Wild animal prevention and control system suitable for extreme environment

Abstract

The invention discloses a wildlife prevention and control system suitable for an extreme environment, and the system comprises a monitoring subsystem which is used for collecting image data of a wildlife appearing and disappearing region; the intelligent identification subsystem is used for receiving the image data, identifying and classifying wild animals in the image through a deep learning model, and outputting an identification result; wherein the deep learning model is integrated with an attention module and is subjected to adversarial training; the ion expelling system is used for starting one or more expelling measures according to the identification result; the transmission subsystem is used for transmitting image data and identification results among the monitoring subsystem, the intelligent identification subsystem, the ion driving system and the remote control end; and the control system is used for calculating a threat level based on the identification result and a preset dynamic threat level model, and controlling the ion expelling system to execute corresponding expelling operation. The problems that an existing system is low in recognition rate and poor in environmental adaptability are solved, and automatic, intelligent and active prevention and control of wild animal threats are achieved.

Description

Technical Field

[0001] This invention relates to a wildlife control system suitable for extreme environments. Background Technology

[0002] As human activities continue to expand, scientific research, resource exploration, and ecological protection are increasingly extending into remote and harsh natural areas, such as plateau regions. In these areas, independently constructed buildings and personnel face a series of unique security challenges, one of the most prominent being the intrusion of large wild animals. Wild animals such as bears and wolves, driven by factors such as searching for food, defending their territory, or simply out of curiosity, may cause serious damage to buildings and, in some cases, directly threaten the lives of those inside.

[0003] Currently, some wildlife control systems attempt to address wildlife threats by triggering alarms through motion detection or using simple deterrent devices (such as bright lights and sounds). However, the fundamental flaw in these systems lies in their severely inadequate front-end sensing and recognition capabilities, directly leading to the blind and inefficient nature of deterrent operations. Specifically: First, the lack of intelligent and accurate identification capabilities leads to false triggering or failure of deterrence actions. Most existing systems rely on simple motion detection or basic image processing algorithms, unable to accurately distinguish between intruding wild animals (such as bears and wolves), harmless animals (such as Tibetan antelopes), or environmental disturbances like rustling grass in complex backgrounds. This lack of identification capability results in two extreme consequences: first, a high false alarm rate, where the system frequently issues deterrence signals to non-threatening targets, wasting valuable energy (such as battery power and water) and potentially causing wild animals to become "habitual" with the deterrence methods, thus losing their deterrent effect; second, a high false negative rate, where the system cannot reliably detect genuine threats under adverse conditions, causing the deterrence system to "fail to activate," creating serious safety hazards. Consequently, even with deterrence devices, the effectiveness and reliability of the entire system are significantly reduced.

[0004] Secondly, the challenges of image quality in extreme environments further amplify the performance limitations of recognition models. In regions like Hoh Xil, known as the "forbidden zone of life," winter temperatures can plummet to below -40°C, often accompanied by severe weather conditions such as low light, blizzards, and dense fog. Ordinary image acquisition equipment struggles to obtain clear images in such environments, and existing recognition models generally lack robust optimization for these extreme conditions. When faced with image noise, motion blur, and target occlusion, the model's recognition accuracy drops sharply. This means that in the harshest moments when the system is most needed to function, its core "eyes" and "brain" (i.e., recognition capabilities) fail first, rendering any subsequent expulsion decisions baseless.

[0005] Furthermore, the lack of model adaptability and data transmission bottlenecks jointly limit the system's practical effectiveness. Existing deep learning models are often trained on general datasets, lacking specialized datasets for wildlife in high-altitude areas and their seasonal characteristics (such as molting). Simultaneously, the models cannot be continuously optimized online based on field data after deployment, making it difficult to adapt to long-term changes in the environment and animal behavior. In addition, poor network signal coverage in remote areas makes stable data transmission and remote model updates difficult. This means that once deployed, the system's recognition performance becomes fixed in an initial state that may not be applicable, ultimately leading to a disconnect between the driven deterrence strategy and real-world needs.

[0006] In summary, the core contradiction of existing technologies lies in the low accuracy and robustness of front-end image recognition, which severely restricts the intelligence and effectiveness of back-end deterrence actions. Therefore, it is necessary to design an integrated prevention and control system that can stably and accurately identify wild animals in extreme environments and execute intelligent, high-success-rate deterrence strategies accordingly. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention provides a wildlife control system suitable for extreme environments. The core of this system lies in achieving high-precision detection and classification of wild animals through optimized intelligent image recognition technology, and intelligently driving them away based on the recognition results.

