Transformer substation AGV full-link intelligent navigation method, device, equipment and medium

Abstract

The invention discloses a substation AGV full-link intelligent navigation method, device and equipment and a medium. The method comprises the following steps: synchronously acquiring multi-modal data through a laser radar and a binocular camera, and preprocessing the multi-modal data; performing semantic segmentation on the processed point cloud and image to generate a semantic occupation grid map containing dynamic obstacle detection and tracking information; calculating a motion residual error based on map features and triggering external parameter self-learning calibration; dynamically fusing laser and visual positioning results through a confidence perception network; constructing and updating a semantic knowledge graph of the transformer substation in real time, and performing global path planning and local re-planning in combination with reinforcement learning; and calculating a multi-modal positioning error at a target point through a visual mark point, a semantic edge feature and force sense feedback, thereby realizing closed-loop fine adjustment and accurate positioning. The full-link intelligent navigation from sensing, positioning, planning to landing is realized, and the adaptability, robustness and operation precision of the AGV in the complex environment of the transformer substation are improved.

Description

Technical Field

[0001] This invention belongs to the field of AGV robot technology, specifically relating to a method, device, equipment, and medium for intelligent navigation of AGVs in substations. Background Technology

[0002] As power systems upgrade towards an "unmanned, intelligent operation and maintenance" model, Automated Guided Vehicles (AGVs) are increasingly being applied to equipment transportation, construction assistance, and intelligent inspection tasks in substations. Substation environments are typically characterized by enclosed spaces, narrow passageways, dense equipment, significant variations in lighting conditions, dynamic obstacles, and strong electromagnetic interference. These factors place high demands on the AGV's navigation system. Especially in certain critical operational scenarios, such as precise docking of switchgear, millimeter-level positioning accuracy is often required.

[0003] Currently, AGV navigation technology mostly employs LiDAR, visual sensors, or a combination thereof to achieve environmental perception and positioning. However, in special environments such as substations, existing navigation solutions still face many challenges. For example, how to achieve stable and accurate positioning in enclosed spaces lacking GPS signals; how to ensure the reliability and robustness of sensor data under conditions of strong electromagnetic interference, metal reflection, and drastic changes in lighting; how to conduct effective environmental understanding and obstacle avoidance decisions in complex scenarios involving both dynamic and static obstacles; and how to achieve high-precision end-point positioning and placement control when performing delicate operations. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, equipment, and medium for intelligent navigation of AGVs in substations throughout the entire process, in order to at least solve or improve one of the problems in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for intelligent navigation of AGVs in a substation throughout the entire process, comprising the following steps: Step 1: Simultaneously acquire multimodal data using LiDAR and binocular camera, and preprocess the multimodal data to obtain preprocessed LiDAR point cloud and visual image; Step 2: Perform semantic segmentation on the preprocessed laser point cloud and visual image obtained in Step 1 to generate a semantic occupancy grid map; and based on the semantic occupancy grid map, detect and track dynamic obstacles to obtain the semantic classification information of the dynamic obstacles and their predicted motion trajectory. Step 3: Based on the semantic occupancy grid map generated in Step 2, calculate the motion residual of the AGV; when the motion residual exceeds the preset threshold, trigger the self-learning calibration process to optimize the extrinsic parameter matrix between the LiDAR and the binocular camera, and obtain the optimized extrinsic parameters; Step 4: Based on the preprocessed laser point cloud and visual image from Step 1, and the optimized extrinsic parameters obtained in Step 3, calculate the laser radar confidence and visual confidence respectively using a confidence-based perception fusion network; dynamically weight and fuse the laser positioning results and visual positioning results according to the confidence to obtain the fused positioning result of the AGV. Step 5: Construct a semantic knowledge graph for the substation; update the semantic knowledge graph based on the semantic occupancy grid map, semantic classification information of dynamic obstacles, and predicted motion trajectories obtained in Step 2; based on the fusion localization results obtained in Step 4 and the updated semantic knowledge graph, use a reinforcement learning model for global path planning, and trigger local replanning when a conflict is detected between a dynamic obstacle and the planned path. Step 6: When the AGV arrives near the target location according to the planned path in Step 5, it calculates the multimodal placement error by identifying preset visual markers, matching semantic edge features, and combining feedback from force sensors; it then performs closed-loop fine-tuning based on the placement error until the placement error is less than the accuracy threshold, thus completing the precise placement.

[0006] Furthermore, step 1 specifically includes the following steps: Step 1.1: Generate a synchronization pulse signal using the FPGA to control the lidar and binocular camera to perform synchronous data acquisition; Step 1.2: Preprocess the acquired raw laser point cloud: First, use statistical filtering to remove isolated points, and then remove electromagnetic interference points based on the principle that the distance deviation of continuous points is consistent with the direction of electromagnetic field strength. Step 1.3: Preprocess the acquired raw visual image: When the light intensity is lower than the set threshold, infrared supplementary lighting is activated, and the Retinex algorithm is combined with semantic region enhancement processing to apply brightness gain to specific gray-level gradient regions.

[0007] Furthermore, step 2 specifically includes the following steps: Step 2.1: Use the improved Fast-SCNN network to perform semantic segmentation on the preprocessed visual image obtained in Step 1. Add a substation semantic attention module to the backbone network of the network and assign higher weights to the three key areas of switch cabinet, channel and cable trench. Step 2.2: Project the semantic segmentation results from Step 2.1 onto the laser point cloud coordinate system to construct a semantic occupancy raster map containing occupancy probability, semantic labels, and risk levels; Step 2.3: Extract dynamic targets from laser point clouds based on temporal point cloud difference method, and perform semantic classification of dynamic obstacles by combining human pose recognition and vehicle contour matching of visual images. Step 2.4: Use Kalman filtering to predict the future trajectory of the classified dynamic obstacles and mark the dynamic obstacle risk area in the semantic occupancy grid map.

