New energy station intrusion detection method and equipment based on linkage technology, and medium

Abstract

The invention discloses a new energy station intrusion detection method based on a linkage technology, equipment and a medium, and relates to the technical field of new energy station safety protection, electronic pulse fences, infrared correlation and video acquisition equipment are deployed, perimeter signals and images are acquired, multi-source data are fused through space-time calibration and a D-S evidence theory, a target is detected by using a YOLOv5 model, and a new energy station intrusion detection result is obtained. Kalman filtering is used for predicting a track, distinguishing normal operation and intrusion behaviors, setting a three-level response mechanism, triggering sound-light alarm, access control blocking and the like, data are stored, a model is optimized regularly, environment self-adaptive adjustment can be achieved, operation areas can be dynamically divided, and the detection reliability is improved. According to the invention, intelligent detection and linkage response of new energy station intrusion are realized, the false alarm rate is significantly reduced, the target detection precision and trajectory prediction accuracy are improved, the response time is shortened, the maintenance cost is reduced, the data security is enhanced, a security guarantee is provided for unattended operation, and the management efficiency and security are improved.

Description

Technical Field

[0001] The present invention relates to the field of new energy station security protection technology, and in particular to a new energy station intrusion detection method, equipment and medium based on linkage technology. Background Art

[0002] With the rapid development of the new energy industry and the continuous expansion of wind farm substations, traditional safety protection models are gradually exposing technical bottlenecks. While the Jingmen Xianghe Wind Farm Phase II facility is capable of operating with minimal personnel, it lacks key technologies such as video alarm linkage, perimeter intrusion alarms, and automatic fire extinguishing. This results in equipment inspections relying on manual round-trips, delayed fault detection, and the inability to meet the real-time and reliability requirements of "unmanned operation." Furthermore, traditional temperature and smoke sensors only trigger alarms in the late smoldering stages of a fire and have detection blind spots. Manual inspections, limited by the number of personnel and expertise, struggle to cover key areas such as high-voltage switchgear and cable trenches, resulting in inefficient troubleshooting of equipment hazards.

[0003] Under the dual pressures of personnel safety and production efficiency, wind farm management urgently needs to transition to an "unmanned" model. Existing technologies, electronic fencing and video surveillance operate independently, resulting in high false alarm rates for intrusion detection. Manual impression taking and static analysis fail to capture mandibular motion, leading to design errors in orthodontic plans. Equipment maintenance relies on scheduled inspections and lacks predictive maintenance methods based on real-time data, resulting in a high risk of unplanned downtime. Furthermore, the lack of localized software and hardware in traditional systems makes it difficult to meet the power industry's requirements for independent control. This is especially true in the large-scale operation of new energy sites, where the need for intelligent and localized security protection is increasingly urgent. Summary of the Invention

[0004] The present invention proposes a new energy station intrusion detection method, device and medium based on linkage technology to solve the problems mentioned in the above-mentioned prior art.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A new energy station intrusion detection method based on linkage technology includes the following steps: S1: Multi-source heterogeneous data collection steps: Deploy a six-wire electronic pulse fence around the perimeter of the new energy station to issue an alarm when the system is short-circuited or disconnected; install a four-beam infrared beam detector to cover the walls and gates; deploy visible light dome cameras and card cameras to collect physical intrusion signal data around the perimeter in real time and transmit it to the data processing center; S2: Intrusion signal fusion step: perform spatiotemporal synchronization calibration on the pulse alarm signal of the electronic fence, the beam blocking coordinates of the infrared beam, and the video frame timestamp; use the DS evidence theory to fuse multi-source data, the formula is , where m(A) is the basic probability assignment of the invasion event after fusion, For the i-th sensor pair Proposition A j The probability assignment is n, the number of sensor categories, represents the probability of the i-th sensor judging an impossible event; S3: Dynamic trajectory prediction and behavior recognition steps: Use the YOLOv5 algorithm to build an intrusion target detection model to identify foreign objects, and use the Kalman filter algorithm to predict the target motion trajectory. The formula is: ,in is the optimal estimate at time k, is the predicted value at time k, z k is the observed value, H k is the observation matrix, K k It is a Kalman gain, combined with the I3D network to analyze target behavior and distinguish normal operations from illegal intrusions; S4: Multi-level linkage response steps: When an intrusion is detected, a three-level response is triggered: In the first level response, the visible light camera automatically zooms to the target area within 2 seconds to capture an image and broadcasts a voice warning; in the second level response, the sound and light alarm is activated and the alarm information is pushed to the security personnel; in the third level response, for continuous intrusion, the grounding switch of the electronic fence is closed within 5 seconds to form a short-circuit alarm loop, and the entrance and exit access of the station is blocked; S5: Data backtracking and model optimization steps: Use blockchain technology to store intrusion event data for ≥90 days; regularly use historical data to train the intrusion detection model, and update the YOLOv5 weight parameters through transfer learning.