[0008] The technical solution adopted in this invention is as follows: An animal hazard control system suitable for extreme environments includes a monitoring subsystem, an intelligent identification subsystem, an ion-fighting system, a transmission subsystem, and a control system. The intelligent identification subsystem, as the core of this invention, is based on AI computing power and incorporates a dedicated animal identification algorithm to analyze image data transmitted by the monitoring subsystem and output identification results. This intelligent identification subsystem supports multi-channel video access and can simultaneously process image information from multiple monitoring points. Specifically, this intelligent identification subsystem uses a deep learning model to quickly and accurately identify large wild animals such as bears and wolves. The deep learning model is based on the YOLOv5s architecture and is specifically designed for extreme environments, including: 1) Integrating a SENet attention module at the end of the backbone network: First, global average pooling is performed on the input feature map to generate channel weight vectors, then weight coefficients are calculated through a fully connected layer and a sigmoid activation function, with a weight scaling factor set to 0.5 to enhance low-light robustness; 2) Adversarial training uses the FGSM method: Model weights are initialized using Xavier; data preprocessing includes normalization to [0,1] and random horizontal flipping; the iterative training process includes forward... The propagation, loss calculation, backpropagation, and weight update processes were performed using the Adam optimizer (learning rate 0.001, batch size 16) for 100 iterations. Loss convergence was defined as the decrease rate of the validation set loss being <0.001 for 10 consecutive iterations. The model training dataset specifically included over 5000 wildlife images collected in the plateau region, labeled with different weather and lighting conditions. Continuous optimization was achieved through an online incremental learning mechanism (based on the Elastic Weighted Written (EWC) algorithm, with key parameters including the Fisher information matrix regularization coefficient λ=0.8 and a learning rate of 0.0005). The model output included animal species confidence and location information for use in the control system's decision-making. The intelligent recognition subsystem was connected to a switch via a Cat5e network cable to receive image data and output recognition results to the control system. To improve recognition accuracy, the intelligent recognition subsystem also had a self-learning function, continuously optimizing the recognition model through sample training.

[0009] In addition, the intelligent recognition subsystem integrates a lightweight environmental perception module that can analyze the levels of snow, fog, and dust in the image in real time, and dynamically adjust image preprocessing parameters (such as contrast and sharpening) and the model's confidence threshold to reduce the false alarm rate under extreme weather conditions. Furthermore, the intelligent recognition subsystem employs an online incremental learning mechanism based on the Elastic Weighted WC (EWC) algorithm: key parameters include the Fisher information matrix regularization coefficient λ = 0.8 (set based on empirical values) and a learning rate of 0.0005; the implementation process is as follows: 1) Real-time acquisition and preprocessing of new image data; 2) Calculation of parameter importance using mini-batch (size = 8): quantifying parameter importance by calculating the second derivative of the loss function (Hessian matrix approximation); 3) Local update of weights through gradient descent: the update formula is... i new = i old - or ∇L( i ),in or 4) Stop updating when the loss function converges (the absolute value of the rate of change of loss for 5 consecutive iterations is <0.01).

[0010] The monitoring subsystem consists of several high-definition intelligent PTZ cameras. Each camera is a DS-2DC6432IW-A model, featuring 32x optical zoom and supporting 1080P resolution imaging. These cameras are IP66 rated for protection and operate within a temperature range of -40℃ to 70℃, enabling them to withstand extreme weather conditions. Each camera incorporates a built-in intelligent tracking algorithm that automatically adjusts the pan / tilt angle based on the trajectory of moving objects, ensuring the target remains within the field of view. The cameras support wide dynamic range, strong light suppression, and backlight compensation, providing clear imaging under various lighting conditions. The monitoring subsystem utilizes RVV2×1.0mm... 2 The cold-resistant power cord is connected to the Huntkey HK600-12UEP server power supply to ensure a stable power supply. Meanwhile, each high-definition intelligent PTZ camera is connected to the switch via Cat 5e cold-resistant network cables to achieve image data aggregation and transmission. Furthermore, the high-definition intelligent PTZ camera integrates low-light night vision functionality, enabling it to capture clear color images even in extremely dark environments with illumination levels below 0.001 lux. Combined with adaptive image enhancement algorithms, it further improves nighttime monitoring performance.

[0011] The ion-repelling system activates corresponding repelling measures based on the identification results from the intelligent identification subsystem. Among these measures, the high-decibel alarm, model DH-AS520, is selected, capable of emitting a sound signal up to 120 decibels, effectively deterring sound-sensitive animals such as wolves. The alarm features IP65 protection, an operating temperature range of -30℃ to 60℃, and is ensured to operate for short periods at low temperatures through an external insulation cover and built-in heating elements. It also utilizes RVV4×0.75mm... 2 The cold-resistant control cable connects to the control system. The automatic sprinkler system is equipped with intelligent nozzles with a range greater than 39 meters and supports angle and range adjustment. The controller of the automatic sprinkler system is connected to the control system via an RS485 communication cable and initiates the sprinkler operation upon receiving a command. The bear-repellent spray device is used to spray a pungent spray containing capsaicin onto the target area.