[0008] Furthermore, step 3 specifically includes the following steps: Step 3.1: Based on the static semantic features in the semantic occupancy grid map, calculate the deviation between the AGV fusion positioning result and the actual semantic features to form a residual vector containing position deviation and attitude deviation; Step 3.2: Set a sliding window to accumulate the residual vectors of multiple consecutive frames. When the mean magnitude of the residual vectors within the window exceeds a preset threshold, trigger the extrinsic parameter self-learning optimization process. Step 3.3: Update the extrinsic parameter matrix between the LiDAR and the binocular camera using an iterative optimization algorithm. The update process is constrained by the maximum amplitude of a single update. Step 3.4: After completing the update of the extrinsic parameter matrix, calculate the reprojection error at the preset calibration verification point. If the reprojection error is less than the set tolerance, the calibration is confirmed to be valid; otherwise, the update result is discarded.

[0009] Furthermore, step 4 specifically includes the following steps: Step 4.1: Construct a confidence-aware fusion network. The network input includes a multi-dimensional feature vector of laser point cloud density, reflectivity variance, scanning stability, image texture entropy, illumination intensity, feature matching success rate, electromagnetic interference intensity, and occlusion rate. Step 4.2: Process the multi-dimensional feature vector through a confidence-aware fusion network to output the LiDAR confidence score and visual confidence score; Step 4.3: Based on the LiDAR confidence level and visual confidence level, dynamically calculate the fusion weight of the LiDAR positioning result and the visual positioning result; Step 4.4: Based on the fusion weight, perform weighted fusion of the laser positioning results and the visual positioning results to obtain the fused positioning result of the AGV; Step 4.5: Monitor confidence level changes, data delays, and abnormal positioning error jumps in real time. When a single sensor anomaly is detected, switch to enhanced positioning mode. When a dual sensor anomaly is detected, trigger manual intervention.

[0010] Furthermore, step 5 specifically includes the following steps: Step 5.1: Construct a semantic knowledge graph for the substation that includes equipment entities, channel entities, obstacle entities, and risk area entities, as well as their adjacent relationships, inclusion relationships, and taboo relationships; Step 5.2: Based on the semantic occupancy grid map, semantic classification information of dynamic obstacles, and predicted motion trajectory obtained in Step 2, update the node attributes and relationships in the semantic knowledge graph; Step 5.3: Construct a state space using the current state of the AGV and the neighboring node attributes of the semantic knowledge graph, and use the Graph-DQN reinforcement learning model to perform path planning in the state space; Step 5.4: Design a reward function that includes distance reduction reward, collision risk penalty, energy consumption penalty, risk area penalty, and time window penalty to optimize the path planning process; Step 5.5: When a dynamic obstacle is detected entering the warning area or its trajectory overlaps with the planned path, local replanning is triggered based on semantic knowledge graph and artificial potential field method; Step 5.6: Verify the feasibility of replanning the path through semantic collision detection to ensure that the path does not pass through risk areas and does not conflict with static semantic features.

[0011] Furthermore, step 6 specifically includes the following steps: Step 6.1: Preset multiple anti-reflective visual markers at the target landing point, extract the semantic edge features of the switch cabinet as visual and semantic references, and install a torque sensor on the AGV steering wheel to construct a force reference. Step 6.2: Calculate the first pose error by visually recognizing visual markers, calculate the second pose error by matching semantic edge features with laser point clouds, and calculate the third pose error by detecting contact state using a torque sensor. Step 6.3: Perform weighted fusion of the first pose error, the second pose error, and the third pose error to obtain the multimodal positioning error; Step 6.4: Based on the multimodal positioning error control, the AGV omnidirectional steering wheel performs vector compensation fine-tuning with a preset step size; Step 6.5: When the modulus of the multimodal placement error is less than the set threshold and the torque sensor detects a continuous and stable contact signal, the placement is determined to be complete.

[0012] In a second aspect, the present invention provides a substation AGV end-to-end intelligent navigation device, comprising: The data acquisition and preprocessing module is used to simultaneously acquire multimodal data through lidar and binocular camera, and preprocess the multimodal data to obtain preprocessed lidar point cloud and visual image; The semantic environment perception and modeling module is used to perform semantic segmentation on the preprocessed laser point cloud and visual image to generate a semantic occupancy grid map, and to detect and track dynamic obstacles based on the semantic occupancy grid map to obtain the semantic classification information of the dynamic obstacles and their predicted motion trajectory. The self-learning calibration module is used to calculate the motion residual of the AGV based on the semantic occupancy grid map. When the motion residual exceeds the preset threshold, the self-learning calibration process is triggered to optimize the extrinsic parameter matrix between the LiDAR and the binocular camera, and obtain the optimized extrinsic parameters. The confidence-adaptive fusion positioning module is used to calculate the LiDAR confidence and visual confidence respectively based on the preprocessed laser point cloud, visual image and optimized extrinsic parameters using the confidence-perception fusion network. The laser positioning result and the visual positioning result are dynamically weighted and fused according to the confidence, so as to obtain the fusion positioning result of AGV. The knowledge graph-driven path planning module is used to construct a semantic knowledge graph for substations. It updates the semantic knowledge graph based on the semantic occupancy grid map, the semantic classification information of dynamic obstacles, and the predicted motion trajectory. Based on the fused localization results and the updated semantic knowledge graph, a reinforcement learning model is used for global path planning. Local replanning is triggered when a conflict between a dynamic obstacle and the planned path is detected. The multimodal constraint precise placement module is used to calculate the multimodal placement error when the AGV arrives near the target location according to the planned path. It identifies preset visual markers, matches semantic edge features, and combines force sensor feedback to perform closed-loop fine-tuning based on the placement error until the placement error is less than the accuracy threshold, thus completing the precise placement.

[0013] In a third aspect, the present invention provides an electronic device including a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement the substation AGV full-link intelligent navigation method described above.

[0014] In a fourth aspect, the present invention provides a computer-readable storage medium storing at least one instruction that, when executed by a processor, implements the substation AGV full-link intelligent navigation method. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a substation AGV end-to-end intelligent navigation method according to an embodiment of the present invention; Figure 2 This is a structural block diagram of a substation AGV end-to-end intelligent navigation device according to an embodiment of the present invention; Figure 3 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0017] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0018] Example 1 This invention constructs an innovative five-layer architecture: semantic enhancement perception, self-learning calibration, confidence-adaptive fusion, knowledge graph planning, and multimodal positioning. Through semantic understanding to associate environmental information, self-learning to reduce manual maintenance, confidence fusion to resist interference, knowledge graph to adapt to scenario constraints, and multimodal verification to ensure positioning accuracy, it achieves intelligent navigation of the entire substation AGV chain.