[0006] Furthermore, it also includes environmental adaptive adjustment steps: deploying temperature and humidity sensors and anemometers. When the ambient humidity is greater than 95% RH or the wind speed is greater than 10m / s, the infrared radiation device beams are automatically increased to 6 beams, and the electronic fence pulse frequency is increased from 1 time / second to 2 times / second.

[0007] Furthermore, it also includes the step of dynamically dividing the working area: through electronic fence zoning control, combined with the maintenance work plan, the safe working area is dynamically delineated on the GIS map; when the operator wearing an RFID work badge enters the unauthorized area, the intrusion alarm is triggered and the high-voltage output of the electronic fence in the area is blocked to avoid the risk of accidental touch.

[0008] Furthermore, in the multi-source heterogeneous data collection step, the electronic fence adopts an AC220V and DC12V dual power supply redundant design and is equipped with a power-off memory module to ensure that the original deployment status is automatically retained after power is restored.

[0009] Furthermore, in the dynamic trajectory prediction and behavior recognition steps, knowledge distillation technology is used to compress the YOLOv5 model parameters, so that the model can be inferred in real time on an edge computing node with a computing power ≥ 2TOPS.

[0010] Furthermore, in the multi-level linkage response step, the sound and light alarm is powered by solar energy and integrated with a GPS positioning module to facilitate rapid positioning of faulty equipment.

[0011] Furthermore, in the data backtracking and model optimization steps, the Kalman filter parameters are adjusted online using the stochastic gradient descent algorithm using the feature vector of the intrusion event, so that the trajectory prediction error is reduced to ≤0.5m.

[0012] Furthermore, the following modules are included: Perception layer module: Electronic pulse fence unit: six-wire front-end fence, equipped with pulse host, RS-485 communication module and IP43 protective chassis, supporting power-off memory function; Infrared radiation unit: Four-beam probe, built-in temperature compensation circuit, support margin coverage; Video acquisition unit: Deploys visible light dome cameras and card cameras, adopts H.265 video compression technology, and supports 25fps frame rate recording; Data processing layer module: Edge computing node: equipped with a lightweight YOLOv5 model and Kalman filter algorithm, equipped with two USB3.0 interfaces and one eSATA interface, supports simultaneous analysis of 8-channel video streams, and realizes target detection and trajectory prediction; Data fusion engine: Deploys the DS evidence theory algorithm to perform time synchronization and confidence calculation on multi-sensor data; Linkage control layer module: Sound and light alarm unit: equipped with LED flashing lights and high-decibel speakers, supporting remote control and local triggering, with a coverage area of ​​≥500㎡; Actuator unit: deploys electronic fence grounding switches and access controllers, and supports linkage with video surveillance systems; Interactive interface: 7-inch touch screen, real-time display of GIS map, alarm log and equipment status, integrated mobile phone APP remote monitoring function; Self-calibration maintenance module: Deploys a laser calibration device to automatically calibrate the distance between geo-fence conductors and the alignment of infrared beams every week, and equips a cleaning robot to regularly clean dust from the sensor surface; Adaptive learning module: Built-in online learning algorithm, real-time analysis of the feature vector of the intrusion event, automatic optimization of the process noise covariance Q and observation noise covariance R of the Kalman filter, and improved trajectory prediction accuracy.

[0013] Furthermore, in the perception layer module, the visible light ball camera supports regional intrusion detection and cross-border detection intelligent functions, with a signal-to-noise ratio of ≥52dB and a fill light illumination distance of ≥50m.

[0014] Furthermore, in the data processing layer module, the edge computing nodes use domestically produced independently controllable hardware, and the operating system is the Kylin / Tongxin domestic system, which meets the safety requirements of the power industry.

[0015] Compared with the existing technology, the beneficial effects of the present invention are: Intelligent inspection technology utilizes micro-track robots and infrared temperature measurement equipment, combined with AI algorithms to intelligently identify equipment defects and human behavior. This reduces inspection efficiency from 15 minutes manually to 20 seconds, addressing the issue of inadequate inspections due to insufficient O&M personnel. Fire monitoring utilizes an infrared dual-view card camera and a new fire extinguishing device to provide early warning of thermal degradation and remotely activate the fire extinguishing device, eliminating the lag of traditional heat and smoke sensors during the smoldering phase and improving initial fire response capabilities.

[0016] Perimeter security utilizes electronic pulse fencing, infrared beams, and video surveillance. Unauthorized intrusions are quickly triggered by audible and visual alarms and access control locks, with a response time of ≤5 seconds, effectively preventing sabotage of power equipment. The system integrates multi-source data and intelligent algorithms, supports data integration with mainstream orthodontic software, and automatically generates reports containing 3D angle data, providing a quantitative basis for treatment planning and reducing the cost of repeated inspections. Furthermore, the system utilizes domestically produced, independently controllable hardware and software (such as Inspur servers, Loongson processors, and the Kylin operating system) to meet power industry safety standards, enhancing system reliability and localization.