[0012] In addition, the ion-repelling system is equipped with a multimodal repelling strategy, which dynamically adjusts the repelling method based on animal behavior feedback. For example, when a single repelling method (such as an audible alarm) fails to effectively repel the animal after two consecutive triggers, the system automatically switches to a combined repelling mode (such as sound + water spray) and optimizes the repelling sequence through a reinforcement learning algorithm to improve the repelling success rate. The reinforcement learning algorithm adopts the DQN framework: the Q network is a three-layer fully connected neural network (input layer dimension = state dimension, 128 hidden nodes using ReLU activation, output layer dimension = number of actions); in the state space, the animal type is encoded using one-hot encoding (e.g., bear = [1,0], wolf = [0,1]), the distance unit is meters, and the number of repelling attempts is an integer; in the action space, 'combination' is defined as starting water spraying 2 seconds after the sound wave; the reward function R = +10 (animal moves 30 meters away) / -5 (animal does not move within 10 minutes). Training used an ε-greedy strategy (ε=0.2), with an experience replay cache size of 1000, a Q-network learning rate of 0.001, and data sourced from a simulated environment generator (built based on historical animal behavior data). The target network was updated every 1000 steps, and training stopped when the average reward changed by less than 1% for 5 consecutive iterations.

[0013] The transmission subsystem utilizes a satellite communication terminal to achieve data transmission between the monitoring subsystem, intelligent identification subsystem, ion-driving system, and remote controller. The satellite communication terminal selected is the VNet800 model, supporting multiple satellite frequency bands with a transmission rate of no less than 2Mbps (upload) and 4Mbps (download). Furthermore, the satellite communication terminal connects to the satellite antenna via a dedicated RF cable and to the switch via a Cat5e network cable to achieve real-time data transmission. The transmission subsystem can operate stably within a temperature range of -40℃ to 70℃, meeting communication requirements in extreme environments. To enhance data transmission reliability, the transmission subsystem also employs forward error correction (FEC) coding and automatic repeat request (ARQ) mechanisms, combined with channel state adaptive modulation technology, to control the data transmission bit error rate to within 10% even in severe weather conditions. -6 The following steps ensure the integrity and real-time performance of critical commands and image data.

[0014] The control system uses the Advantech IPC-610L industrial control computer as its core device. It receives the identification results from the intelligent identification subsystem and issues control commands according to preset rules. By adding an external insulation box and a built-in heating module, the industrial control computer extends its operating temperature range to -40℃ to 70℃, ensuring stable operation in extreme low temperatures. It supports multiple serial ports and multiple network ports and connects to a switch via Cat5e network cables to achieve data exchange with other subsystems.

[0015] The control system receives status information from each subsystem via a data interface and stores it in local storage. Based on data, algorithms, and logic, it sends scheduling or control commands to each subsystem to coordinate their operation. The control software is based on a modular architecture, including a data acquisition module, a decision engine, and an execution control module. The data acquisition module receives monitoring and identification data in real time. The decision engine uses a dynamic threat level model based on multi-factor weighted assessment. Assessment factors include animal species, movement speed vector (unit: m / s), group size, historical behavior records, and real-time environmental factors (such as wind speed, unit: level). The comprehensive threat score is calculated using the following formula: S =0.3× C type +0.4× V speed +0.2× N group +0.1× W wind ,in C type The species coefficients are determined based on historical attack data statistics: bear = 1.0, wolf = 0.8, other animals = 0.5; the coefficient weights are determined through linear regression analysis of historical datasets (sample size ≥ 1000); the response strategy threshold is set as follows: S'Monitoring and tracking' will be activated when the time is <30; 'monitoring and tracking' will be activated when S <45 triggers 'audible warning', 45≤ S <60 initiates 'physical removal', S ≥60 triggers 'chemical removal', with the threshold selected based on ROC curve testing (using a simulated dataset, AUC≥0.85).

[0016] The energy supply and storage system includes grid power, solar power, and a diesel backup system. The primary energy source is an off-grid solar photovoltaic system, powering the deep well pumps, booster pumps, and terminal heating cables. The total power and battery capacity of the solar photovoltaic system are precisely configured based on the power consumption of all equipment and daily operating hours to ensure it can withstand consecutive cloudy days. Specifically, the total power of the solar panels and the battery capacity are calculated based on the local average effective sunshine hours and the system's average daily power consumption. For example, in the Hoh Xil region, with an average of 5 hours of effective sunshine per day, if the equipment's average daily power consumption is 25 kWh, then a solar panel array of approximately 15 kW (generating approximately 75 kWh of electricity per day) and at least 60 kWh of lithium-ion batteries are required to support continuous system operation for more than two days without sunshine. The diesel generator serves as backup power, manually started when there are consecutive cloudy days and the battery is depleted, providing emergency power to the system. The energy supply and storage system is also equipped with a UPS (Uninterruptible Power Supply) to ensure continuous operation of critical equipment during power switching.

