutions (Pareto solutions) that simultaneously satisfy the shortest path length, the lowest energy consumption and the highest coverage efficiency in the solution space. For example, in a farmland scene, a certain path may have a shorter path length but need to pass through a dense obstacle area (high energy consumption), and another path may have a longer length but move along the ridge (low energy consumption), both of which are Pareto solutions. The schemes in the Pareto solution set are comprehensively evaluated by using an entropy weight method, which objectively allocates weights according to the information entropy of each optimization variable. The smaller the information entropy (the higher the variable dispersion degree), the greater the weight. For example, if the power shortage of the device is the main problem in the historical data, the weight of the energy consumption variable is higher, and the path scheme with the lowest energy consumption is finally selected to generate the optimization path instruction containing the device moving speed adjustment amount (such as from 2 m / s to 1.5 m / s to save power) and the trajectory complexity parameter (such as the path bending index).

[0066] In practical applications, the cooperative relationship modeling module and the path optimization module need to interact data in real time: the cooperative characteristic parameters (such as the increase of the communication delay between devices) may trigger the path optimization module to re-plan the path to avoid communication interruption; the path optimization result (such as the movement of a device to a new location) will change the spatial layout between devices, thereby affecting the update of the cooperative characteristic parameters. For example, when the communication delay between two traps leads to a decrease in the cooperative efficiency, the path optimization module may adjust the movement path of one of the two traps to shorten the distance between them to reduce the delay, while ensuring that the path adjustment does not exceed the energy consumption threshold through the constraint definition sub-module.

[0067] In terms of hardware implementation, the wireless sensor network of the cooperative relationship modeling module adopts the low-power wide-area network (LPWAN) technology to adapt to the signal coverage requirements of large-scale scenes such as farmland; the singular value decomposition and cross-correlation analysis are implemented through the floating-point operation unit of the edge computing node (such as Raspberry Pi). The artificial potential field method and particle swarm optimization algorithm of the path optimization module are implemented through high-performance programming languages such as C++, and parallel computing is used to accelerate the evaluation of candidate paths. Embodiments

[0068] This embodiment relates to the specific structure and workflow of the decision fusion module and the execution control module, and describes the technical implementation details in combination with the mouse trapping scene in the warehouse environment.

[0069] As the high-level decision layer of the system, the decision fusion module is responsible for integrating multi-source heterogeneous data to generate comprehensive decision indicators, and is composed of a data compression sub-module, an interaction analysis sub-module, and an indicator calculation sub-module. The data compression sub-module processes multi-source data from the environment perception map, target recognition data, risk distribution map, etc. These data may contain tens of dimensions of features such as temperature, humidity, target size, activity frequency, etc., and there is strong correlation and redundancy. Independent component analysis (ICA) algorithm is used to extract key feature vectors, which assumes that multi-source data is a linear combination of independent hidden components, and separates independent principal components by maximizing non-Gaussianity. For example, in the warehouse environment, temperature and humidity data are usually highly correlated, and independent component analysis can separate them into independent components reflecting the comprehensive state of environmental temperature and humidity and local fluctuation components, and the top 3-5 principal components (cumulative contribution rate exceeds 85%) are selected by feature contribution rate to generate a reduced dimension feature matrix, reducing the subsequent calculation amount while retaining the core information.

[0070] The interaction analysis submodule analyzes the non-linear interaction between parameters based on the reduced dimension feature matrix using the Spearman rank correlation algorithm. Taking the rodent trapping scenario as an example, the correlation between the target appearance frequency and the risk area activity density, the device communication delay and the path complexity is analyzed: if the Spearman correlation coefficient is 0.7, it indicates that there is a strong positive correlation between the two (e.g. the higher the activity density, the higher the target appearance frequency); if the coefficient is -0.5, it indicates a negative correlation (e.g. as the communication delay increases, the path planning complexity decreases). By constructing a parameter dependency network graph, nodes represent parameters (such as "risk response weight" and "synergy efficiency coefficient"), and edges represent the strength of correlation (represented by line thickness or color depth), the core parameter group affecting the trapping effect is identified, such as the strong correlation parameter chain formed by "risk area activity density-target size-trapping intensity", which provides key reference for decision indicator calculation.