[0017] Through full-process intelligent management and control, this application has achieved the leap from "few people on duty" to "unmanned" in wind farm substations. It not only reduces labor costs and improves production efficiency, but also greatly reduces the risk of safety accidents through early warning and intelligent handling, providing a replicable technical template for the large-scale and intelligent development of new energy stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic block diagram of the new energy station intrusion detection method based on linkage technology proposed in the present invention; Figure 2 This is a schematic block diagram of the new energy station intrusion detection system based on linkage technology proposed by the present invention; Figure 3 This is a bar chart comparing the accuracy of traditional manual inspection and this intelligent inspection; Figure 4 The following is a line chart comparing the accuracy of perimeter intrusion identification between the traditional solution and this solution. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like to indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0021] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the said features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be a connection between the two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below with reference to the accompanying drawings.

[0022] Reference Figures 1 to 4 A new energy station intrusion detection method based on linkage technology is characterized by comprising the following steps: Multi-source heterogeneous data collection steps: Deploy six-wire electronic pulse fences, four-beam infrared radiation devices, visible light dome cameras and card cameras at the perimeter of new energy stations to collect physical intrusion signal data around the perimeter in real time and transmit it to the data processing center.

[0023] A six-wire electronic pulse fence is installed atop the substation perimeter wall. The pulse voltage of each conductor is set to 5kV to 10kV in high-voltage mode and 1kV to 2kV in low-voltage mode. Switching between high and low voltage is accomplished via an EPF-6L pulse controller, with a switching time of less than one second. The fence conductors are constructed of 2.5mm diameter 304 stainless steel, with 15cm spacing between conductors. The length of a single protection zone is limited to 50m. In the event of an open or short circuit, the pulse controller sends an alarm signal via the RS-485 bus at a baud rate of 9600bps. Leveraging hardware spacing and signal transmission delay algorithms, the controller achieves a positioning accuracy of ±0.5m.

[0024] The PA-120 four-beam infrared probe has a nominal detection range of 120m and can provide 100% coverage of a 60m long fence. The probe spacing should not exceed 50m. The probe has a built-in temperature compensation circuit and operates stably in environments ranging from -30°C to +65°C. An alarm is triggered if any two or more beams are blocked. Optical alignment technology and signal processing algorithms ensure coordinate error is kept to within 1°.

[0025] The VS-200 visible light dome camera and the VC-100 compact camera were deployed. The dome camera features a 2-megapixel resolution, 23x optical zoom, IP66 protection, 360° rotation, 300 preset positions, and a zoom speed of as fast as 2 seconds. The compact camera offers a 1080P resolution and IP67 protection. The dome camera's video stream is compressed using H.265 encoding and transmitted to edge computing nodes via a 5G network. By optimizing network bandwidth and encoding efficiency, transmission latency is kept below 200ms.

[0026] Intrusion signal fusion steps: perform spatiotemporal synchronization calibration on the pulse alarm signal of the electronic fence, the beam blocking coordinates of the infrared beam, and the video frame timestamp; use the DS evidence theory to fuse multi-source data, the formula is , where m(A) is the basic probability assignment of the invasion event after fusion, For the i-th sensor pair Proposition A j The probability assignment is n, the number of sensor categories, The probability value assigned to the event that sensor type i determines is impossible. Time synchronization is performed on the three sensor data types using edge computing nodes (Inspur NF5280M5 servers and Loongson 3C5000 processors): geofence alarm signal timestamp accuracy is 1ms, infrared beam occlusion event time error is ≤5ms, and video frame (25fps) timestamp error is ≤40ms. A unified spatiotemporal coordinate system (error ≤10ms) is constructed using the Hardware Time Protocol (PTP) and software interpolation algorithms.

[0027] Assume that electronic fence (E), infrared radiation (I), and video detection (V) are independent evidence sources, and the basic probability assignments for intrusion event A are , non-intrusion events are assigned The fusion formula is: Here, m(A) represents the basic probability of an intrusion event after fusion, and the denominator is the conflict coefficient, which is used to eliminate data conflicts. When m(A) ≥ 0.8, it is considered a valid intrusion. Field measurements show that the false alarm rate has dropped from 5 per day in traditional solutions to 0.1 per day.

[0028] Dynamic trajectory prediction and behavior recognition steps: The edge computing node runs a lightweight YOLOv5s model to perform real-time detection on the video stream, identifying targets such as people, animals, and kites. Detection latency is ≤ 200ms (determined by the model's lightweightness and hardware computing power). The Kalman filter algorithm is used to update the target state, and the state equation is: ,in, is the optimal estimated value of the target state at time k (including parameters such as position and speed); is the target state prediction value at time k, estimated by the previous moment Calculated by the state transfer matrix F; The target coordinates (observation values) detected in the video at time k; is the observation matrix, which is used to map the state space to the observation space; is the Kalman gain, and the calculation formula is ,in is the prediction covariance, and R is the observation noise covariance. For targets moving at speeds ≤5 m / s, the prediction error is ≤0.8 m (determined by sensor accuracy and the number of algorithm iterations). The I3D network analyzes the target's motion features for 10 consecutive frames, extracting spatiotemporal convolutional feature vectors that are then input into the fully connected layer for classification (normal operation / illegal intrusion).