[0019] like Figure 1 As shown, a substation AGV end-to-end intelligent navigation method includes the following steps: Step 1: Simultaneously acquire multimodal data using LiDAR and binocular camera, and preprocess the multimodal data to obtain preprocessed LiDAR point cloud and visual image; Step 2: Perform semantic segmentation on the preprocessed laser point cloud and visual image obtained in Step 1 to generate a semantic occupancy grid map; and based on the semantic occupancy grid map, detect and track dynamic obstacles to obtain the semantic classification information of the dynamic obstacles and their predicted motion trajectory. Step 3: Based on the semantic occupancy grid map generated in Step 2, calculate the motion residual of the AGV; when the motion residual exceeds the preset threshold, trigger the self-learning calibration process to optimize the extrinsic parameter matrix between the LiDAR and the binocular camera, and obtain the optimized extrinsic parameters; Step 4: Based on the preprocessed laser point cloud and visual image from Step 1, and the optimized extrinsic parameters obtained in Step 3, calculate the laser radar confidence and visual confidence respectively using a confidence-based perception fusion network; dynamically weight and fuse the laser positioning results and visual positioning results according to the confidence to obtain the fused positioning result of the AGV. Step 5: Construct a semantic knowledge graph for the substation; update the semantic knowledge graph based on the semantic occupancy grid map, semantic classification information of dynamic obstacles, and predicted motion trajectories obtained in Step 2; based on the fusion localization results obtained in Step 4 and the updated semantic knowledge graph, use a reinforcement learning model for global path planning, and trigger local replanning when a conflict is detected between a dynamic obstacle and the planned path. Step 6: When the AGV arrives near the target location according to the planned path in Step 5, it calculates the multimodal placement error by identifying preset visual markers, matching semantic edge features, and combining feedback from force sensors; it then performs closed-loop fine-tuning based on the placement error until the placement error is less than the accuracy threshold, thus completing the precise placement.

[0020] In one embodiment, step 1 specifically includes the following steps: Step 1.1: Generate a synchronization pulse signal using the FPGA to control the lidar and binocular camera to perform synchronous data acquisition; Step 1.2: Preprocess the acquired raw laser point cloud: First, use statistical filtering to remove isolated points, and then remove electromagnetic interference points based on the principle that the distance deviation of continuous points is consistent with the direction of electromagnetic field strength. Step 1.3: Preprocess the acquired raw visual image: When the light intensity is lower than the set threshold, infrared supplementary lighting is activated, and the Retinex algorithm is combined with semantic region enhancement processing to apply brightness gain to specific gray-level gradient regions.

[0021] In one embodiment, step 2 specifically includes the following steps: Step 2.1: Use the improved Fast-SCNN network to perform semantic segmentation on the preprocessed visual image obtained in Step 1. Add a substation semantic attention module to the backbone network of the network and assign higher weights to the three key areas of switch cabinet, channel and cable trench. Step 2.2: Project the semantic segmentation results from Step 2.1 onto the laser point cloud coordinate system to construct a semantic occupancy raster map containing occupancy probability, semantic labels, and risk levels; Step 2.3: Extract dynamic targets from laser point clouds based on temporal point cloud difference method, and perform semantic classification of dynamic obstacles by combining human pose recognition and vehicle contour matching of visual images. Step 2.4: Use Kalman filtering to predict the future trajectory of the classified dynamic obstacles and mark the dynamic obstacle risk area in the semantic occupancy grid map.

[0022] In one embodiment, step 3 specifically includes the following steps: Step 3.1: Based on the static semantic features in the semantic occupancy grid map, calculate the deviation between the AGV fusion positioning result and the actual semantic features to form a residual vector containing position deviation and attitude deviation; Step 3.2: Set a sliding window to accumulate the residual vectors of multiple consecutive frames. When the mean magnitude of the residual vectors within the window exceeds a preset threshold, trigger the extrinsic parameter self-learning optimization process. Step 3.3: Update the extrinsic parameter matrix between the LiDAR and the binocular camera using an iterative optimization algorithm. The update process is constrained by the maximum amplitude of a single update. Step 3.4: After completing the update of the extrinsic parameter matrix, calculate the reprojection error at the preset calibration verification point. If the reprojection error is less than the set tolerance, the calibration is confirmed to be valid; otherwise, the update result is discarded.

[0023] In one embodiment, step 4 specifically includes the following steps: Step 4.1: Construct a confidence-aware fusion network. The network input includes a multi-dimensional feature vector of laser point cloud density, reflectivity variance, scanning stability, image texture entropy, illumination intensity, feature matching success rate, electromagnetic interference intensity, and occlusion rate. Step 4.2: Process the multi-dimensional feature vector through a confidence-aware fusion network to output the LiDAR confidence score and visual confidence score; Step 4.3: Based on the LiDAR confidence level and visual confidence level, dynamically calculate the fusion weight of the LiDAR positioning result and the visual positioning result; Step 4.4: Based on the fusion weight, perform weighted fusion of the laser positioning results and the visual positioning results to obtain the fused positioning result of the AGV; Step 4.5: Monitor confidence level changes, data delays, and abnormal positioning error jumps in real time. When a single sensor anomaly is detected, switch to enhanced positioning mode. When a dual sensor anomaly is detected, trigger manual intervention.

[0024] In one embodiment, step 5 specifically includes the following steps: Step 5.1: Construct a semantic knowledge graph for the substation that includes equipment entities, channel entities, obstacle entities, and risk area entities, as well as their adjacent relationships, inclusion relationships, and taboo relationships; Step 5.2: Based on the semantic occupancy grid map, semantic classification information of dynamic obstacles, and predicted motion trajectory obtained in Step 2, update the node attributes and relationships in the semantic knowledge graph; Step 5.3: Construct a state space using the current state of the AGV and the neighboring node attributes of the semantic knowledge graph, and use the Graph-DQN reinforcement learning model to perform path planning in the state space; Step 5.4: Design a reward function that includes distance reduction reward, collision risk penalty, energy consumption penalty, risk area penalty, and time window penalty to optimize the path planning process; Step 5.5: When a dynamic obstacle is detected entering the warning area or its trajectory overlaps with the planned path, local replanning is triggered based on semantic knowledge graph and artificial potential field method; Step 5.6: Verify the feasibility of replanning the path through semantic collision detection to ensure that the path does not pass through risk areas and does not conflict with static semantic features.