[0071] The indicator calculation submodule quantifies the contribution probability of each parameter to the decision result based on the parameter dependency network graph using the logistic regression model. The logistic regression model converts the linear combination to a probability value between 0 and 1 through the sigmoid function, for example, inputting risk response weight, synergy efficiency coefficient and other parameters, outputting the decision probability of "starting multi-device collaborative trapping". Through probability weighted fusion, the decision probabilities of each parameter are weighted and summed according to their importance (determined by the weight of the Spearman correlation coefficient or expert experience), generating a fusion decision indicator containing risk response weight and synergy efficiency coefficient. For example, in a high-risk rodent activity area, the risk response weight accounts for 60%, the synergy efficiency coefficient accounts for 40%, and when the comprehensive indicator exceeds the threshold, a high-intensity trapping strategy is triggered.

[0072] The execution control module serves as the execution layer of the system, generating specific control instructions based on the fusion decision indicator, composed of a sequence generation submodule, a feedback adjustment submodule, and an optimization learning submodule. The sequence generation submodule uses the tabu search algorithm to search for the optimal action combination of the trapping device. The tabu search algorithm avoids repeated search of visited solutions by maintaining a "tabu list", for example, in the case of trapping rodents between warehouse shelves, device actions include moving to point A, starting ultrasonic repelling, releasing bait, etc. The algorithm searches different action combinations through iteration, evaluates their impact on trapping effectiveness (such as covering the target area and trapping delay time), balances global exploration and local development ability through tabu length strategy (such as prohibiting repeated access to the last 5-step action combination), and generates an initial action sequence, such as "move to the corner of the shelf → release grain bait → delay 10 minutes and start the trapping net".

[0073] The feedback adjustment submodule dynamically adjusts the trapping parameters based on the coverage effect detection data (such as the number of target entries detected by the infrared sensor) during execution, using a proportional-integral-derivative (PID) controller. For example, if the target capture rate is lower than expected at the preset trapping intensity, the PID controller calculates the execution deviation (the difference between the actual capture rate and the target value), quickly responds to the deviation through the proportional term, eliminates steady-state error through the integral term, and predicts the deviation trend through the derivative term, adjusting the correction coefficient of the trapping intensity parameter (such as increasing the bait release amount by 20%) and the dynamic threshold of the action range (such as expanding the ultrasonic repelling range to 3 meters around). This process operates in real-time closed loop, updating control parameters every 5 minutes to ensure that the trapping strategy adapts to changes in target behavior.

[0074] The optimization learning submodule builds a deep Q network (DQN) model based on historical regulation data to improve the system's adaptive ability. Deep Q network combines deep learning and reinforcement learning, taking the action sequence of the trapping device as input and the trapping effect evaluation value (such as the capture amount per unit time) as the reward signal, and iteratively updating the action value function through the policy gradient algorithm. For example, in a warehouse scenario, the system records the capture amount corresponding to different action combinations (such as "move + bait + repel" and "fixed position + double bait"), and the model learns the optimization rule "prefer to use the move + bait strategy in high-risk areas" through training. As historical data accumulates, the model gradually converges to the optimal strategy, generating control execution instructions containing trapping intensity adaptive parameters (such as dynamically adjusting bait concentration according to target density) and action range dynamic thresholds (such as expanding the light trapping range at night), achieving autonomous optimization of the trapping strategy.

[0075] In the specific application of warehouse mouse trapping, the collaborative process of the decision fusion module and the execution control module is as follows: when the data compression submodule detects that the temperature in a certain area abnormally rises (possibly due to the gathering activities of mice), the target recognition data shows that the frequency of rodent targets appears 5 times / hour, and the risk distribution map shows that this area is an active hotspot and threatens to spread to the warehouse entrance, the interactive analysis submodule identifies the strong association of "temperature-target frequency-risk diffusion", the index calculation submodule generates the fusion decision index (risk response weight 0.7, synergy efficiency coefficient 0.6), and triggers the execution control module to start the emergency trapping scheme. The sequence generation submodule generates the action sequence according to the tabu search algorithm: 3 trapping devices are moved to the three entrances of the hotspot area, and high-concentration baits are released synchronously and low-frequency ultrasonic interference is started (to hinder mouse communication); the feedback adjustment submodule detects through the infrared sensor that the target activity frequency in the device coverage range decreases by 40% within 2 hours, but the capture target has not been reached, so the bait is added with pheromone attractant (adjusting the trapping intensity parameter) through the PID controller, and the ultrasonic interference range is narrowed to 1 meter around the bait (focusing on the action range); the optimization learning submodule records this time of regulation data, updates the deep Q network model, so that the system preferentially adopts the combination strategy of "bait + pheromone + local interference" in subsequent similar scenarios.