[0029] Multi-level linkage response steps: When an intrusion is detected, three levels of response are triggered according to the risk level.

[0030] Level 1 Response (Suspected Intrusion): When the video detection target confidence level reaches 0.7 or above, the visible light dome camera immediately activates its automatic zoom function, focusing on the target area within 2 seconds and capturing a 1920×1080 HD image. Simultaneously, the station broadcasts a preset voice warning at 85dB volume. The entire process from detection to response takes no more than 3 seconds, quickly alerting potential threats.

[0031] Level 2 Response (Confirmed Intrusion): When the theoretical support m(A) of the evidence, derived from multi-source data fusion, exceeds 0.8, the system determines an intrusion is confirmed. The electronic fence immediately switches to 10kV high-voltage mode to enhance deterrence. The audible and visual alarms activate simultaneously, with the LED flashing at a strong 1500cd intensity and the speaker emitting a 100dB siren. The system pushes an alarm message containing GIS coordinates and the zone number to the security personnel's mobile phone via the 5G network. The entire process is completed within 5 seconds, ensuring personnel are promptly aware of the danger.

[0032] Level 3 Response (Continued Intrusion): If the intrusion persists for more than 30 seconds, the system enters the highest level of response. It automatically closes the grounding switch of the electronic fence in the intrusion area, generating a short-circuit alarm with a 30ms short-circuit tripping speed. The access control system is also activated, blocking the nearest entrance and exit within 1 second, cutting off the intrusion path. The video surveillance automatically records the data, which is encrypted and uploaded to the cloud in real time. The entire response, from detection to completion, takes less than 8 seconds, effectively blocking the intrusion and preserving evidence.

[0033] Data backtracking and model optimization steps: Blockchain technology is used to perform SHA-256 hash storage on alarm data (including sensor signals, video clips, and response records). Each block contains the hash value of the previous block, a timestamp, and a data summary to ensure that the data cannot be tampered with. The storage period is ≥ 90 days (meeting the safety audit requirements of the power industry).

[0034] Every month, 100,000 images (including over 5,000 intrusion samples) are randomly sampled from historical data. Transfer learning is used to fine-tune the YOLOv5 model, focusing on optimizing the accuracy of detecting foreign objects (such as kites and bird nests). Furthermore, intrusion trajectory data is used to update the Kalman filter parameters Q (process noise covariance) and R (observation noise covariance), reducing the prediction error from 0.8m to 0.5m.

[0035] The present invention also includes an environmental adaptive adjustment step: achieving environmental adaptive adjustment through multi-type sensor data collection and an intelligent decision-making system. The deployed temperature and humidity sensors utilize high-precision digital sensor modules based on temperature-sensitive resistors and capacitive humidity sensing elements, coupled with a 16-bit ADC chip to achieve precise temperature measurements with an accuracy of ±0.5°C and humidity with an accuracy of ±5%RH. The anemometer utilizes a three-cup mechanical structure, converting cup rotation signals into digital pulses via a photoelectric encoder. Combined with a microprocessor's frequency-to-speed conversion algorithm, this achieves a resolution of 0.1 m / s.

[0036] When the environmental monitoring module receives abnormal data indicating humidity >95% RH or wind speed >10m / s, the system triggers a three-level response mechanism. First, it sends a command to the infrared beam transmitter via the Modbus RTU protocol, activating the redundant beam transmitter and expanding the original four infrared beams into a six-beam three-dimensional protection array. The spacing between adjacent beams is reduced to 15cm, forming an interlaced protection network, effectively reducing the probability of false beam triggering caused by water mist, sand and dust. Second, the electronic fence controller uses PWM pulse modulation technology to increase the pulse frequency from 1 time / second to 2 times / second, while maintaining a safe voltage range of 5000-10000V, using short-duration high-frequency pulses to enhance intrusion detection sensitivity. Finally, the system uses the Kalman filter algorithm to perform real-time noise reduction on the sensor data, combines historical environmental data to establish a Bayesian prediction model, and dynamically adjusts the alarm threshold to ensure that the false alarm rate is stably controlled at less than 0.1 times / day in extreme environments.

[0037] The present invention also includes a step for dynamically demarcating work areas: the electronic fence utilizes a segmented pulse host, enabling refined control of individual defense zones ≤ 50m via the RS485 bus. Each zone has a built-in dual-path signal verification module, supporting multiple alarm detection types, such as disconnection, short circuit, and network contact. The system connects to the station's operations and maintenance management platform to acquire maintenance work order data, including operator information, time range, and work area. Using a coordinate conversion algorithm, this work order area information is mapped to a GIS map to generate dynamic safe work areas.

[0038] For personnel location, RFID badges integrate UHF chips, which, in conjunction with RFID readers deployed at key nodes on the site, achieve centimeter-level positioning accuracy. If a person enters an unauthorized area, the system uses a multi-source data fusion algorithm (RFID positioning data + geo-fence alarm signals) for secondary confirmation to avoid false alarms. Once an intrusion is confirmed, the system immediately implements dual protection measures: First, a command is sent to the geo-fence controller in the corresponding area, which disconnects the high-voltage output circuit via a solid-state relay, retaining the alarm function. Second, the intrusion location is marked on a GIS map in real time, and the video surveillance system is linked to retrieve on-site footage. AI visual analysis technology verifies the intrusion and simultaneously sends a graphic alarm message to the security center.