[0025] In one embodiment, step 6 specifically includes the following steps: Step 6.1: Preset multiple anti-reflective visual markers at the target landing point, extract the semantic edge features of the switch cabinet as visual and semantic references, and install a torque sensor on the AGV steering wheel to construct a force reference. Step 6.2: Calculate the first pose error by visually recognizing visual markers, calculate the second pose error by matching semantic edge features with laser point clouds, and calculate the third pose error by detecting contact state using a torque sensor. Step 6.3: Perform weighted fusion of the first pose error, the second pose error, and the third pose error to obtain the multimodal positioning error; Step 6.4: Based on the multimodal positioning error control, the AGV omnidirectional steering wheel performs vector compensation fine-tuning with a preset step size; Step 6.5: When the modulus of the multimodal placement error is less than the set threshold and the torque sensor detects a continuous and stable contact signal, the placement is determined to be complete.

[0026] In a more specific embodiment, a substation AGV end-to-end intelligent navigation method is also provided, including the following steps: Step A: Perform semantic enhancement perception to achieve dual-dimensional environmental perception of geometric and semantic information.

[0027] Step B: Perform self-learning calibration to achieve sensor parameter calibration that maintains long-term accuracy without relying on manual intervention.

[0028] Step C: Adaptive confidence fusion is adopted to achieve multi-sensor fusion positioning based on dynamic adjustment of weights according to the environment.

[0029] Step D: Run knowledge graph-driven path planning to achieve intelligent planning based on prior knowledge and with rapid transferability.

[0030] Step E: Complete the precise placement of multimodal constraints to achieve highly reliable and accurate final positioning and execution.

[0031] Optionally, step A may include the following steps: A1. Synchronous acquisition and preprocessing of multimodal data.

[0032] Specifically, a 16-line LiDAR is installed at the top center of the AGV. The LiDAR parameters are: Velodyne VLP-16, scanning frequency 10Hz, ranging accuracy ±2cm, electromagnetic insulation level IP67. A global shutter binocular camera is installed at the front of the AGV. The parameters are: Baslerac A2500-14gm, resolution 1280×720, frame rate 30fps, baseline 120mm, and support for 850nm infrared illumination. A 10Hz synchronization pulse signal is generated through FPGA to ensure that the time stamp error between the laser and vision is <1ms.

[0033] Statistical filtering combined with electromagnetic interference removal is used. First, isolated points with fewer than 10 neighboring points are removed. Then, electromagnetic interference points are removed based on the principle that the distance deviation of three consecutive points is greater than ±5cm and the direction is consistent with the electromagnetic field strength. In low light conditions (<50 lux), infrared supplementary lighting is activated. Retinex combined with semantic region enhancement algorithm is used to increase the brightness gain to 3.5 for the metal texture area of ​​the switch cabinet (grayscale gradient > 80) to avoid feature point loss.

[0034] A2. Semantic segmentation and map construction of substations.

[0035] An improved Fast-SCNN network is adopted, and a substation semantic attention module is added to the backbone to give higher weights to three key regions: switchgear (metal edge features), passage (flat ground features), and cable trench (concave contour features). The semantic segmentation results are projected onto the laser point cloud coordinate system to construct a semantic grid map with a resolution of 5cm×5cm. Each grid contains the occupancy probability (0-1), semantic label (e.g., switch cabinet-1, channel-2), and risk level (e.g., high voltage switch area risk=1.0, channel risk=0.2). The map is updated every 0.5s.

[0036] A3. Semantic Detection and Tracking of Dynamic Obstacles: By using the "temporal point cloud difference method" (comparing point clouds of two frames before and after, clustering those with a movement distance > 0.1m as dynamic targets), combined with human pose recognition and vehicle contour matching of visual images, construction workers (height 1.5-1.8m, speed 0.5-1.5m / s), tool vehicles (size 0.8-1.2m×0.5-0.8m), and mobile equipment (such as temporary power distribution boxes) can be distinguished. Kalman filtering is used to predict the trajectory of a dynamic target within the next 3 seconds (position error < ±10cm), and dynamic obstacle risk areas are marked in the semantic map to provide real-time constraints for path planning.

[0037] Optionally, step B may include the following steps: B1. Real-time calculation of motion residuals.

[0038] Based on the static semantic features in the semantic map (such as the right-angled edge of the switch cabinet and the equipment mounting hole), the deviation between the AGV fusion positioning result and the actual semantic features is calculated to form a residual vector R(t)=(Δx,Δy,Δθ), where Δx / Δy is the position deviation and Δθ is the attitude deviation, and the calculation frequency is 10Hz. Set a sliding window (window size = 100 frames), calculate the mean residual value R_avg = (Δx_avg, Δy_avg, Δθ_avg) within the window, and trigger self-learning calibration when the modulus of R_avg is greater than 2mm. B2. External parameter self-learning optimization model.

[0039] The extrinsic parameter matrix [R|T] (a 3×4 matrix, where R is the rotation matrix and T is the translation vector) of the laser and camera is iteratively optimized using the following formula: [R|T]_new=[R|T]_old+α ΔR_t Where α is the self-learning rate coefficient (dynamically adjusted according to the residual size: α=0.1 when residual > 3mm, α=0.05 when residual 2-3mm), and ΔR_t is the time-cumulative residual (ΔR_t=ΣR_avg). Δt, where Δt is the window time interval); To avoid abrupt changes in external parameters, a maximum single update amplitude is set (ΔR≤0.5°, ΔT≤1mm) to ensure a smooth calibration process.

[0040] B3. Self-verification and anomaly correction: After each self-learning calibration is completed, the system automatically samples 5 preset calibration verification points in the substation (such as equipment mounting holes with known coordinates), calculates the reprojection error (error of laser point cloud projection onto the image plane) after updating the external parameters, and confirms that the calibration is effective if RMSE < 0.5 pixels. If RMSE > 0.5 pixels, it is judged as an abnormal residual (such as interference from dynamic obstacles), the current update result is discarded, and the residual of the next sliding window is recalculated.

