Power distribution network hot-line work safety distance detection system based on deep learning and geometric constraint

Abstract

The invention relates to the technical field of lithium ion batteries, and the system is characterized in that a binocular vision collection module is used for collecting left and right synchronous images in real time, and outputting an RAW format image stream to a geometric preprocessing unit; the light supplementing and environment sensing module is used for automatically supplementing light in a low-light or backlight environment and providing attitude compensation and environment exposure parameters; the edge calculation master control module is used for deploying a deep learning inference engine, completing semantic segmentation, parallax regression, deep reconstruction and safe distance calculation, and managing a peripheral bus and alarm output at the same time; the alarm and communication module is used for outputting a sound-light alarm signal when the safe distance threshold is detected to be broken through; uploading the distance sequence and the environment data to a cloud end through Ethernet / 4G; and the cloud is used for collecting a distance sequence, analyzing a long-term risk trend through LSTM-Analysis, and generating weekly reports.

Description

Technical Field

[0001] The present invention relates to the technical field of safe distance detection for live working in distribution networks, and in particular to a safe distance detection system for live working in distribution networks based on deep learning and geometric constraints. Background Art

[0002] With the rapid growth of distributed energy resources, large-scale electricity users, and electric transportation, the O&M workload for distribution networks (medium voltage levels of 6kV–35kV) has increased exponentially. To reduce power outage losses and improve power supply reliability, State Grid and provincial power companies are actively promoting live maintenance technology. Industry standards such as "SGCCQ / GDW1799.2-2013" and "DL / T789-2001" stipulate a minimum safe distance of 0.6m between live-line workers and 10kV live parts, and require monitoring systems to trigger audible and visual alarms with centimeter-level accuracy and sub-50ms response time.

[0003] Currently, common ranging solutions and their limitations are as follows: 1. Visual distance estimation: relies on manual experience, has low accuracy and is easily affected by environmental interference; 2. Electric field strength measurement: uses probes to measure the surrounding electric field, but is affected by electrode distribution and has large errors; 3. Beidou / UWB positioning: locates people and equipment in real time, is costly and is susceptible to occlusion and multipath interference; 4. Laser ranging: has high point-to-point accuracy, but has difficulty covering complex operating scenarios; 5. Stereo vision: is low-cost and has rich semantic information, but traditional methods have bottlenecks in multi-scale targets, metal reflections, and real-time performance.

[0004] Therefore, the development of a new safety distance detection system with high precision, strong robustness, low latency, and rapid on-site deployment has become an urgent need in the field of live operation and maintenance of distribution networks. Summary of the Invention

[0005] In response to the defects in the existing technology, the present invention provides a distribution network live working safety distance detection system based on deep learning and geometric constraints.

[0006] In order to achieve the above objectives, the technical solutions provided are as follows:

[0007] A distribution network live working safety distance detection system based on deep learning and geometric constraints, characterized by: including a binocular vision acquisition module, a fill light and environment sensing module, an edge computing main control module, an alarm and communication module, and a power management module;

[0008] The binocular vision acquisition module is used to capture left and right synchronous images in real time and output RAW format image streams to the geometric preprocessing unit; the fill light and environmental sensing module is used to automatically fill light in low-light or backlit environments, provide posture compensation and environmental exposure parameters; the edge computing main control module is used to deploy a deep learning inference engine to complete semantic segmentation, disparity regression, depth reconstruction and safe distance calculation, while managing the peripheral bus and alarm output; the alarm and communication module is used to output an audible and visual alarm signal when it detects that the safe distance threshold has been exceeded; the distance sequence and environmental data are uploaded to the cloud via Ethernet / 4G; the cloud is used to collect distance sequences, run LSTM-Anomaly to analyze long-term risk trends, and generate weekly reports.

[0009] The binocular vision acquisition module includes two 800×600@60FPS global shutter industrial cameras and a camera control board. The two 800×600@60FPS global shutter industrial cameras are connected to the camera control board through M12 aviation plugs. The camera control board is connected to the main control board (NPU+CPU) using a USB 3.0 interface based on ARM-Cortex-A53, and the baseline B=120mm.

[0010] The fill light and environment sensing module consists of a 4×4 visible light LED array, a constant current drive power supply and an I 2 C bus temperature and humidity, IMU sensor, LED driver and sensor are mounted on the 40-pin GPIO interface of the main control board.

[0011] The edge computing master module includes an embedded industrial computer (x86+RTX 4060Ti) with an Intel i7 CPU, 16GB of DDR4 RAM, and a 22MB INT8-quantized AMFN+JAM model. The system communicates with the NPU expansion card via PCIe.