[0039] In the multi-source heterogeneous data acquisition step described in this invention, the dual-power redundant system utilizes a relay switching circuit. AC220V mains power, after EMI filtering, bridge rectification, and DC-DC conversion, powers the system and simultaneously charges a DC12V backup lithium battery pack. When a mains power outage is detected, the relay switches to battery power within 5ms, ensuring uninterrupted system operation. The power-off memory module, based on a ferroelectric random access memory (FRAM) design, uses a dedicated power monitoring chip to monitor system voltage in real time. When the voltage drops below a threshold, the current deployment status (including the zone activation list, alarm thresholds, and pulse parameters) is immediately written to the FRAM in 100μs or less. After the system restarts, the main controller reads the FRAM data via the I2C bus, restoring the system to its pre-power-off state within 3 seconds. The power management unit also integrates overvoltage protection (clamping voltage 30V), overcurrent protection (current limit 2A), and surge suppression (8 / 20μs waveform, nominal discharge current 5kA) circuitry to ensure stable operation in complex power grid environments.

[0040] In the dynamic trajectory prediction and behavior recognition steps described in this invention, knowledge distillation technology is used to lightweight the YOLOv5 model. Using YOLOv5s as the teacher model, a student model is designed. Through feature map distillation (using the L2 loss function) and classification probability distillation (temperature parameter T=4), the model parameters are compressed from 7.2M to 4.3M, reducing the number of floating-point operations (FLOPs). The quantization process uses 8-bit integer precision, combined with the TensorRT inference engine for operator fusion and memory optimization, enabling the model to achieve 30 FPS real-time inference on edge computing nodes with 2TOPS computing power, such as the NVIDIA Jetson Nano. The behavior recognition algorithm utilizes a spatiotemporal feature fusion architecture, feeding object detection results from five consecutive frames into an LSTM network to extract motion trajectory features. The system also implements a dynamic ROI (region of interest) adaptive adjustment mechanism, automatically adjusting the detection area based on scene complexity, further reducing computational overhead.

[0041] In the multi-level linkage response step described in this invention, the energy management system of the sound and light alarm utilizes maximum power point tracking (MPPT) technology. The solar panel output is converted via a step-up DC-DC converter. The energy storage unit utilizes a lithium iron phosphate battery pack with a capacity designed to meet 72 hours of operation in continuous rainy weather. The battery management system (BMS) integrates overcharge protection (charge cut-off voltage 3.65V / cell), over-discharge protection (discharge cut-off voltage 2.5V / cell), and balancing circuits. The GPS positioning module utilizes a multi-frequency, multi-constellation receiver (supporting GPS L1 / L5 and Beidou B1I / B2I). Utilizing RTK differential positioning technology, it achieves positioning accuracy of 2cm in open environments and ≤5m in urban canyon environments. The alarm signal utilizes FSK modulation, operates at a frequency of 433MHz, has a transmission power of 100mW, and has a communication range of ≥1km. The device also integrates an accelerometer and microphone, supporting vibration alarm and voice broadcast functions, and can switch alarm modes via remote commands.

[0042] In the data backtracking and model optimization steps of this invention, the system first establishes a 12-dimensional state vector containing the target's position, velocity, and acceleration, and predicts the next position using the state transition matrix. During the measurement update phase, the error between the actual detected and predicted positions is used to dynamically adjust the Kalman filter's process noise covariance matrix Q and measurement noise covariance matrix R using the SGD algorithm. Specifically, the feature vectors of historical intrusion events (speed range 0.5-8 m / s, turn angle ≤ 120° / s) are input into a neural network, which outputs the optimal parameter combination. During the online optimization process, the learning rate is set to 0.01, and Mini-Batch SGD (batch size 32) is used, with parameters updated every 50 new events. Experimental data shows that the optimized trajectory prediction error is ≤0.3m in straight-line motion scenarios and ≤0.5m in turning scenarios. The system also implements an anomaly detection mechanism that automatically triggers model retraining when the prediction error exceeds 1m for five consecutive frames.