[0017] The water supply and insulation system prioritizes groundwater as its source, obtained through drilling wells. The water storage tank is made of stainless steel and equipped with a ≥100mm thick closed-cell rubber-plastic insulation board for external insulation, further protected by a galvanized steel or aluminum shell. Internal heating utilizes a stainless steel immersion heater, requiring 1.5kW to 2kW of heating power per cubic meter of water to maintain the water temperature between 5℃ and 10℃. The water storage tank is equipped with a small circulating water pump that operates periodically to ensure water circulation and uniform temperature, and includes low-temperature and low-level alarm systems. Pipeline heating and insulation are achieved through auxiliary electric heating. The water supply pipes are reinforced with a ≥60mm thick rubber-plastic insulation layer and protected with a sheet metal shell to prevent physical damage. Specifically, anti-freeze thermostats are installed at the pipe ends, activating heating only when the temperature drops below 5℃ to minimize energy consumption. Pipelines are laid either buried below the frost line or overhead, with a burial depth of at least 3 meters, and PE-RT is the preferred material. The water supply system primarily uses solar-powered booster pumps, supplemented by terminal electric heating. Deep well pumps draw groundwater into insulated water tanks buried underground. The booster pumps are powered by the solar system and controlled by a level controller and pressure sensor. The level controller has set high and low thresholds; when the water level drops below 50%, the deep well pump starts, and when it reaches 90%, it stops. The pressure sensor maintains the pipeline pressure at 0.3–0.5 MPa to ensure the range of the spray nozzles. Simultaneously, the system is equipped with a leak detection sensor to monitor pipeline integrity in real time and automatically shuts off the water supply and issues an alarm upon detecting a leak.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) In terms of image recognition, this invention achieves high-precision and robust detection and classification of wild animals through an intelligent recognition subsystem optimized for extreme environments. Specifically, the deep learning model used in this invention is based on a lightweight YOLOv5s architecture and is specifically optimized, integrating the SENet channel attention module (with a weight scaling factor set to 0.5) to improve the extraction capability of key features. Furthermore, the model generates diverse perturbation samples through an adversarial training strategy using FGSM, effectively enhancing its generalization ability and stability under harsh imaging conditions such as low light, blizzards, and dense fog, thereby overcoming the recognition degradation problem caused by image noise, motion blur, and partial occlusion of the target. The model training dataset specifically contains more than 5,000 high-quality labeled images of wild animals in the plateau region, comprehensively covering different seasons, various typical weather conditions (such as rain, snow, and dust storms), and complex lighting conditions (such as dawn and dusk, and backlighting), ensuring the model's strong adaptability to the characteristics of animals in specific regions. Test data shows that the system achieves recognition accuracy of 97.2% for wild bears and 95.8% for wolves, and even in snowy conditions with extremely low visibility, the recognition accuracy remains consistently above 93.5%. Compared with traditional image recognition methods, this system significantly reduces the false alarm rate and false negative rate by 82% and 75% respectively in the same test scenario. Therefore, the optimized model of this invention significantly improves recognition accuracy in extreme environments.

[0019] (2) This invention introduces a real-time environment perception module and an online incremental learning mechanism (based on the elastic weight solidification EWC algorithm, with key parameter λ=0.8), which directly serves the image recognition process and improves the model's real-time optimization capability and anti-interference ability in dynamic environments. Specifically, the environment perception module can use image analysis technology to quantitatively evaluate the snowfall intensity, fog concentration, and dust level in the image in real time, and dynamically optimize the image preprocessing parameters (such as adaptive adjustment of local contrast and nonlinear sharpening intensity) and the model output confidence threshold accordingly; while the online incremental learning mechanism can continuously learn and adapt to the feature drift of animal appearance caused by seasonal molting, body shape changes, etc., ensuring that the model can continue to evolve during long-term deployment, significantly improving the recognition model's real-time optimization capability and anti-interference ability in dynamic environments.

[0020] (3) This invention constructs a multi-modal intelligent deterrence mechanism based on animal behavior principles. This mechanism accurately identifies the unique "fear weaknesses" of different species (such as specific sound wave frequencies and light patterns) through deep learning and integrates a dynamic threat assessment system. The system utilizes an integrated sensor network (including thermal imaging and motion sensors) and intelligent algorithms deployed in the wild to monitor animal activity trajectories, group size, and approach speed in real time, and automatically matches and triggers the optimal deterrence method (such as directional sound waves, strong light of specific frequency bands, or odor spray). In a closed test field that highly simulates the real wild environment, the overall success rate of this deterrence mechanism is close to 95%, achieving efficient and relatively accurate proactive intervention against threats to wild animals.

[0021] (4) At the data transmission level, this invention utilizes an industrial-grade ruggedized satellite communication terminal to construct the core transmission subsystem. This terminal supports mainstream satellite communication frequency bands such as L / S / C and maintains stable operation within an extreme temperature range (-40℃ to 70℃). It employs forward error correction coding and adaptive modulation technology to ensure low-latency and high-reliability transmission of image data streams and control commands in remote, uninhabited areas, overcoming the limitations of insufficient signal coverage in vast uninhabited areas by traditional cellular or wireless networks, and achieving more reliable remote data transmission.

[0022] (5) The control system of the present invention realizes intelligent decision-making based on a multi-factor weighted evaluation model. The model analyzes in real time multiple information from the identification subsystem, environmental perception module, eviction status feedback and energy storage unit operation data, etc. Through the preset decision rule engine, it automatically generates the optimal response strategy and schedules the cooperative work of each subsystem (such as adjusting the identification frequency, switching the eviction mode, and starting the backup energy), which reduces the dependence on manual monitoring and intervention and improves the autonomy and response efficiency of the system.