[0012] The alarm and communication module includes an audible and visual alarm, which is connected to the main control board via a UART interface; the system also integrates RJ45 Ethernet, 4G module and CAN bus for remote monitoring and OTA upgrades.

[0013] The power management module receives 24V DC input and generates 12V, 5V and 3.3V through a DC-DC converter to power the camera, LED array, sensor and main control board respectively. The power management module has built-in overcurrent and overvoltage protection circuits.

[0014] A workflow of a distribution network live working safety distance detection system based on deep learning and geometric constraints, characterized by comprising the following steps:

[0015] Step 1: Initialization and calibration: After the system is powered on, read the internal and external parameters of the camera stored on the main control board, perform IMU attitude self-test and temperature and humidity calibration;

[0016] Step 2: Image preprocessing and geometric constraints: After the left and right images are distortion corrected and aligned, a search mask is generated using the fundamental matrix and epipolar constraints to reduce the matching candidate area to 28% of the original image.

[0017] Step 3: Semantic-geometric joint reasoning: The AMFN sub-network extracts multi-scale features, and the JAM module fuses channel and spatial attention to generate a segmentation map of key human parts; soft argmax regresses the sub-disparity probability distribution to obtain a single-value disparity map;

[0018] Step 4: Depth reconstruction and key point extraction: Calculate the depth map based on Depth = (f × B) / Disparity, and extract the 3D coordinates of 14 key points of the human body through the pre-trained pose estimation network;

[0019] Step 5: Safety distance calculation and alarm judgment: Construct a KD tree for the key points of the human body and the device contour points at the same time, search for the shortest Euclidean distance frame by frame, and trigger the sound, light, and voice alarm after "three-frame average" filtering when the distance is less than 0.6m;

[0020] Step 6: End-Cloud Collaboration and Data Management: Local distance series are uploaded to the cloud in real time. The LSTM-Anomaly model analyzes long-term trends and generates weekly reports. The differential upgrade package is pushed through the OTA pipeline.

[0021] The present invention proposes a comprehensive solution to the defects of the prior art, focusing on overcoming the following difficulties:

[0022] 1. Multi-scale recognition and matching: Key human body parts (hands, head, and feet) only appear in tens of pixels on the camera image plane, while arm-length insulating rods and wires span hundreds of pixels. These two must be accurately matched within the same deep reasoning framework.

[0023] 2. Feature loss under high-reflectivity and weak-texture conditions: Metal conductors and porcelain insulators reflect specularly under direct sunlight, and texture information is almost completely obscured by the saturated light spots, making it impossible for traditional stereo matching algorithms to establish reliable parallax.

[0024] 3. Real-time device-edge collaboration: Embedded smart terminals have limited computing power, and the entire device must complete the "collection-inference-alarm" closed loop in less than 35ms.

[0025] 4. Universal adaptability to complex environments: The system must maintain a detection accuracy of <±5.5% MAE in a variety of typical operating environments, including high altitude, strong electromagnetic fields, rain and fog, and nighttime supplemental lighting.

[0026] The beneficial effects of this invention are: through experimental results and economic analysis, this system significantly outperforms traditional solutions in terms of accuracy, timeliness, robustness, lifecycle cost, and ergonomic comfort. The system's "end-cloud collaboration + modular evolution" architecture has enabled a technological leap from "passive protection" to "active early warning" and finally to "intelligent closed-loop" for power grid operations in Gansu Province. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the overall system architecture and signal flow of the present invention;

[0028] Figure 2 A schematic diagram comparing the working principle of the geometric constraint module and the matching mask effect of the present invention;

[0029] Figure 3 This is a schematic diagram of the AMFN structure + channel number marking of the present invention;

[0030] Figure 4 Schematic diagram of the JAM module of the present invention;

[0031] Figure 5 This is a schematic diagram of an example of segmentation-depth dual output of the present invention;

[0032] Figure 6 The training loss multi-curve of the present invention;

[0033] Figure 7 This is a schematic diagram of the on-site point layout and camera field of view coverage of the present invention;

[0034] Figure 8 This is a screenshot of the distance detection interface of the present invention. DETAILED DESCRIPTION

[0035] The following, in conjunction with the accompanying drawings, describes in detail the circuit structure, operating principle, signal flow, and the relationships between the various modules involved in the present invention, so that those skilled in the art can implement the technical solutions of the various claims of the present invention based on the contents of this embodiment. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. 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.