[0043] The present invention includes the following modules: Perception layer module: The six-wire fence adopts a 15cm equidistant conductor layout, and the length of a single defense zone is strictly controlled within 50m to ensure uniform distribution of the electric field. The pulse host supports 200m RS-485 bus cascade and uses the ModbusRTU protocol to achieve multi-machine communication. The protective chassis is IP43 rated, equipped with UV-proof ABS shell and silicone sealing ring, which can resist solid foreign objects with a diameter of ≥1mm and water splashing from any direction. The power supply system uses AC220V mains as the main power supply, which is powered by EMI filtering and bridge rectification. The system is powered by a DC12V lead-acid battery as a backup power supply with a battery life of ≥8 hours. The power-off memory module uses the AT24C02EEPROM chip, which immediately writes the deployment status when a voltage fluctuation is detected, and the data retention life is ≥10 years; The four-beam sensor uses the PA-120 model, employing digital pulse phase modulation technology with a transmit power of ≥50mW. The receiver features a built-in automatic gain control (AGC) circuit. In rainy or foggy weather, the ambient light sensor triggers a signal enhancement mode, automatically increasing receiver sensitivity. The device also features a built-in temperature compensation circuit, ensuring a false trigger rate of ≤0.05 times per day in ambient temperatures of -25°C to +65°C. The mounting bracket, constructed of hot-dip galvanized steel, stands 2.5m tall. The center of the beam is 1.8m above the ground, creating a detection zone with a vertical height of 0.8m, ensuring that at least two beams are blocked when a person crosses it.

[0044] The visible light dome camera is equipped with a 1 / 1.8" ProgressiveScan CMOS sensor, supports H.265+ encoding, and boasts a signal-to-noise ratio of ≥52dB. Its fill illumination system utilizes an array of white LEDs with an effective illumination range of ≥50m and supports automatic day / night switching (using an ICR dual filter). Its preset positioning accuracy is ≤0.1°, and it supports 360° continuous rotation and a -20° to 90° pitch. The compact camera uses an M12 macro lens with a 4mm focal length, a working distance of 10cm to 50cm, an F1.8 aperture, and supports 1080p video capture at 60fps. With an IP67 rating and a fully sealed design featuring a built-in moisture-proof silicone layer, it is suitable for close-range inspection scenarios such as fence nodes and terminal boxes.

[0045] The data processing layer module, centered around edge computing nodes, utilizes the Inspur NF5280M5 server, powered by a Loongson 3C5000 processor with a quad-core architecture and a clock speed of 2.5GHz. Equipped with 8GB of DDR4 memory and a 1TB SSD, it supports parallel analysis of eight 1080P video streams and features an expandable GPU acceleration module via a PCIe 3.0 interface. The system utilizes a modular design with two reserved PCIe slots and four SATA ports for future feature upgrades.

[0046] The software architecture is built on the Tongxin server operating system and utilizes a microservices architecture to enhance system stability. Core algorithms include YOLOv5 target detection, Kalman filter trajectory prediction, and the DS evidence theory decision model. Developed in C++ and incorporating OpenMP multi-threading technology, the YOLOv5 model is optimized using the TensorRT inference engine, achieving real-time latency of ≤200ms. The system also features a built-in automatic update mechanism that remotely pushes algorithm upgrades to ensure detection accuracy.

[0047] The data fusion engine ensures data consistency through a time synchronization module, achieving nanosecond-level clock synchronization across multiple sensors based on the IEEE 1588 protocol. A Lagrange interpolation algorithm is used to compensate for network transmission delays, keeping time errors to ≤10ms. The confidence calculation module employs a weighted fusion strategy, dynamically assigning weights based on sensor characteristics: 0.4 for geofence, 0.3 for infrared radiation, and 0.3 for video detection.

[0048] Linkage control layer module: The sound and light alarm unit utilizes a modular design. The LS-100 LED strobe light features a built-in constant-current drive circuit, a stable 2Hz flash frequency, and an effective illumination distance of up to 50m. PWM dimming technology allows for adjustable flash intensity from 500cd to 1500cd, and supports remote command control. The SA-120 high-decibel speaker features a built-in digital amplifier chip, an output power of 120W, and a coverage area of ​​500m² or greater. It supports SD card storage for MP3-formatted voice files, receives broadcast commands via an RS485 interface, and has a volume adjustment range of 60dB to 100dB. It also includes speech synthesis functionality for real-time alarm tone generation.

[0049] The actuator unit's electronic fence grounding switch utilizes a DC electromagnetic design with a built-in bistable coil, a tripping time of ≤30ms, and silver alloy contacts with a lifespan of ≥100,000 cycles. It supports both remote control and local manual operation, and features status feedback, reporting switch position in real time via auxiliary contacts. The access control controller, an AC-01 model, supports dual authentication using ISO14443 RFID cards and QR codes. A built-in high-definition camera, linked to the video surveillance system, automatically captures images of individuals when a card is swiped and uploads them to the management platform, with a response time of ≤1 second.

[0050] The local touch screen of the interactive interface uses a 7-inch capacitive touch screen with a resolution of 1024×600 and supports ten-point touch. The screen uses a tempered glass panel with a surface hardness of ≥7H and a protection level of IP65. Developed based on the Linux kernel, the system displays GIS maps, sensor status, alarm history and other information in real time, supports touch-based arming / disarming operations, and has an operation delay of ≤500ms. The mobile app is developed based on the Android / iOS platform and adopts the MVVM architecture design. It supports real-time video preview, alarm push, device control and other functions. Data transmission uses the SSL / TLS1.3 encryption protocol, with a response delay of ≤2 seconds, and supports offline operation and message caching.