[0023] (6) In terms of energy security, the energy supply, storage and water supply system of the present invention adopts solar energy as the main power source, diesel generator as backup and multiple insulation design. This design ensures that the system can maintain the stable operation of key equipment (such as heating components, communication terminals and computing units) even when encountering continuous rain / snow and extremely low temperature environments below -30℃, providing a solid energy foundation for the continuous execution of core functions (especially real-time image recognition with high computing power requirements).

[0024] In summary, this invention uses image recognition technology as its core and deeply integrates key subsystems such as environmental perception, intelligent deterrence, reliable transmission, intelligent decision-making, and highly robust energy to construct a highly automated and reliable intelligent wildlife control system suitable for extreme outdoor environments such as plateaus and extreme cold. Attached Figure Description

[0025] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0026] This invention provides a wildlife control system suitable for extreme environments such as the Hoh Xil region. It achieves effective monitoring, automatic identification, and intelligent removal of wild animals through a monitoring subsystem, an intelligent identification subsystem, an ion-repelling system, a transmission subsystem, and a control system. The following is in conjunction with... Figure 1 The specific embodiments of the present invention are described in detail below.

[0027] In the overall system layout, high-definition intelligent PTZ cameras, serving as the front-end equipment of the monitoring subsystem, are installed at key locations around the building for 24 / 7, blind-spot-free monitoring of the surrounding environment. Each high-definition intelligent PTZ camera uses the DS-2DC6432IW-A model, featuring 32x optical zoom and 1080P resolution imaging, while also supporting wide dynamic range, strong light suppression, and backlight compensation. The equipment operates within a temperature range of -40℃ to 70℃, meeting the requirements of the extreme climatic conditions in the Hoh Xil region. The high-definition intelligent PTZ camera incorporates an intelligent tracking algorithm that adjusts the pan-tilt angle based on the trajectory of moving objects, ensuring the target remains within the field of view. (RVV2×1.0mm) 2 Cold-resistant power cords connect the high-definition intelligent PTZ cameras to the server power supply to ensure a stable power supply. In addition, each high-definition intelligent PTZ camera is connected to the switch via Cat5e cold-resistant network cables to achieve image data aggregation and transmission. The high-definition intelligent PTZ cameras are also equipped with a defrosting function, using built-in heated lenses and periodically activated fans to prevent lens frost or ice buildup in extremely cold environments, ensuring image clarity.

[0028] The core equipment of the intelligent identification subsystem is based on AI computing power and incorporates a dedicated animal recognition algorithm for analyzing image data transmitted by the monitoring subsystem. The intelligent identification subsystem supports multi-channel video access and can simultaneously process image information from multiple monitoring points. Specifically, the intelligent identification subsystem uses a deep learning model optimized based on the YOLOv5s architecture to quickly and accurately identify large wild animals such as bears and wolves, with a response time controlled within 5 seconds. Specific optimizations of the deep learning model include: integrating a SENet attention module at the end of the backbone network to enhance the model's robustness under low-light conditions; and employing the FGSM method for adversarial training to improve the model's resistance to image noise and blur. The model training dataset contains 5000 labeled images, covering animal morphology under different seasons and weather conditions. After deployment, the model continuously receives monitoring data for online optimization, achieving an accuracy rate of over 95%. Test data shows that the model's accuracy rate for identifying bears is 97.2%, for wolves it is 95.8%, and even in snowy conditions, the accuracy rate remains above 93.5%. The intelligent recognition subsystem is connected to the switch via a Cat5e network cable, receives image data, and outputs the recognition results to the control system.

[0029] To improve the adaptability of the recognition system in extreme weather conditions, the intelligent recognition subsystem also integrates a lightweight environmental perception module. This module can analyze the concentration levels of snow, fog, and dust in the image in real time, and dynamically adjust image preprocessing parameters (such as contrast enhancement intensity and nonlinear sharpening factor) and the output confidence threshold of the deep learning model accordingly, thereby effectively reducing the false alarm rate under complex weather conditions.

[0030] To improve recognition accuracy and adapt to seasonal changes in animal appearance, the intelligent recognition subsystem features an online incremental learning mechanism and employs the Elastic Weighted WC (EWC) algorithm. Its implementation process is as follows: 1) Real-time acquisition and preprocessing of new image data; 2) Calculation of parameter importance using mini-batches (size = 8); 3) Local updates of weights via gradient descent; 4) Stopping updates when the loss function converges (absolute value of the rate of change of loss < 0.01 for 5 consecutive iterations). Key parameters include: Fisher information matrix regularization coefficient λ = 0.8, and learning rate 0.0005. In this invention, the intelligent recognition subsystem adopts a distributed computing architecture, balancing the recognition task load across multiple GPU cores to process multiple video streams in parallel, ensuring a response time of less than 5 seconds even in high-concurrency scenarios.