[0036] Example 1

[0037] 1. Hardware composition and circuit structure

[0038] See also Figure 1 (system overall architecture and signal flow) and Figure 7 (On-site layout & camera field of view coverage diagram) This system mainly consists of the following hardware modules:

[0039] 1. Binocular vision acquisition module

[0040] Circuit structure: Two 800×600@60FPS global shutter industrial cameras are connected to a camera control board via M12 connectors. The camera control board uses an ARM-Cortex-A53-based USB 3.0 interface to connect to the main control board (NPU + CPU). Baseline B = 120mm.

[0041] Function: Real-time acquisition of left and right synchronous images; output of RAW format image stream to the geometry pre-processing unit.

[0042] 2. Fill light and environment sensing module

[0043] Circuit structure: 4×4 visible light LED array, constant current drive power supply and I 2 The LED driver and sensor are mounted on the 40-pin GPIO interface of the main control board.

[0044] Function: Automatically fill in light in low-light or backlit environments; provides posture compensation and ambient exposure parameters.

[0045] 3. Edge computing master control module

[0046] Circuit Structure: Embedded industrial computer (x86 + RTX 4060Ti) with an Intel i7 CPU, 16GB of DDR4, and a 22MB INT8 quantized AMFN + JAM model. The system communicates with the NPU expansion card via PCIe.

[0047] Function: Deploys a deep learning inference engine to perform semantic segmentation, disparity regression, depth reconstruction, and safety distance calculation; also manages peripheral buses and alarm outputs.

[0048] 4. Alarm and communication module

[0049] Circuit structure: The sound and light alarm is connected to the main control board through the UART interface; the system also integrates RJ45 Ethernet, 4G module and CAN bus for remote monitoring and OTA upgrade.

[0050] Function: Outputs audible and visual alarm signals when the safety distance threshold is detected to be exceeded; uploads distance sequence and environmental data to the cloud via Ethernet / 4G.

[0051] 5. Power management module

[0052] Circuit Structure: 24V DC input, generated by a DC-DC converter to 12V, 5V, and 3.3V, respectively, to power the camera, LED array, sensor, and main control board. Built-in overcurrent and overvoltage protection circuits.

[0053] Embodiment 2

[0054] A power distribution network live working safety distance detection system based on deep learning and geometric constraints, comprising the following framework: a perception layer, a geometric constraint layer, a semantic-geometric joint deep network, a safety distance calculation, an edge computing deployment and a cloud;

[0055] The perception layer adopts an 800x600@60FPS industrial-grade global shutter binocular camera + visible light fill light array, with a baseline B = 120mm, integrated IMU and temperature and humidity sensors, to realize attitude compensation and environment adaptive exposure;

[0056] The geometric constraint layer constructs the essential matrix and the fundamental matrix through the camera calibration matrix, generates a refined search with a mask using the epipolar constraint, reduces the matching candidate area to 28% of the original area, and significantly reduces the backend calculation amount (formula (1)-formula (7);

[0057] E = TR (1)

[0058] In the formula, E is the essential matrix, the determinant of which is related to the rotation (R) and translation (T) between the cameras.

[0059]

[0060] The essential matrix E can be calculated from the camera intrinsic matrix and the fundamental matrix F. l and Kr are the intrinsic parameters of the left and right cameras, respectively.

[0061]

[0062] The essential matrix E is normalized according to the intrinsic matrix to obtain the normalized matrix φ.

[0063]

[0064] This is a general expression of the epipolar constraint. The pixel coordinates of the same physical point in the left and right images must satisfy this constraint.

[0065] ax+bx+c = 0 (5)

[0066] The standard form of the binary linear equation is a, b, and c, which are the parameters in the linear equation.

[0067] A1U R +B1V R +C1 = 0 (6)

[0068] The epipolar line equation in the right image corresponding to the point (U r V r ).

[0069] A2U L +B2V L +C2=0 (7)

[0070] The left image midpoint (U l V l ) is the corresponding polar line equation.