[0051] Self-calibration maintenance module: Laser calibration device: The laser calibration device is deployed at both ends of the electronic fence, and the calibration procedure is automatically performed every week. A semiconductor laser with a wavelength of 650nm is used to detect the spacing and verticality of the conductors through the triangulation principle, with a measurement accuracy of ±1mm. When a spacing deviation of >2mm is detected, the control system drives the stepper motor to adjust the fence bracket. The stepper motor adopts subdivision drive technology with a step angle of 1.8°. Micron-level adjustment is achieved through a precision transmission mechanism, and the calibration error is ≤0.5%. The infrared beam probe has a built-in laser aligner, which automatically emits a visible laser beam during calibration. The beam offset is detected by the PSD position sensor, and the angle of the transmitting end is dynamically adjusted in combination with the servo control system to ensure that the parallelism of the four beams is ≤0.1°.

[0052] The CR-01 intelligent cleaning robot utilizes a tracked mobile platform equipped with a brushless DC motor and a climbing capability of ≤30°. Its navigation system integrates ultrasonic sensors and lidar for 360-degree obstacle avoidance. The cleaning system is equipped with a micro vacuum cleaner with a maximum suction force of 5000Pa and a three-stage filtration design, effectively collecting particles larger than 0.3μm. It also features a soft-bristle cleaning assembly and a dual-axis robotic arm for multi-angle cleaning. The robot automatically initiates a cleaning program every two weeks, deep-cleaning the sensor surface along a preset route. A single cleaning cycle takes approximately two hours. After cleaning, a transmittance sensor is used to verify the optical sensor's transmittance. The system supports remote monitoring of cleaning progress and includes fault self-diagnosis and alarm functions.

[0053] Adaptive learning module: The online learning algorithm uses the stochastic gradient descent (SGD) algorithm to monitor the intrusion event feature vector (such as target speed v, intrusion path angle \(\theta\), and residence time t) in real time, and dynamically adjust the Kalman filter parameters: when \(v>3m / s\), increase the process noise covariance Q (such as from 10 -4 Increase to\(5×10 -4 ), improve tracking sensitivity; when the video detection confidence c < 0.7, increase the observation noise covariance R (such as from 10 -3Increased to 8×10 -3 ), reducing the observation weight.

[0054] Behavior recognition optimization relies on a long short-term memory (LSTM) network to build a predictive model. The system conducts deep learning on historical intrusion path data and integrates it with geographic information systems (GIS) to generate risk heat maps. For high-risk areas such as fence corners and equipment blind spots, the detection confidence threshold is automatically increased from 0.7 to 0.8.

[0055] In the perception layer module of this invention, the visible light dome camera uses a 1 / 1.8" CMOS sensor (2 megapixels), supports H.265+ encoding and 3D-DNR noise reduction, and has a signal-to-noise ratio of ≥52dB. Regional intrusion detection, based on a deep learning algorithm, supports eight custom polygonal areas and three levels of sensitivity adjustment; cross-border detection supports the setting of four bidirectional warning lines. The fill light system uses an array of infrared and white light, with an infrared fill light range of 100m and a white light range of 50m, and automatic ICR filter switching. A built-in three-axis gyroscope provides electronic image stabilization and 360° horizontal rotation (maximum speed of 240° / s).

[0056] In this invention, the edge computing nodes in the data processing layer module utilize an Inspur NF5280M6 server and a Loongson 3A5000 processor (quad-core, 1.8GHz), 8GB of ECC memory, and a 256GB SSD. They are equipped with the Kirin V10SP1 operating system and meet Level 3 security requirements. They integrate a proprietary and controllable inference engine and support TensorFlow model conversion. They utilize a microservices architecture (Docker + Kubernetes), support SM4 encryption (128-bit key), and comply with the DL / T860 communication protocol.

[0057] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A new energy station intrusion detection method based on linkage technology, characterized in that: The following steps are involved: S1: Multi-source heterogeneous data collection steps: Deploy a six-wire electronic pulse fence around the perimeter of the new energy station to issue an alarm when the system is short-circuited or disconnected; install a four-beam infrared beam detector to cover the walls and gates; deploy visible light dome cameras and card cameras to collect physical intrusion signal data around the perimeter in real time and transmit it to the data processing center; S2: Intrusion signal fusion step: perform spatiotemporal synchronization calibration on the pulse alarm signal of the electronic fence, the beam blocking coordinates of the infrared beam, and the video frame timestamp; use the DS evidence theory to fuse multi-source data, the formula is , where m(A) is the basic probability assignment of the invasion event after fusion, For the i-th sensor pair Proposition A j The probability assignment is n, the number of sensor categories, represents the probability assignment of the event judged as impossible by the i-th sensor; S3: Dynamic trajectory prediction and behavior recognition steps: Use the YOLOv5 algorithm to build an intrusion target detection model to identify foreign objects, and use the Kalman filter algorithm to predict the target motion trajectory. The formula is: ,in is the optimal estimate at time k, is the predicted value at time k, z k is the observed value, H k is the observation matrix, K k It is a Kalman gain, combined with the I3D network to analyze target behavior and distinguish normal operations from illegal intrusions; S4: Multi-level linkage response steps: When an intrusion is detected, a three-level response is triggered: In the first level response, the visible light camera automatically zooms to the target area within 2 seconds to capture an image and broadcasts a voice warning; in the second level response, the sound and light alarm is activated and the alarm information is pushed to the security personnel; in the third level response, for continuous intrusion, the grounding switch of the electronic fence is closed within 5 seconds to form a short-circuit alarm loop, and the entrance and exit access of the station is blocked; S5: Data backtracking and model optimization steps: Use blockchain technology to store intrusion event data for ≥90 days; regularly use historical data to train the intrusion detection model, and update the YOLOv5 weight parameters through transfer learning.