[0041] Optionally, step C may include the following steps: C1, Confidence-AwareFusionNet (CAF-Net) confidence modeling.

[0042] The architecture employs a 2-layer CNN combined with a 1-layer fully connected layer. The input is the following feature vector: point cloud density (per m 2 Point count), reflectivity variance (distinguishing between metal and non-metal), scan stability (overlap rate of point clouds in 5 consecutive frames); image texture entropy (measuring texture richness), illumination intensity (lux), ORB feature matching success rate (number of matches / extractions per frame); electromagnetic interference intensity (V / m), occlusion rate (percentage of laser / visual area occluded). CAF-Net outputs laser confidence C_L (0-1) and visual confidence C_V (0-1) with an update cycle of 100ms. When the illumination is <20 lux, C_V drops to 0.2, and when the electromagnetic interference is >80V / m, C_L drops to 0.5.

[0043] C2, Dynamic Weighted Fusion Algorithm: Sensor weights are dynamically assigned based on confidence levels, using the following formula: w_L=C_L / (C_L+C_V), w_V=C_V / (C_L+C_V); The fused localization result is X_f=(x_f, y_f, θ_f)=w_L X_L+w_V X_V; where X_L is the laser positioning result and X_V is the visual positioning result). Scene adaptation rules: Low light scene (light < 20 lux), C_V < 0.3, w_L increased to 0.8-0.9, with priority given to laser; Strong electromagnetic scene (E > 80V / m), C_L < 0.6, w_V increased to 0.6-0.7, with priority given to visual semantic features; Normal scene (light 20-60 lux, E < 60V / m), w_L = 0.5-0.6, w_V = 0.4-0.5, with balanced fusion.

[0044] C3. Anomaly detection and fault tolerance.

[0045] Anomaly identification: Real-time monitoring of three types of anomalies: sudden drop in confidence (e.g., C_L / C_V decreases by >0.3 within 100ms), data delay (laser / visual data delay >50ms), and positioning error jump (X_f deviation from the previous frame >5mm); Fault tolerance strategy: In single-mode emergency mode, if one type of sensor malfunctions, it automatically switches to "enhanced positioning" of another type of sensor. When the laser is malfunctioning, positioning is based on visual markers in the semantic map (such as equipment nameplates); when the vision is malfunctioning, positioning is based on laser semantic features (such as the edge of the switch cabinet), ensuring that the positioning error is <±4mm. If both types of sensors malfunction due to manual intervention, the system will issue an alarm and switch to "manual remote control mode" to prevent the AGV from going out of control.

[0046] Optionally, step D may include the following steps: D1. Construct a semantic knowledge graph for substations.

[0047] Graph structure: A three-level node system is constructed around entities, relationships, and attributes. Among them, Primary nodes include equipment (switchgear / transformer), passageways, obstacles (static / dynamic), and risk areas (high-voltage areas / energized areas). Relationships are defined as follows: adjacent (e.g., switch cabinet A - adjacent - passage B), included (e.g., passage B - included - cable trench C), and prohibited (e.g., AGV - prohibited - high voltage area D). Attribute information: Each node includes spatial coordinates (e.g., switch cabinet center (x, y)), passage cost (e.g., energy consumption coefficient of channel B = 1.0), and time constraints (e.g., switch cabinet installation window [T1=14:00, T2=14:30]). With each update of the semantic map, the graph is updated synchronously with dynamic obstacle nodes (such as construction worker E - position (x1, y1) - speed 0.8 m / s) to ensure the timeliness of knowledge.

[0048] D2, Graph-DQN reinforcement learning path planning.

[0049] The state space is input as a 6-dimensional state vector S = [AGV current coordinates, semantic attributes of adjacent nodes (such as channel risk level), remaining power, time difference ΔT from the installation window, and distance d from dynamic obstacles]. The operating space outputs a combination of forward step length (0.1-0.5m), turning angle (0-10°), and speed (0.3-1m / s) to ensure that the AGV can move flexibly in narrow passages. Reward function optimization: Introducing a reward function constrained by knowledge graphs. R=α d_reduce-β collision-γ energy-δ risk-ε T_penalty Wherein, d_reduce represents the reduction in distance between the AGV and the target point (positive reward); collision represents the collision risk (a penalty is triggered if the distance to an obstacle is less than 0.5m); energy represents energy consumption (calculated based on path length and slope); risk represents the risk penalty (triggered when entering a risky area, with higher risk levels resulting in greater penalties); T_penalty represents the window period penalty (the penalty coefficient ε increases from 0.1 to 0.3 when ΔT < 10min); α=0.3, β=0.25, γ=0.2, δ=0.15, and ε=0.1 are empirical weights; Training and migration are carried out iteratively in the digital twin environment of the substation; when the environment layout changes (such as adding switchgear), only the knowledge graph nodes need to be updated, and there is no need to retrain.

[0050] D3, Dynamic Local Replanning.

[0051] Triggering conditions: Local replanning is triggered when a dynamic obstacle enters the "warning zone (distance from AGV < 0.8m)" or when the obstacle's trajectory overlaps with the AGV's path. The replanning algorithm uses a knowledge graph combined with an artificial potential field method. Based on the channel-obstacle relationship in the graph, it assigns low repulsive force to channels and high repulsive force to obstacles in the potential field, thereby quickly generating obstacle avoidance paths. Path verification is performed through semantic collision detection (verifying whether the path passes through a risk area or conflicts with static semantic features).

[0052] Optionally, step E may include the following steps: E1. Construct a multimodal positioning benchmark.

[0053] For visual reference, three anti-reflective visual markers are pre-set at the switch cabinet installation location. These markers are made of QR codes combined with a black light-absorbing frame structure (QR code diameter 8cm, frame width 2cm, reflectivity <5%), and are distributed in an isosceles triangle (spacing 1.5-2m). Semantic benchmarks are used to extract semantic edge features of the switchgear (such as right-angled edges of the cabinet and circular features of mounting holes) as auxiliary benchmarks for placement. Force reference: A torque sensor is installed on the AGV steering wheel. When the AGV comes into contact with the switch cabinet, the sudden change in torque serves as a physical signal for the completion of the positioning.

[0054] E2, Multimodal error fusion calculation.