[0071] In a binocular vision system, using epipolar geometry constraints to limit the search area for pixel matching points in the RGB images captured by the left and right cameras is the key to achieving efficient and accurate feature matching. Specifically, after completing the binocular camera calibration in the laboratory using MATLAB, we can obtain the intrinsic parameters (intrinsic parameters) and extrinsic parameters (extrinsic parameters), including the rotation matrix and translation matrix, of the left and right cameras. These parameters are first used to calculate the essential matrix (Essential Matrix, Formula 1), which reflects the spatial relationship between the two camera coordinate systems; then the fundamental matrix (Fundamental Matrix, Formula 2) is derived using the intrinsic matrix. The fundamental matrix describes the homography relationship between the corresponding pixel points between the two cameras at the imaging plane level, laying the mathematical foundation for establishing pixel-level geometric constraints. By derivation of Formula (3), the fundamental matrix F can be directly calculated from the calibration parameters. Next, using the fundamental matrix F, we can establish the so-called epipolar constraint (Essential Matrix, Formula 4): that is, the matching point of any point on the left image on the right image must fall on a specific epipolar line. After this constraint relationship is expressed as an algebraic equation, it can be further written as the line equation of the epipolar line (Formula 5), ​​whose parameters a, b, and c are directly obtained from the basic matrix and the specific pixel position, thus simplifying the original two-dimensional problem of searching for matching points on the entire imaging plane to a one-dimensional problem of searching only on a straight line. This method not only greatly reduces the computational complexity, but also eliminates a large number of possible mismatches. Subsequently, Formulas (6) and (7) respectively give the epipolar analytical expressions on the imaging planes of the left and right cameras, which facilitates practical calculations and engineering implementation. Through this geometric constraint based on the basic matrix and epipolar equation, the search range of potential homonymous points can be effectively narrowed, greatly improving the accuracy and robustness of feature point matching. Especially in large outdoor scenes and complex environments, this method can also reduce the impact of interference such as illumination changes and repeated textures on matching. Finally, this epipolar constraint module inputs the calculated constraint band or matching candidate area into the subsequent feature extraction module, which not only simplifies the matching calculation and improves the accuracy, but also provides a solid foundation for subsequent applications such as three-dimensional reconstruction and distance measurement. Therefore, formulas (1)-(7) jointly realize the geometric constraints (epipolar constraints) that map the 3D calibration parameters to the 2D image plane, greatly reducing the search space of matching points and filtering out false matches, providing a basis for subsequent feature extraction and matching. This is also the core geometric mechanism of the binocular vision matching algorithm.

[0072] The semantic-geometric joint deep network includes an adaptive multi-scale feature network AMFN (Adaptive Multi-scale Feature Network), a joint attention module JAM (Joint Attention Module) and a dual-branch multi-task head;

[0073] The safety distance calculation simultaneously projects 14 key points of the human body and the device contour point set into 3-D space, constructs a KD tree to search for the shortest Euclidean distance frame by frame, and triggers an alarm in real time in the interface if the limit safety distance is exceeded. The built-in time-distance dual threshold filter (3-frame average) eliminates occasional jitter false alarms;

[0074] The edge computing deployment uses edge computing (ARM+NPU) and an x86+RTX 4060Ti edge box industrial computer. The INT8 quantization model size of the edge computing (ARM+NPU) is 22MB. Examples of suppliers of the x86+RTX 4060Ti edge box industrial computer include Aetina AIP-E600-4060Ti, Advantech MIC-770V3+SKY-QM4600 GPU module, and ASUS IoT PE1000U-4060Ti. The computer has a size of 220×200×70mm and a total power consumption of 110W. It has a CUDA 4352 core and 16GB of GDDR6. It supports Ubuntu 22.04+TensorRT 9.x, with a single-frame inference latency of ≈14ms and end-to-end latency reduced to 24ms.

[0075] The cloud is used to collect distance sequences, run LSTM-Anomaly to analyze long-term risk trends, and generate weekly reports.

[0076] Specifically, the adaptive multi-scale feature network AMFN (Adaptive Multi-scale Feature Network) is based on the ResNet-34 backbone, and adds dynamic dilated convolution (dilation∈{1,2,3}) to each level of feature map to achieve size-adaptive receptive field; the sampling position is fine-tuned through deformable convolution to improve the detection rate of oblique insulating rods; the joint attention module JAM (Joint Attention Module) introduces SE-Channel Attention to capture the global distributed channel weights, and superimposes CBAM-Spatial Attention to highlight edges and sharp points, so that the wire contour is still recognizable against a highly reflective background; the dual-branch multi-task head includes a segmentation branch and a disparity branch. The segmentation branch adopts YOLOv8-seg head+Dice / CE hybrid loss, and the disparity branch adopts Soft argmax disparity regression, and Sub-pixel cost volume encodes micro-disparity.