[0031] The ion-repelling system activates corresponding repelling measures based on the identification results from the intelligent identification subsystem. The high-decibel alarm, model DH-AS520, emits a sound signal up to 120 decibels, effectively repelling sound-sensitive animals such as wolves. The high-decibel alarm features IP65 protection and an operating temperature range of -30℃ to 60℃, extended to -40℃ via an external insulation cover and built-in heating element. It also utilizes RVV4×0.75mm... 2 The cold-resistant control cable connects to the industrial control computer. The automatic sprinkler system uses a ZDMS0.8 / 20S model, equipped with intelligent nozzles, offering long range and adjustable angle and range. It initiates spraying upon receiving commands. The bear-repellent spray system includes a built-in reservoir, DC power pump, and control solenoid valve, all working together to spray a capsaicin-containing irritant spray onto the target area through a fan-shaped nozzle. The system also features a WX-XQZ bear deterrent device, equipped with dual visible light and infrared thermal imaging cameras and a built-in analysis algorithm to collaboratively detect and locate the position and distance of bear targets. This deterrent device receives commands from the control system via an RS485 communication bus. The reservoir and external piping are equipped with electrically heated insulation layers to prevent freezing at low temperatures.

[0032] Furthermore, the ion-repelling system integrates a multimodal repelling strategy and is optimized using a reinforcement learning algorithm based on the DQN framework. The state space of this reinforcement learning algorithm includes animal type (encoded by one-hot encoding), distance from the building (in meters), and the number of repelling triggers; the action space includes single repelling methods (e.g., sound only) and combined methods (e.g., water spraying after a 2-second delay following sound); the reward function is defined as: +10 when the animal moves 30 meters away, and -5 if it does not move within 10 minutes. The algorithm is trained using a simulated environment generator (built based on historical animal behavior data) and can dynamically adjust the repelling sequence based on animal behavior feedback. Performance tests show that under the reinforcement learning-optimized multimodal repelling strategy, the overall repelling success rate of the system is close to 95%. For example, for bear targets, the effective repelling distance of the bear spray device is 8–15 meters, the spray coverage angle is 60 degrees, and the duration of a single spray is 5 seconds. After a successful repelling, in 95% of cases, the bear leaves the warning area within 5 minutes and does not return within 24 hours. The ion-repelling system also integrates a sound library containing various deterrent sounds for different animals. By analyzing the animal's reaction, it dynamically selects the most effective sound segment, thus effectively improving the targeting of the repellent.

[0033] The transmission subsystem uses a satellite communication terminal to achieve data transmission between the monitoring subsystem, intelligent identification subsystem, ion-driving system, and remote controller. The satellite modem is a VNet800 model, supporting multiple satellite frequency bands with a transmission rate of at least 2Mbps (upload) and 4Mbps (download). The satellite modem connects to the satellite antenna via a dedicated RF cable and to the switch via a Cat5e network cable to achieve real-time data transmission. The transmission subsystem can operate stably within a temperature range of -40℃ to 70℃, meeting communication requirements in extreme environments. In this invention, the transmission subsystem supports data compression and encrypted transmission, using the H.265 video coding standard to compress the video stream by more than 50%, and employing the AES-256 encryption algorithm to ensure data security and prevent the leakage of sensitive information. To enhance reliability, the transmission subsystem also employs forward error correction (FEC) coding and automatic repeat request (ARQ) mechanisms, combined with channel state adaptive modulation technology, to control the data transmission error rate to within 10% even in severe weather conditions. -6 the following.

[0034] The control system uses an industrial control computer as its core device, responsible for receiving the identification results from the intelligent identification subsystem and issuing control commands according to preset rules. The industrial control computer, equipped with an external insulation box and a built-in heating module, ensures stable operation within a temperature range of -40℃ to 70℃. It supports multiple serial ports and multiple network ports for expansion and connects to a switch via Cat 5e network cable to achieve data interaction with other subsystems. The control system receives status information from each subsystem through a data interface and stores it in local storage. Based on the animal species and its proximity, combined with a dynamic threat level model, the control system activates the corresponding ion-repelling system and resumes normal monitoring mode after the animal is driven away. The dynamic threat level model is based on a multi-factor weighted assessment, with assessment factors including animal species, movement speed vector, group size, historical behavior records, and real-time environmental factors (such as wind speed). The comprehensive threat score calculation formula is as follows: S =0.3× C type +0.4× V speed +0.2× N group +0.1× W wind ,in C type is the species coefficient (bear = 1.0, wolf = 0.8, other animals = 0.5). V speed For the speed of movement, N group For the group size, W wind For wind speed factors. The response strategy threshold is set as follows: S'Monitoring and tracking' will be activated when the time is <30; 'monitoring and tracking' will be activated when S <45 triggers 'audible warning', 45≤ S <60 activates 'physical dispersal' (i.e., water spraying). S ≥60 triggers 'chemical repellency' (i.e., spraying an irritant containing capsaicin).

[0035] The control software's workflow includes: initialization phase detecting the status of each subsystem; operation phase cyclically receiving identification data and updating the threat assessment every 100ms; decision logic outputting expulsion commands based on the aforementioned threat assessment model; and a log module recording all events, supporting remote querying and diagnostics. The control system receives software update packages via satellite link and performs periodic system health checks, automatically identifying hardware faults or performance degradation, and generating maintenance reports to send to the management center.