[0055] The total error E_total is calculated using a weighted fusion method. E_total=w1 (Δx1,Δy1,Δθ1)+w2 (Δx², Δy², Δθ²) + w³ F_error Where w1=0.5 (visual weight), w2=0.3 (semantic weight), w3=0.2 (force weight), and F_error is the force deviation vector.

[0056] The visual error is calculated using the PnP algorithm to determine the relative pose error (Δx1, Δy1, Δθ1) between the AGV and the visual markers. Semantic error is calculated by matching the AGV laser point cloud with the semantic edge features of the switch cabinet and calculating the position deviation (Δx2, Δy2, Δθ2). Force error: When the torque sensor detects a contact signal, the contact deviation between the AGV and the cabinet is calculated (if only one side is in contact, there is a deviation of Δθ).

[0057] E3. Closed-loop fine-tuning and positioning confirmation: Fine-tuning control, based on E_total, the AGV omnidirectional steering wheel performs vector compensation in 0.3mm steps (e.g., if Δx=2mm, it will fine-tune forward by 2mm, and if Δθ=0.5°, it will fine-tune the steering by 0.5°), with a fine-tuning frequency of 10Hz; Positioning determination: when the mold length of E_total is < ±2mm and the torque sensor detects a stable contact signal (lasting 2s), the positioning is determined to be complete. Data feedback: After placement is completed, the system records data such as placement error and sensor confidence level, which are used to optimize CAF-Net and the self-learning calibration model to achieve continuous iteration.

[0058] The following section will provide a further explanation of this solution using a specific implementation scenario.

[0059] Automatic handling and placement of substation switchgear I. Scene Parameters Environment: The substation is an enclosed indoor space with a passageway width of 1.8m, a switch cabinet spacing of 1m, a light intensity of 15 lux (nighttime emergency lights), and an electromagnetic field strength of E=90V / m (near the high-voltage switch). The dynamic obstacle is two construction workers carrying metal tools (moving at a speed of 0.8m / s). The task is for the AGV to move the switch cabinet from the storage area (point A, coordinates (0,0,0°)) to the installation position (point B, coordinates (30m,5m,90°)) within the window period [T1=14:00,T2=14:30], with a positioning accuracy of ≤±2mm; The AGV is configured with dimensions of 1.2m×0.8m×0.6m, omnidirectional drive (maximum speed 1m / s), and is equipped with a VLP-16 lidar, a Basler binocular camera, and a steering wheel torque sensor.

[0060] II. Implementation Steps.

[0061] 1. Semantic perception and mapping (Step A): Data collected during AGV no-load cruise was processed with laser noise reduction, retaining 98% of valid points. Visual enhancement improved the contrast of metallic textures by 65%. Improved Fast-SCNN segmentation divides the switch cabinet area (label 1), channel (label 2), and high-voltage risk area (label 3) to construct a semantic raster map; Two construction workers were detected dynamically, and Kalman filtering was used to predict their trajectories (they will pass through the midpoint of the passage (15m, 2.5m) at 14:10).

[0062] 2. Self-learning calibration (Step B): The system read the historical residual (R_avg=1.8mm for the first 100 frames) and did not trigger self-learning. Five verification points were sampled, and the reprojection error RMSE was 0.3 pixels, confirming the validity of the extrinsic parameters.

[0063] 3. Confidence fusion localization (step C): CAF-Net input features: laser point cloud density 200 points / m², visual texture entropy 80, E=90V / m; output C_L=0.55, C_V=0.45. Weight calculation: w_L=0.55, w_V=0.45, fusion positioning result X_f error ±2.8mm; At 14:08, E suddenly rose to 95V / m, C_L dropped to 0.5, the weights were adjusted to w_L=0.5 and w_V=0.5, and the error stabilized at ±3mm.

[0064] 4. Knowledge Graph Planning (Step D): Knowledge graph node, point A - adjacent - channel 2, channel 2 - taboo - high pressure zone 3 (distance 1m), point B - window period [14:00, 14:30]; The Graph-DQN planned path is: A → Channel 2 → Detour construction personnel (arriving at the detour point (14m, 3m) at 14:10) → B, with a path length of 32.5m and an estimated arrival time of 14:22. At 14:15, the construction workers changed direction, triggering a local replanning (a new path was generated in 0.25 seconds), increasing the detour distance by 0.8m, but still meeting the window period.

[0065] 5. Multimodal placement (step E): At 14:22, the AGV arrived near point B and visually identified three anti-reflective markers. The values ​​of Δx1 = 2.1 mm, Δy1 = 1.8 mm, and Δθ1 = 0.3° were calculated. Laser matching was used to determine the edge features of the switch cabinet, and the calculated values ​​were Δx2=1.9mm, Δy2=2.0mm, and Δθ2=0.2°. Fusion error E_total=(0.5×2.1+0.3×1.9,0.5×1.8+0.3×2.0,0.5×0.3+0.3×0.2)=(1.62mm,1.5mm,0.21°); The steering wheel was fine-tuned in 0.3mm increments. At 14:22:10, E_total < ±2mm, and the torque sensor detected stable contact (5.2N). m), the placement is complete.

[0066] Example 2 like Figure 2 As shown, based on the same inventive concept as the above embodiments, the present invention also provides a substation AGV end-to-end intelligent navigation device, comprising: The data acquisition and preprocessing module is used to simultaneously acquire multimodal data through lidar and binocular camera, and preprocess the multimodal data to obtain preprocessed lidar point cloud and visual image; The semantic environment perception and modeling module is used to perform semantic segmentation on the preprocessed laser point cloud and visual image to generate a semantic occupancy grid map, and to detect and track dynamic obstacles based on the semantic occupancy grid map to obtain the semantic classification information of the dynamic obstacles and their predicted motion trajectory. The self-learning calibration module is used to calculate the motion residual of the AGV based on the semantic occupancy grid map. When the motion residual exceeds the preset threshold, the self-learning calibration process is triggered to optimize the extrinsic parameter matrix between the LiDAR and the binocular camera, and obtain the optimized extrinsic parameters. The confidence-adaptive fusion positioning module is used to calculate the LiDAR confidence and visual confidence respectively based on the preprocessed laser point cloud, visual image and optimized extrinsic parameters using the confidence-perception fusion network. The laser positioning result and the visual positioning result are dynamically weighted and fused according to the confidence, so as to obtain the fusion positioning result of AGV. The knowledge graph-driven path planning module is used to construct a semantic knowledge graph for substations. It updates the semantic knowledge graph based on the semantic occupancy grid map, the semantic classification information of dynamic obstacles, and the predicted motion trajectory. Based on the fused localization results and the updated semantic knowledge graph, a reinforcement learning model is used for global path planning. Local replanning is triggered when a conflict between a dynamic obstacle and the planned path is detected. The multimodal constraint precise placement module is used to calculate the multimodal placement error when the AGV arrives near the target location according to the planned path. It identifies preset visual markers, matches semantic edge features, and combines force sensor feedback to perform closed-loop fine-tuning based on the placement error until the placement error is less than the accuracy threshold, thus completing the precise placement.