[0077] Example 3: Working Principle and Processing Flow

[0078] The complete process of the entire system from power-on to alarm is as follows Figure 1 As shown:

[0079] 1. Initialization and calibration

[0080] After the system is powered on, it reads the internal and external parameters of the camera stored on the main control board; performs IMU attitude self-test and temperature and humidity calibration.

[0081] 2. Image preprocessing and geometric constraints

[0082] After the left and right images are aligned with the distortion correction, the fundamental matrix and epipolar constraint (EP Module, Figure 2 ) Generate a search mask and reduce the matching candidate area to 28% of the original image.

[0083] 3. Semantic-geometric joint reasoning

[0084] The AMFN sub-network extracts multi-scale features, and the JAM module fuses channel and spatial attention to generate a segmentation map of key parts of the human body; the Soft argmax regresses the sub-disparity probability distribution to obtain a single-value disparity map ( Figures 3-4 ).

[0085] 4. Depth reconstruction and key point extraction

[0086] The depth map is calculated based on Depth = (f × B) / Disparity; the three-dimensional coordinates of 14 key points of the human body are extracted through the pre-trained posture estimation network ( Figure 5 ).

[0087] 5. Training results of semantic segmentation network,

[0088] Segmenting the key parts of the personnel from the power scene Figure 6 )

[0089] 6. Safety distance calculation and alarm judgment

[0090] The human body key points and device contour point set are simultaneously constructed K-D tree, and the shortest Euclidean distance is searched frame by frame; when lower than 0.6m, trigger "three frame average" filtered sound and light + voice alarm Figures 7-8 ).

[0091] 7. End-cloud collaboration and data management

[0092] The local distance sequence is uploaded to the cloud in real time, the LSTM-Anomaly model analyzes the long-term trend and generates a weekly report; through the OTA pipeline, the differential upgrade package is pushed.

[0093] 8. Signal flow direction and functional relationship between modules

[0094] Reference Figure 1 The signal flow direction arrow:

[0095] Image signal: camera→camera control board→USB3.0→NPU inference engine→depth map.

[0096] Sensing signal: IMU / temperature and humidity→I 2 C→CPU; for exposure correction and filter compensation.

[0097] Control signal: CPU→LED drive→light compensation lamp; CPU→UART→sound and light alarm.

[0098] Communication signal: main control board→Ethernet / 4G→cloud; for data upload and OTA.

[0099] Each module is connected through standard interfaces (USB, I 2 C, UART, Ethernet, CAN), and the signal processing link is clear:

[0100] Acquisition→correction→constraint screening (geometric preprocessing);

[0101] Feature extraction→attention fusion→disparity regression (deep neural network);

[0102] Coordinate reconstruction→tree structure query→threshold judgment (safety distance calculation);

[0103] Local alarm→data upload→cloud analysis (end-cloud closed loop);

[0104] This embodiment involves a binocular camera baseline B=120mm, a fill light array, an IMU attitude compensation, an AMFN and JAM module, and a KD tree shortest search method.

[0105] Example 4

[0106] This embodiment describes the comprehensive performance of the "deep learning + geometric constraints + multi-sensor fusion" safety distance detection system (hereinafter referred to as this system) in various scenarios and complex climatic conditions within the jurisdiction of Gansu Electric Power Company, highlighting its technological breakthroughs and engineering value compared to traditional methods.

[0107] 1 Experimental design and boundary conditions

[0108] 1.1 Test Scenario: This experiment selected typical power transmission and distribution scenarios in Gansu Province, including three types of applications: 10kV overhead lines, 35kV distribution transformer leads in mountainous areas, and busbars of box-type substations in urban core areas. A total of 32 test lines were covered in Lanzhou, Jiuquan, Tianshui, Longnan and other plateau climate zones with dry and cold weather, frequent sandstorms, and strong sunshine.

[0109] 1.2 Control algorithm:

[0110] Traditional stereo vision: SGBM+WLS-filter;

[0111] Rotating LiDAR (16 lines): Voxel Grid + ICP;

[0112] Electric field strength probe: domestically produced 400kV / m high voltage electric field meter;

[0113] Manual experience method: On-site judgment by senior operation and maintenance personnel selected by Gansu Electric Power Company.

[0114] 1.3 Device Platform: All algorithms were uniformly deployed on an x86-64 industrial control platform for fair comparison testing. This system also evaluated the end-side response latency on an Aetina AIP-E600-4060 Ti industrial computer.

[0115] 2 Index system and evaluation criteria

[0116] 2.1 Accuracy Indicators: The depth accuracy is evaluated using the Euclidean minimum distance MAE and RMSE. Based on the 0.6m warning line specified in Q / GDW1799.2-2013, the Recall@0.6m and F1-score are collected to evaluate the detection performance.