[0036] The energy supply and storage system includes mains power, solar power, and a diesel backup system. The primary energy source is an off-grid solar photovoltaic system, powering deep well pumps, booster pumps, and terminal heating cables. The total power and battery capacity of the solar photovoltaic system are precisely configured according to the power requirements of the equipment and daily operating hours to ensure it can cope with consecutive cloudy days. For example, for a system with an average daily power consumption of 25 kWh, under the condition of an average of 5 hours of effective sunshine per day in the Hoh Xil region, a 15 kW solar panel array (approximately 75 kWh of daily power generation) and a 60 kWh lithium battery pack (using lithium iron phosphate, operating temperature -20℃ to 60℃, and optimized charging efficiency through an MPPT controller) are configured. A diesel generator serves as backup power, manually started when there are consecutive cloudy days and the battery is depleted, providing emergency power to the system. The energy supply and storage system is equipped with a UPS (Uninterruptible Power Supply) to ensure continuous operation of critical equipment during power switching. Energy efficiency tests show that the solar power system generates an average of 45-65 kWh per day in high-altitude areas, which fully meets the system's daily power consumption requirement of 25 kWh. The battery can support the system to run continuously for 72 hours when fully charged. The diesel generator ran for a total of 42 hours during the test, consuming 28 liters of diesel fuel, which proves the rationality and economy of the system's energy configuration.

[0037] The water supply and insulation system prioritizes groundwater as its source, obtained through drilling wells. The water storage tanks are made of stainless steel and equipped with ≥100mm thick closed-cell rubber-plastic insulation boards for external insulation. Internal heating utilizes stainless steel immersion heaters, requiring 1.5kW to 2kW of heating power per cubic meter of water to maintain the water temperature between 5℃ and 10℃. Pipeline heating and insulation are achieved through auxiliary electric heating; the water supply pipes are reinforced with ≥60mm thick rubber-plastic insulation layers and equipped with sheet metal protective shells to prevent physical damage. Anti-freeze thermostats are installed at the pipe ends, activating heating only when the temperature drops below 5℃ to minimize energy consumption. The water supply system primarily uses solar-powered booster pumps, supplemented by terminal electric heat tracing. Deep well pumps draw groundwater to insulated water tanks buried underground. The booster pumps are powered by the solar system and controlled by level controllers and pressure sensors. In its implementation, a low-temperature resistant deep well pump was selected, with a working depth of 50 meters and a head of 40 meters. The insulated water tank, with a volume of 2 cubic meters, was installed 3 meters below the frost line, with an internal heater power of 3kW, controlled by a temperature controller. The heating cable, with a power density of 15W / m, was laid along the entire length of the water supply pipeline and its start / stop was triggered by a temperature sensor. According to insulation performance tests, in an extreme environment of -35℃, the water temperature inside the tank remained at 6.2±1.5℃, while the water temperature at the end of the pipeline remained at 4.8±1.2℃, with no freezing observed. This indicates that the insulation measures of the water supply and insulation system met the design requirements, ensuring stable water temperature. Furthermore, the system's average daily water consumption was 0.8 cubic meters, and its energy consumption was 12.5 kWh, with the heating system accounting for 68% of the energy consumption, meeting the expected energy-saving standards. In this invention, the water supply and insulation system uses phase change material (PCM) as an additional insulation layer. PCM plates are wrapped around the outer walls of pipes and water tanks. By utilizing its latent heat of phase change to absorb or release heat, temperature fluctuations can be effectively smoothed and heating energy consumption can be reduced.

[0038] During system operation, high-definition intelligent PTZ cameras first acquire image information of the surrounding area of ​​the building and transmit the image data to the switch via Category 5e cold-resistant network cable. The switch forwards the image data to the intelligent recognition subsystem, which analyzes and identifies targets in the images using a deep learning model. If a large wild animal such as a bear or wolf is detected entering the set distance range, the intelligent recognition subsystem sends the identification result to the control system. The control system activates the corresponding ion-repelling system based on the assessment results of the dynamic threat level model, such as activating a high-decibel alarm or bear-repelling spray device. Simultaneously, the control system issues commands to activate the video recording system and PTZ tracking function to ensure that the target object is continuously recorded and tracked. After the ion-repelling system is activated, it drives the animal away by spraying irritating spray or emitting a high-decibel sound. Once the animal leaves the set distance range, the control system issues commands to restore each subsystem to normal monitoring mode.