[0067] Example 3 As shown in Figure X, the present invention also provides an electronic device 100 for realizing a full-link intelligent navigation method for AGVs in substations; The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.

[0068] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the substation AGV full-link intelligent navigation method of Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.

[0069] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0070] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.

[0071] The memory 101 in the electronic device 100 stores multiple instructions to implement a substation AGV full-link intelligent navigation method, and the processor 102 can execute multiple instructions to achieve the following: Step 1: Simultaneously acquire multimodal data using LiDAR and binocular camera, and preprocess the multimodal data to obtain preprocessed LiDAR point cloud and visual image; Step 2: Perform semantic segmentation on the preprocessed laser point cloud and visual image obtained in Step 1 to generate a semantic occupancy grid map; and based on the semantic occupancy grid map, detect and track dynamic obstacles to obtain the semantic classification information of the dynamic obstacles and their predicted motion trajectory. Step 3: Based on the semantic occupancy grid map generated in Step 2, calculate the motion residual of the AGV; when the motion residual exceeds the preset threshold, trigger the self-learning calibration process to optimize the extrinsic parameter matrix between the LiDAR and the binocular camera, and obtain the optimized extrinsic parameters; Step 4: Based on the preprocessed laser point cloud and visual image from Step 1, and the optimized extrinsic parameters obtained in Step 3, calculate the laser radar confidence and visual confidence respectively using a confidence-based perception fusion network; dynamically weight and fuse the laser positioning results and visual positioning results according to the confidence to obtain the fused positioning result of the AGV. Step 5: Construct a semantic knowledge graph for the substation; update the semantic knowledge graph based on the semantic occupancy grid map, semantic classification information of dynamic obstacles, and predicted motion trajectories obtained in Step 2; based on the fusion localization results obtained in Step 4 and the updated semantic knowledge graph, use a reinforcement learning model for global path planning, and trigger local replanning when a conflict is detected between a dynamic obstacle and the planned path. Step 6: When the AGV arrives near the target location according to the planned path in Step 5, it calculates the multimodal placement error by identifying preset visual markers, matching semantic edge features, and combining feedback from force sensors; it then performs closed-loop fine-tuning based on the placement error until the placement error is less than the accuracy threshold, thus completing the precise placement.

[0072] Example 4 If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0073] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A substation AGV full-link intelligent navigation method, characterized in that, The method comprises the following steps: Step 1, synchronously collecting multi-modal data by laser radar and binocular camera, and pre-processing the multi-modal data to obtain pre-processed laser point cloud and visual image; Step 2, performing semantic segmentation on the pre-processed laser point cloud and visual image obtained in step 1 to generate a semantic occupancy grid map; and based on the semantic occupancy grid map, detecting and tracking dynamic obstacles to obtain semantic classification information and a predicted motion trajectory of the dynamic obstacles; Step 3, based on the semantic occupancy grid map generated in step 2, calculating the motion residual of the AGV; when the motion residual exceeds a preset threshold, triggering a self-learning calibration process to optimize the extrinsic parameter matrix between the laser radar and the binocular camera to obtain an optimized extrinsic parameter; Step 4, based on the pre-processed laser point cloud and visual image in step 1 and the optimized extrinsic parameter obtained in step 3, using a confidence perception fusion network to calculate laser radar confidence and visual confidence respectively; according to the confidence, dynamically weighting and fusing the laser positioning result and the visual positioning result to obtain a fusion positioning result of the AGV; Step 5, constructing a substation semantic knowledge graph; based on the semantic occupancy grid map, the semantic classification information and the predicted motion trajectory of the dynamic obstacles obtained in step 2, updating the semantic knowledge graph; based on the fusion positioning result obtained in step 4 and the updated semantic knowledge graph, using a reinforcement learning model to perform global path planning, and when a conflict between the dynamic obstacle and the planned path is detected, triggering local re-planning; Step 6, when the AGV reaches the vicinity of the target point according to the planned path in step 5, calculating a multi-modal landing error by recognizing a preset visual marker point, matching semantic edge features and combining force sensor feedback; performing closed-loop fine tuning according to the landing error until the landing error is less than a precision threshold, and completing accurate landing.

2. The substation AGV full-link intelligent navigation method according to claim 1, characterized in that, Step 1 specifically comprises the following steps: Step 1.1, generating a synchronization pulse signal by FPGA to control the laser radar and the binocular camera to perform synchronous data collection; Step 1.2, pre-processing the collected original laser point cloud: first removing isolated points by statistical filtering, and then removing electromagnetic interference points based on the principle that the distance deviation of consecutive points is consistent with the direction of the electromagnetic field; Step 1.3, pre-processing the collected original visual image: starting infrared fill light when the light intensity is lower than a set threshold, and applying brightness gain to specific gray gradient regions by using the Retinex algorithm combined with semantic region enhancement processing.

3. The substation AGV full-link intelligent navigation method according to claim 2, characterized in that, Step 2 specifically comprises the following steps: Step 2.1, using an improved Fast-SCNN network to perform semantic segmentation on the pre-processed visual image obtained in step 1, and adding a substation semantic attention module in the backbone network of the network to give higher weights to three types of key regions, namely switch cabinet, passageway and cable trench; Step 2.2, projecting the semantic segmentation result of step 2.1 to the laser point cloud coordinate system to construct a semantic occupancy grid map containing occupancy probability, semantic label and risk level; Step 2.3, extract dynamic targets from laser point cloud based on time point cloud difference method, and combine human pose recognition and vehicle contour matching of visual image to classify dynamic obstacles semantically; Step 2.4, predict the future motion trajectory of the classified dynamic obstacle by Kalman filter, and mark the dynamic obstacle risk area in the semantic occupancy grid map.