[0117] 2.2 Timeliness Indicators: Test the system’s single-frame inference time and end-to-end alarm delay.

[0118] 2.3 Robustness Indicator: MAE drift is evaluated under eight common weather conditions and three vibration levels (≤0.2g, 0.5g, 1g) on ​​the Gansu Plateau; ΔMAE ≤ +50% is considered satisfactory.

[0119] 2.4 Economic Indicators: The project's payback ratio is assessed using life cycle cost (LCC) and internal rate of return (IRR), taking into account initial installation CAPEX, operation and maintenance OPEX, and a 15-year lifecycle.

[0120] 3 Core Results and Data Analysis

[0121] 3.1 Precision Breakthrough: In 14,196 frames of high-contrast lighting data, the MAE of this system was only 0.028m, a 31.7% reduction compared to LiDAR and a 77.1% reduction compared to stereo vision.

[0122] 3.2 Improved Timeliness: Using the RTX 4060Ti platform, this system achieves an end-to-end 24ms response time. It can still issue an early alarm when a person approaches at a speed of 1.8m / s, leaving a manual response margin of more than 120ms.

[0123] 3.3 Robustness Verification: Under high vibration (1g) conditions, the MAE of this system only increases to 0.058m, which is significantly better than the 0.312m of traditional stereo vision.

[0124] 3.4 Long-term reliability: After one month of field deployment at a 10kV station in eastern Gansu, a total of 8.2×10 7 Frame. Test results are excellent

[0125] 4 Engineering significance and safety gains

[0126] 4.1 Decrease in accident rate: According to statistics from the operation and maintenance platform of Gansu Electric Power Company, the number of "approaching live parts" accidents on the pilot lines in 2023-2024 dropped from an average of 5 per year to 0, and accidental collisions during construction dropped from 14 to 2, with the overall accident rate dropping by 87%.

[0127] 4.2 Power outage loss savings: This system has an average of less than 3 false alarms per year, avoiding isolation power outages of more than 60 hours. Based on the average load and electricity price of 10kV users, the annual loss savings are approximately RMB 7.1 million.

[0128] 4.3 Human factors engineering optimization: The new helmet-mounted integrated machine replaces the traditional binocular solution, reducing the equipment weight by 54%, effectively extending the working time and reducing operator fatigue.

[0129] 5Sustainable operation and maintenance and upgrade potential

[0130] This system's OTA solution supports differential upgrades, using RSA-4096 digital signatures to ensure security. Even in mountainous areas of Gansu with weak 4G networks, upgrades take less than 75 seconds. By leveraging a cloud-based algorithm warehouse and AutoML pipeline, models tailored to specific tower structures can be generated within 48 hours, significantly reducing the burden of manual annotation.

[0131] In summary, experimental results and economic analysis demonstrate that this system significantly outperforms traditional solutions in terms of accuracy, timeliness, robustness, lifecycle cost, and ergonomic comfort. The system's "end-cloud collaboration + modular evolution" architecture has enabled a technological leap from "passive protection" to "active early warning" and finally to "intelligent closed-loop" operations in Gansu Province's power grid.

[0132] Explanation of terms:

[0133] 1. Live working: refers to the operation method of inspecting, repairing, inspecting and testing live distribution network lines or equipment without shutting down the power.

[0134] 2. Safety Clearance: The minimum distance workers must maintain near live equipment. This article uses the 0.6m standard in Q / GDW 1799.2-2013 as the benchmark.

[0135] E=TR (1)

[0136] E is the essential matrix, whose determinant is related to the rotation (R) and translation (T) between cameras.

[0137]

[0138] The essential matrix E can be calculated from the camera intrinsic parameter matrix and the basic matrix F. l and Kr are the intrinsic parameters of the left and right cameras respectively.

[0139]

[0140] The essential matrix E is normalized according to the internal parameter matrix to obtain the normalized matrix φ.

[0141] 3. Epipolar Constraint: In stereo vision, the pixel-corresponding search bandwidth obtained from the imaging geometry of the binocular camera can simplify matching. See formulas (4) to (7). ——GC module

[0142]

[0143] This is the general expression for the epipolar constraint. The pixel coordinates of the same physical point in both the left and right images must satisfy this constraint.

[0144] ax+bx+c=0 (5)

[0145] The standard form of a linear equation of two variables a, b, c are the parameters in the linear equation.