[0039] During the system installation and debugging phase, device connection and functional testing are required. Debugging the monitoring subsystem includes network connectivity testing, image quality testing, PTZ control testing, and tracking function testing. Debugging the intelligent recognition subsystem includes device connection testing, recognition model loading and training, recognition accuracy testing, and response speed testing. Recognition model loading and training must be configured and verified according to the adversarial training and online incremental learning (EWC) process described in the invention. During debugging, it is necessary to verify whether the environmental perception module can correctly adjust preprocessing parameters and confidence thresholds based on simulated snow and fog images. Debugging the ion drive system includes high-decibel alarm debugging, automatic sprinkler device debugging, bear-repelling spray device debugging, and linkage testing. In the linkage testing, it is necessary to verify whether the multimodal repulsion strategy can dynamically adjust the repulsion sequence based on simulated animal behavior feedback. Debugging the transmission subsystem includes satellite signal strength testing, data transmission rate testing, and transmission stability testing. Debugging the control system includes device control testing, data processing and storage testing, and linkage logic testing. The linkage logic testing should focus on verifying the dynamic threat level model based on multi-factor weighted assessment and its response thresholds. The overall system debugging requires simulated scenario testing, extreme environment testing, and long-term operation testing to ensure system stability and reliability.

[0040] In extreme environment testing, the system must operate under winter or severe weather conditions to verify its performance in environments with low temperatures, strong winds, and blizzards. For example, in an environment of -30°C, the system must operate continuously for more than 72 hours without human intervention. Additional insulation measures (such as insulated boxes and heating modules) ensure the equipment operates normally at ultra-low temperatures. Regularly check the operating status of each device, data transmission, and storage, and record any problems encountered during system operation. Analyze and resolve any problems promptly to ensure the system can operate stably and reliably for extended periods.

[0041] Compared with existing technologies, this invention has outstanding advantages in extreme environment adaptability, identification accuracy, system integration and automation level, and solves the problem of threats from wild animals to independent and remote buildings and their inhabitants, and has high practical application value.

[0042] The above are merely preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but which still solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A wildlife control system suitable for extreme environments, characterized in that, include: The monitoring subsystem is used to collect image data of areas where wild animals are present. The intelligent identification subsystem communicates with the monitoring subsystem to receive image data and identify and classify wild animals in the images using a deep learning model, outputting the identification results; the deep learning model integrates an attention module and has been trained adversarially. The ion-repelling system communicates with the intelligent identification subsystem to initiate one or more ion-repelling measures based on the identification results; The transmission subsystem is used to transmit image data and recognition results between the monitoring subsystem, the intelligent identification subsystem, the ion-driving system, and the remote control terminal. The control system communicates with the intelligent identification subsystem and the ion removal system to calculate the threat level based on the identification results and the preset dynamic threat level model, and to control the ion removal system to perform the corresponding removal operation.

2. A wildlife control system suitable for extreme environments according to claim 1, characterized in that, The deep learning model is based on the YOLOv5s architecture and integrates an SENet attention module at the end of the backbone network. The weight scaling factor of the SENet attention module is set to 0.

5. The deep learning model uses the FGSM method for adversarial training and the Adam optimizer for iterative training.

3. A wildlife control system suitable for extreme environments according to claim 2, characterized in that, The intelligent recognition subsystem also includes an environmental perception module and an online incremental learning mechanism. The environmental perception module is used to analyze the levels of snow, fog, and dust in the image in real time and dynamically adjust the image preprocessing parameters. The online incremental learning mechanism is based on the elastic weight merging algorithm, with a Fisher information matrix regularization coefficient λ=0.8 and a learning rate of 0.0005.

4. A wildlife control system suitable for extreme environments according to claim 3, characterized in that, The ion-repelling system includes a high-decibel alarm, an automatic water spray device, and a bear-repelling spray device.

5. A wildlife control system suitable for extreme environments according to claim 4, characterized in that, The ion removal system is also equipped with a multimodal ion removal strategy. The multimodal ion removal strategy adopts a reinforcement learning algorithm based on the DQN framework. Its Q network is a three-layer fully connected neural network, trained using the ε-greedy strategy, and the experience replay cache size is 1000.

6. A wildlife control system suitable for extreme environments according to claim 5, characterized in that, The transmission subsystem uses a satellite communication terminal and employs forward error correction coding and automatic retransmission request mechanisms to control the data transmission error rate to below 10⁻⁶.

7. A wildlife control system suitable for extreme environments according to claim 6, characterized in that, The dynamic threat level model of the control system is based on multi-factor weighted evaluation. Its comprehensive threat score calculation formula is S=0.3×Ctype +0.4×Vspeed +0.2×Ngroup +0.1×Wwind, where Ctype is the animal species coefficient, Vspeed is the movement speed, Ngroup is the group size, and Wwind is the wind speed factor.

8. A wildlife control system suitable for extreme environments according to claim 7, characterized in that, It also includes an energy supply and storage system, which comprises a solar photovoltaic system, a battery bank, and a diesel backup generator, wherein the total power of the solar photovoltaic system and the battery capacity are configured based on the local average effective sunshine hours and the system's average daily power consumption.

9. A wildlife control system suitable for extreme environments according to any one of claims 1 to 8, characterized in that, It also includes a water supply insulation system, which includes an underground water source, an insulated water storage tank, a pipe insulation structure, and a heating device. The insulated water storage tank is made of stainless steel and equipped with a closed-cell rubber and plastic insulation board with a thickness of not less than 100mm. The pipe insulation structure includes a rubber and plastic insulation layer with a thickness of not less than 60mm and a sheet metal protective shell.

Images