4. The substation AGV full-link intelligent navigation method according to claim 3, characterized in that, Step 3 specifically includes the following steps: Step 3.1, based on the static semantic features in the semantic occupancy grid map, calculate the deviation between the AGV fusion positioning result and the actual semantic features to form a residual vector containing position deviation and attitude deviation; Step 3.2, set a sliding window to accumulate the residual vectors of continuous multiple frames, and when the mean length of the residual vectors in the window exceeds the preset threshold, trigger the external parameter self-learning optimization process; Step 3.3, update the external parameter matrix between the laser radar and the binocular camera through the iterative optimization algorithm, and the update process is limited by the maximum amplitude constraint of single update; Step 3.4, after completing the external parameter matrix update, calculate the re-projection error at the preset calibration verification point, and when the re-projection error is less than the set tolerance, confirm that the calibration is valid, otherwise discard the update result.

5. The substation AGV full-link intelligent navigation method according to claim 4, characterized in that, Step 4 specifically includes the following steps: Step 4.1, construct a confidence perception fusion network, the input of the network includes multi-dimensional feature vectors of laser point cloud density, reflectivity variance, scanning stability, image texture entropy, light intensity, feature matching success rate, electromagnetic interference intensity and occlusion rate; Step 4.2, process the multi-dimensional feature vectors through the confidence perception fusion network to output the laser radar confidence and the visual confidence; Step 4.3, dynamically calculate the fusion weight of the laser positioning result and the visual positioning result according to the laser radar confidence and the visual confidence; Step 4.4, based on the fusion weight, the laser positioning result and the visual positioning result are weighted and fused to obtain the fusion positioning result of the AGV; Step 4.5, real-time monitor the confidence mutation, data delay and positioning error jump abnormal situation, switch to enhanced positioning mode when single sensor anomaly is detected, and trigger manual intervention when double sensor anomaly is detected.

6. The substation AGV full-link intelligent navigation method according to claim 5, characterized in that, Step 5 specifically includes the following steps: Step 5.1, construct a substation semantic knowledge graph containing device entities, channel entities, obstacle entities and risk area entities, as well as their adjacent relationships, inclusion relationships and taboo relationships; Step 5.2, based on the semantic occupancy grid map obtained in step 2, the semantic classification information and the predicted motion trajectory of the dynamic obstacle, update the node attributes and relationships in the semantic knowledge graph; Step 5.3, construct a state space based on the current state of the AGV and the adjacent node attributes of the semantic knowledge graph, and use the Graph-DQN reinforcement learning model to plan the path in the state space; Step 5.4, design a reward function containing distance reduction reward, collision risk penalty, energy consumption penalty, risk area penalty and time window penalty to optimize the path planning process; Step 5.5, when detecting that the dynamic obstacle enters the warning area or its motion trajectory overlaps with the planned path, trigger local re-planning based on the semantic knowledge graph and artificial potential field method; Step 5.6, verify the feasibility of the re-planned path by semantic collision detection, ensure that the path does not pass through the risk area and does not conflict with static semantic features.

7. The substation AGV full-link intelligent navigation method according to claim 6, characterized in that, Step 6 specifically includes the following steps: Step 6.1, preset a plurality of anti-glare visual markers at the target landing site, extract the semantic edge features of the switch cabinet as visual and semantic reference, and install a torque sensor on the AGV rudder to build a force sense reference; Step 6.2, calculate the first pose error by visual recognition of the visual marker points, calculate the second pose error by matching the semantic edge features of the laser point cloud, and calculate the third pose error by detecting the contact state through the torque sensor; Step 6.3, weighted fusion of the first, second and third pose errors to obtain a multi-modal landing error; Step 6.4, based on the multi-modal landing error, control the AGV omni-directional rudder to perform vector compensation fine tuning with a preset step size; Step 6.5, when the modulus of the multi-modal landing error is less than the set threshold and the torque sensor detects a continuous and stable contact signal, it is determined that the landing is completed.

8. A substation AGV full-link intelligent navigation device, characterized in that, Comprise: a data acquisition and preprocessing module for synchronously acquiring multi-modal data through a laser radar and a binocular camera, and preprocessing the multi-modal data to obtain preprocessed laser point cloud and visual images; a semantic environment perception and modeling module for performing semantic segmentation on the preprocessed laser point cloud and visual images to generate a semantic occupancy grid map, and detecting and tracking dynamic obstacles based on the semantic occupancy grid map to obtain semantic classification information and predicted motion trajectories of the dynamic obstacles; a self-learning calibration module for calculating the motion residual of the AGV based on the semantic occupancy grid map, triggering a self-learning calibration process when the motion residual exceeds a preset threshold to optimize the extrinsic matrix between the laser radar and the binocular camera, and obtaining an optimized extrinsic parameter; a confidence adaptive fusion positioning module for calculating laser radar confidence and visual confidence respectively using a confidence perception fusion network based on the preprocessed laser point cloud and visual images and the optimized extrinsic parameter, dynamically weighting the laser positioning result and the visual positioning result according to the confidence, and obtaining a fusion positioning result of the AGV; a knowledge graph driven path planning module for constructing a substation semantic knowledge graph, updating the semantic knowledge graph based on the semantic occupancy grid map, the semantic classification information of the dynamic obstacles and the predicted motion trajectories, performing global path planning based on the fusion positioning result and the updated semantic knowledge graph using a reinforcement learning model, and triggering local re-planning when a conflict between the dynamic obstacles and the planned path is detected; a multi-modal constraint accurate landing module for calculating a multi-modal landing error by recognizing the preset visual marker points, matching the semantic edge features and combining the force sensor feedback when the AGV reaches the vicinity of the target point according to the planned path, performing closed-loop fine tuning according to the landing error until the landing error is less than the accuracy threshold, and completing accurate landing.

9. An electronic device, comprising: comprise a processor and a memory, the processor is used to execute the computer program stored in the memory to realize the substation AGV full-link intelligent navigation method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores at least one instruction, and the at least one instruction is executed by the processor to implement the substation AGV full-link intelligent navigation method in any one of claims 1 to 7.

Images