[0146] A1U R +B1V R +C1=0 (6)

[0147] The midpoint of the right image (U r V r ) is the corresponding polar line equation.

[0148] A2U L +B2V L +C2=0 (7)

[0149] The left image midpoint (U l V l ) is the corresponding polar line equation.

[0150] 4. Geometric Constraint Module (GCModule): A pre-screening module that uses epipolar constraints and camera calibration parameters to reduce the amount of matching calculations.

[0151] 5. Adaptive Multi-scale Feature Network (AMFN): A deep learning sub-network that fuses features of different resolutions through dynamic dilation convolution.

[0152] F l ∈Conv 1×1 (F)(8) Apply 1×1 convolution to the input feature map F to obtain the feature map F l . Used to adjust the number of channels.

[0153]

[0154] Perform global average pooling on the feature map F in the spatial dimensions (height and width) to retain the global statistical information of each channel, which facilitates the subsequent attention mechanism to extract key features.

[0155] N s =sigmoid(Conv 1×1 (F c )) (11)

[0156] N s : Attention or normalized features obtained through convolution and sigmoid activation function.

[0157] For a specific channel feature F c Perform 1×1 convolution and then normalize it through the sigmoid activation function, which is often used to generate spatial or channel attention weights.

[0158]

[0159] N s It takes itself as input, and after the product of sigmoid and itself, a weight map W is obtained, which is used for feature weighting or attention fusion.

[0160]

[0161] The output feature is the result of weighted sum with weight W. In practical terms, W and 1-W are applied to different parts of F respectively and then added together.

[0162] 6. Joint Attention Module (JAM): It simultaneously introduces channel attention and spatial attention to enhance the feature expression of key areas.

[0163] 7. Disparity Regression: The process of predicting the sub-disparity probability distribution for each pixel under soft argmax and regressing the single-valued disparity. — Formula (14) - Formula (15)

[0164]

[0165] d: Disparity, which refers to the horizontal distance difference between the corresponding points on the image planes of the left and right cameras for the same spatial point.

[0166] This formula calculates the final disparity d by multiplying each possible disparity value d by its softmax-normalized confidence score and summing the results to obtain the weighted average disparity. Softmax is often used in neural networks to output probability distributions, giving higher-confidence disparities a greater contribution.

[0167]

[0168] Used to calculate the depth Z, whose value is proportional to the focal length f and baseline T, and inversely proportional to the disparity d.

[0169] 8. Depth Reconstruction: Calculate the true depth value based on the relationship between parallax, baseline, and focal length:

[0170] Depth=(f×B) / Disparity. —Formula (15)

[0171]

[0172] 9. KD-Tree (K-Dimensional Tree): A multidimensional spatial index structure used for fast nearest neighbor queries. This paper uses it to search for the shortest Euclidean distance between a person and a device.

[0173] 10. Smooth L1 Loss: A regression loss function that combines the advantages of L1 and L2, improving the robustness of disparity prediction. — Formulas (16)-(17)

[0174]

[0175] D: represents the directed distance from a point (x, y, z) in space to the plane $Ax+By+Cz+d=0.

[0176] A, B, C, d: Parameters of the plane equation, which are the components and intercept of the normal vector respectively.

[0177] (x,y,z): The coordinates of a point in space.

[0178] This formula calculates the distance from a point to a plane in three-dimensional space. The numerator is the value obtained by substituting the coordinates of the point into the plane equation, and the denominator is the normalized value of the plane normal vector.

[0179]

[0180] L seg : Segmentation loss function, which is used to measure the gap between the segmentation result and the true label.

[0181] CrossEntropy(y,y^): Cross entropy loss function, which measures the difference between the predicted probability distribution and the true distribution.

[0182] Dice(y,y^): Dice coefficient loss, commonly used in binary or multi-classification tasks such as medical image segmentation, reflects the degree of overlap between the predicted area and the true area. The larger the value, the better the overlap.

[0183] y: ground truth label.

[0184] hat{y}: model output or predicted label (prediction)

[0185] 11. Dice Loss: A measure of the overlap between the predicted segmentation and the true segmentation. The smaller the value, the more accurate the segmentation. — Formula (18)

[0186]

[0187] LDisparity: Disparity loss function (Disparity Loss), used to measure the accuracy of disparity estimation.

[0188] i: The index of the pixel point. The subscript iii indicates the traversal and summation of all pixels.

[0189] d i : The true disparity label of the iii-th pixel.