[0076] In terms of hardware implementation, the independent component analysis and Spearman correlation analysis of the decision fusion module are run on the edge server through the Python scientific computing library (such as Scikit-learn), supporting batch data processing; the tabu search algorithm and PID control of the execution control module are implemented through the PLC (programmable logic controller), ensuring real-time control accuracy; the deep Q network model is deployed on the cloud server, and model updating is realized through the timing synchronization of historical data. The whole module combination realizes the intelligent generation and dynamic adjustment of trapping strategies through the three-level mechanism of data fusion, real-time feedback and learning optimization, improving the trapping efficiency and resource utilization rate in complex environments.

[0077] It should be noted that, in this text, relationship terms such as first and second are only used to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0078] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A trapping equipment collaborative control system based on multi-sensor fusion, characterized in that: The system comprises: The multi-source data acquisition module performs multi-dimensional scanning of the trapping area based on environmental perception signals, extracts spatiotemporal features from the scanned data, and generates an environmental perception map. The target recognition module, based on the environmental perception map, performs spectrum analysis on the target reflection signal, identifies the target dynamic behavior characteristics, classifies the target type, and generates target recognition data; The risk assessment module, based on target identification data, should collect activity frequency data in the trapping area, divide the risk level areas, and use the probability model to invert the evolution trend of potential threats to generate a risk distribution map; The collaborative relationship modeling module collects the state signals of the trapping devices based on the risk distribution map, separates the intrinsic mode components of the device interactions, extracts the spatial correlation features of the components, and generates collaborative characteristic parameters; The path optimization module dynamically plans the movement path of the trapping device based on the collaborative characteristic parameters, constructs the cost and coverage objective functions based on the historical trapping data, and uses a multi-constraint optimization strategy to balance the conflicting conditions in the path planning and generate the optimized path instructions; The decision fusion module uses the evidence reasoning system to perform confidence weighting on multi-source data based on environmental perception maps, target recognition data, risk distribution maps, collaborative characteristic parameters, and optimized path instructions. It uses Pearson correlation analysis to determine the interaction between parameters and generate fusion decision indicators. The execution control module matches the action sequence of the trapping device based on the fusion decision-making indicators, adjusts the trapping intensity and scope of action in combination with the real-time feedback mechanism, optimizes the execution efficiency through model predictive control, and generates control execution instructions.

2. The trapping device collaborative control system based on multi-sensor fusion according to claim 1 is characterized by: The environmental perception map specifically includes regional temperature distribution, humidity gradient and light intensity; the target identification data includes target size parameters, motion category and occurrence frequency; the risk distribution map specifically refers to activity hotspot density and threat propagation direction; the collaborative characteristic parameters include inter-device communication delay and resource allocation weight; the optimized path instruction includes device movement speed adjustment amount and trajectory complexity parameter; the fusion decision indicator includes risk response weight and collaborative efficiency coefficient; the control execution instruction specifically includes trapping intensity adaptive parameter and action range dynamic threshold.

3. The trapping device collaborative control system based on multi-sensor fusion according to claim 1 is characterized by: The multi-source data acquisition module includes a signal acquisition submodule, a data preprocessing submodule, and a feature generation submodule; The signal acquisition submodule synchronously acquires environmental perception signals based on multiple groups of infrared sensors, generates a time-frequency feature map using a wavelet transform algorithm, eliminates signal drift errors through baseline correction, extracts time series data of key environmental parameters, and generates an initial perception matrix; The data preprocessing submodule uses the principal component analysis algorithm to perform noise reduction and fusion on the data of different sensors based on the initial perception matrix, calculates the covariance matrix between the data, eliminates redundant interference in the acquisition process, and generates refined perception data; The feature generation submodule is based on the refined perception data, uses the gradient boosting algorithm to calculate the spatial feature distribution, divides the perception abnormality area through cluster analysis, combines the polynomial regression to fit the benchmark model and calculates the deviation of each feature to generate the environmental perception map.

4. The trapping device collaborative control system based on multi-sensor fusion according to claim 1, characterized in that: The target recognition module includes a spectrum analysis submodule, a type classification submodule, and a feature quantization submodule; The spectrum analysis submodule uses millimeter-wave radar to collect target reflection signals, reconstructs signal frequency domain features through Fourier transform algorithm, and uses bandpass filtering to extract the spectrum difference between target and background to generate target contour information; The type classification submodule uses the K-means algorithm to segment overlapping target areas based on target contour information, extracts the motion trajectory, speed and direction characteristics of each independent area, and uses random forest to build a multi-class classification model to distinguish insects, rodents and birds, and generate target category labels; The feature quantification submodule uses a regional statistical algorithm to calculate the proportion of each type of target based on the target category label, combines simulated annealing to optimize the weight distribution strategy, comprehensively evaluates the impact of the target on the trapping effect, and generates target recognition data.