2. The new energy station intrusion detection method based on linkage technology according to claim 1 is characterized in that: It also includes environmental adaptive adjustment steps: deploying temperature and humidity sensors and anemometers. When the ambient humidity is greater than 95% RH or the wind speed is greater than 10m / s, the infrared beam of the device is automatically increased to 6 beams, and the pulse frequency of the electronic fence is increased from 1 time / second to 2 times / second.

3. The new energy station intrusion detection method based on linkage technology according to claim 1 is characterized in that: It also includes the step of dynamically dividing the working area: through electronic fence zoning control, combined with the maintenance work plan, the safe working area is dynamically delineated on the GIS map; when the operator wearing the RFID work badge enters the unauthorized area, the intrusion alarm is triggered and the high-voltage output of the electronic fence in the area is blocked to avoid the risk of accidental touch.

4. The new energy station intrusion detection method based on linkage technology according to claim 1 is characterized in that: In the multi-source heterogeneous data collection step, the electronic fence adopts an AC220V and DC12V dual power supply redundant design and is equipped with a power-off memory module to ensure that the original deployment status is automatically retained after power failure is restored.

5. The new energy station intrusion detection method based on linkage technology according to claim 1 is characterized in that: In the dynamic trajectory prediction and behavior recognition steps, knowledge distillation technology is used to compress the YOLOv5 model parameters, so that the model can be inferred in real time on edge computing nodes with a computing power of ≥2TOPS.

6. The new energy station intrusion detection method based on linkage technology according to claim 1 is characterized in that: In the multi-level linkage response step, the sound and light alarm is powered by solar energy and integrated with a GPS positioning module to facilitate rapid positioning of faulty equipment.

7. The new energy station intrusion detection method based on linkage technology according to claim 1 is characterized in that: In the data backtracking and model optimization steps, the Kalman filter parameters are adjusted online using the stochastic gradient descent algorithm using the feature vector of the intrusion event, so that the trajectory prediction error is reduced to ≤0.5m.

8. A system using the new energy station intrusion detection method based on linkage technology according to any one of claims 1 to 7, characterized in that: Includes the following modules: Perception layer module: Electronic pulse fence unit: six-wire front-end fence, equipped with pulse host, RS-485 communication module and IP43 protective chassis, supporting power-off memory function; Infrared radiation unit: Four-beam probe, built-in temperature compensation circuit, support margin coverage; Video acquisition unit: Deploys visible light dome cameras and card cameras, adopts H.265 video compression technology, and supports 25fps frame rate recording; Data processing layer module: Edge computing node: equipped with a lightweight YOLOv5 model and Kalman filter algorithm, equipped with two USB3.0 interfaces and one eSATA interface, supports simultaneous analysis of 8-channel video streams, and realizes target detection and trajectory prediction; Data fusion engine: Deploys the DS evidence theory algorithm to perform time synchronization and confidence calculation on multi-sensor data; Linkage control layer module: Sound and light alarm unit: equipped with LED flashing light and high-decibel speaker, supports remote control and local triggering, and covers an area of ​​≥500㎡; Actuator unit: deploys electronic fence grounding switches and access controllers, and supports linkage with video surveillance systems; Interactive interface: 7-inch touch screen, real-time display of GIS map, alarm log and equipment status, integrated mobile phone APP remote monitoring function; Self-calibration maintenance module: Deploys a laser calibration device to automatically calibrate the distance between geo-fence conductors and the alignment of infrared beams every week, and equips a cleaning robot to regularly clean dust from the sensor surface; Adaptive learning module: Built-in online learning algorithm, real-time analysis of the feature vector of the intrusion event, automatic optimization of the process noise covariance Q and observation noise covariance R of the Kalman filter, and improved trajectory prediction accuracy.

9. The system of the new energy station intrusion detection method based on linkage technology according to claim 8 is characterized in that: In the perception layer module, the visible light ball camera supports regional intrusion detection and cross-border detection intelligent functions, with a signal-to-noise ratio ≥52dB and a fill light illumination distance ≥50m.

10. The system of the new energy station intrusion detection method based on linkage technology according to claim 8 is characterized in that: In the data processing layer module, the edge computing nodes use domestically produced independently controllable hardware, and the operating system is the Kylin / Tongxin domestic system, which meets the safety requirements of the power industry.

Images