[0190] d^ i : The network predicted disparity for the iii-th pixel.

[0191] This loss function is a pixel-by-pixel regression loss. For each pixel, the error between the predicted disparity and the true disparity is calculated and introduced into the smooth L1 loss, which is then summed over all pixels.

[0192] 12. MAE (Mean Absolute Error): Mean absolute error indicator, which is 5.46% in this study.

[0193] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A distribution network live working safety distance detection system based on deep learning and geometric constraints, characterized by: It includes binocular vision acquisition module, fill light and environment sensing module, edge computing main control module, alarm and communication module and power management module; The binocular vision acquisition module is used to capture left and right synchronous images in real time and output RAW format image streams to the geometric preprocessing unit; the fill light and environmental sensing module is used to automatically fill light in low-light or backlit environments, provide posture compensation and environmental exposure parameters; the edge computing main control module is used to deploy a deep learning inference engine to complete semantic segmentation, disparity regression, depth reconstruction and safe distance calculation, while managing the peripheral bus and alarm output; the alarm and communication module is used to output an audible and visual alarm signal when it detects that the safe distance threshold has been exceeded; the distance sequence and environmental data are uploaded to the cloud via Ethernet / 4G; the cloud is used to collect distance sequences, run LSTM-Anomaly to analyze long-term risk trends, and generate weekly reports.

2. The system for detecting safe distances for live working in distribution networks based on deep learning and geometric constraints according to claim 1, characterized in that: The binocular vision acquisition module includes two 800×600@60FPS global shutter industrial cameras and a camera control board. The two 800×600@60FPS global shutter industrial cameras are connected to the camera control board through M12 aviation plugs. The camera control board is connected to the main control board (NPU+CPU) using a USB 3.0 interface based on ARM-Cortex-A53, and the baseline B=120mm.

3. The distribution network live working safety distance detection system based on deep learning and geometric constraints according to claim 1 is characterized by: The fill light and environment sensing module consists of a 4×4 visible light LED array, a constant current drive power supply and an I 2 C bus temperature and humidity, IMU sensor, LED driver and sensor are mounted on the 40-pin GPIO interface of the main control board.

4. The system for detecting safe distances for live working in distribution networks based on deep learning and geometric constraints according to claim 1, characterized in that: The edge computing main control module includes an embedded industrial computer (x86+RTX 4060Ti), which contains an Intel i7 CPU, 16GB DDR4 and a 22MB INT8 quantized AMFN+JAM model.

5. The distribution network live working safety distance detection system based on deep learning and geometric constraints according to claim 1 is characterized by: The alarm and communication module includes an audible and visual alarm, which is connected to the main control board via a UART interface; the system also integrates RJ45 Ethernet, 4G module and CAN bus for remote monitoring and OTA upgrades.

6. The distribution network live working safety distance detection system based on deep learning and geometric constraints according to claim 1 is characterized by: The power management module receives 24V DC input and generates 12V, 5V and 3.3V through a DC-DC converter to power the camera, LED array, sensor and main control board respectively. The power management module has built-in overcurrent and overvoltage protection circuits.

7. A workflow of a distribution network live working safety distance detection system based on deep learning and geometric constraints, characterized in that: The steps include: Step 1: Initialization and calibration: After the system is powered on, read the internal and external parameters of the camera stored on the main control board, perform IMU attitude self-test and temperature and humidity calibration; Step 2: Image preprocessing and geometric constraints: After the left and right images are distortion corrected and aligned, a search mask is generated using the fundamental matrix and epipolar constraints to reduce the matching candidate area to 28% of the original image. Step 3: Semantic-geometric joint reasoning: The AMFN sub-network extracts multi-scale features, the JAM module fuses channel and spatial attention to generate a segmentation map of key human parts, and soft argmax regresses the sub-disparity probability distribution to obtain a single-value disparity map; Step 4: Depth reconstruction and key point extraction: Calculate the depth map based on Depth = (f × B) / Disparity, and extract the 3D coordinates of 14 key points of the human body through the pre-trained pose estimation network; Step 5: Calculate safety distance and determine alarm: Build a KD tree using the key points of the human body and the device's contour points. Search for the shortest Euclidean distance frame by frame. When the distance is less than 0.6m, an audible, visual, and voice alarm is triggered after "three-frame average" filtering. Step 6: End-cloud collaboration and data management: The local distance series is uploaded to the cloud in real time. The LSTM-Anomaly model analyzes long-term trends and generates weekly reports. The differential upgrade package is pushed through the OTA pipeline.

Images