5. The trapping device collaborative control system based on multi-sensor fusion according to claim 1, characterized in that: The risk assessment module includes an activity monitoring submodule, a region division submodule, and a threat inversion submodule; The activity monitoring submodule uses an ultrasonic sensor array to scan the trapping area and converts the raw sound wave data into activity frequency values ​​through a signal amplification algorithm to generate a high-resolution activity distribution map; The region division submodule uses the dynamic threshold method to divide the high-risk core area and the diffusion area based on the activity distribution map, combines the morphological opening operation to remove noise interference in the segmented area, and generates risk level area boundary data; The threat inversion submodule constructs a Markov chain probability transfer model based on the risk level area boundary data, uses the Monte Carlo method to iteratively solve the coupling relationship between activity frequency and threat, calculates the equivalent threat distribution through the risk propagation equation, and generates a risk distribution map.

6. The trapping device collaborative control system based on multi-sensor fusion according to claim 1, characterized in that: The collaborative relationship modeling module includes a state acquisition submodule, a pattern decomposition submodule, and an association analysis submodule; The state acquisition submodule is based on a wireless sensor network, synchronously collects the state signals of each node of the trapping device, eliminates transmission delay interference through Kalman filtering, and generates multi-channel state time domain signals; The mode decomposition submodule decomposes the signal into multiple eigenmode components based on the state time domain signal by applying the singular value decomposition technique, removes the false mode components by energy threshold screening, and retains the effective components reflecting the real interaction of the devices; The correlation analysis submodule calculates the spatial correlation and phase difference of each component based on the effective mode component by cross-correlation analysis, identifies the collaborative dependency characteristics between devices through coherence analysis, and generates collaborative characteristic parameters.

7. The trapping device collaborative control system based on multi-sensor fusion according to claim 1, characterized in that: The path optimization module includes a trajectory planning submodule, a constraint definition submodule, and a multi-objective solution submodule; The trajectory planning submodule generates the initial device movement path based on the trapping area topology using an artificial potential field method, dynamically adjusts the attraction weights of the path nodes through a potential field gradient update mechanism, and generates a set of candidate paths; The constraint definition submodule builds an energy consumption model and a coverage prediction model based on historical trapping data, defines path length, turning frequency and communication load as optimization variables, and generates multidimensional constraint conditions; The multi-objective solution submodule applies the particle swarm optimization algorithm to solve the Pareto frontier solution set based on the candidate path set and the constraint conditions, selects the path solution with the best comprehensive performance through the entropy weight method, and generates the optimized path instructions.

8. The trapping device collaborative control system based on multi-sensor fusion according to claim 1, characterized in that: The decision fusion module includes a data compression submodule, an interactive analysis submodule, and an indicator calculation submodule; The data compression submodule uses independent component analysis to extract key feature vectors based on multi-source heterogeneous data, screens the main independent components by feature contribution rate to reduce data dimensions, and generates a dimension-reduced feature matrix; The interaction analysis submodule calculates the interaction strength between parameters based on the dimensionality reduction feature matrix and applies the Spearman rank correlation algorithm to construct a parameter dependency network diagram to identify the core parameter group that affects the trapping effect; The indicator calculation submodule is based on the parameter dependency network diagram, uses a logistic regression model to quantify the decision probability of each parameter, generates a comprehensive evaluation index through probability weighted fusion, and generates a fusion decision index.

9. The trapping device collaborative control system based on multi-sensor fusion according to claim 1, characterized in that: The execution control module includes a sequence generation submodule, a feedback adjustment submodule, and an optimization learning submodule; The sequence generation submodule uses a tabu search algorithm to search for the optimal action combination of the trapping device based on the fusion decision index, and balances the global exploration and local development capabilities through the tabu length strategy to generate the initial action sequence; The feedback adjustment submodule estimates the execution deviation based on the coverage effect detection data during the execution process using a proportional-integral-derivative controller, and dynamically adjusts the correction coefficients of the trapping intensity and range parameters; The optimization learning submodule constructs a deep Q network model based on historical control data, iteratively updates the action value function through the policy gradient algorithm, optimizes the adaptive adjustment rules of the execution parameters, and generates control execution instructions.

10. The trapping device collaborative control system based on multi-sensor fusion according to claim 4, characterized in that: The K-means algorithm adopts a multi-scale distance measurement strategy, calculates the similarity characteristics of target regions through Mahalanobis distance, combines density clustering to enhance regional separability, and generates independent target region boundaries.

Images