Intelligent inspection robot path optimization method and system based on edge reasoning model

Abstract

The invention provides an intelligent inspection robot path optimization method and system based on an edge inference model, and the method comprises the steps: obtaining an environment feature topological graph of a target inspection region, and determining an initial edge inference model; incremental training is carried out on the initial edge reasoning model through the real-time environment perception data flow, and a dynamic reasoning model adaptive to the current environment characteristics is generated; performing priority scoring on each path node in the environment characteristic topological graph based on a dynamic reasoning model, and generating an initial optimization path sequence; triggering a feedback calibration mechanism of the dynamic reasoning model according to the environment perception data flow updated in real time, performing dynamic path node replacement on the initial optimization path sequence, and generating a final inspection path; and controlling the intelligent inspection robot to execute an inspection task according to the final inspection path, and continuously collecting new environment sensing data streams in the inspection task process to update a parameter set of the dynamic reasoning model. According to the invention, the adaptability and reliability of path planning to a complex dynamic environment can be improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and more particularly, to an intelligent inspection robot path optimization method and system based on an edge inference model. Background Art

[0002] Currently, the path optimization methods of intelligent inspection robots mostly rely on static environmental maps or preset rules for path planning, lacking the ability to perceive dynamic environmental features in real time and make adaptive adjustments. For example, traditional methods usually generate inspection paths based on fixed topological structures or historical environmental data, and select path nodes through preset obstacle avoidance rules or energy consumption thresholds. However, when facing real-time changes in obstacle distribution, sudden environmental events, or complex and variable energy consumption scenarios, such methods are often difficult to effectively update the path decision logic, resulting in low obstacle avoidance efficiency or insufficient energy utilization rate of the inspection path. In addition, the evaluation of path nodes in the prior art mostly uses single-dimensional indicators (such as the shortest distance or the lowest energy consumption), lacking the spatio-temporal correlation analysis of multi-modal environmental features (such as dynamic obstacle density, ground reflection intensity, and thermal radiation changes), which is likely to cause local optimality of path planning or inaccurate global risk assessment. In addition, the conventional path optimization model usually keeps the parameters fixed after deployment and cannot perform incremental learning through real-time environmental data streams, resulting in deviations between the model inference results and the current environmental features. Especially in long-term operation, problems such as path node score failure and dynamic connection weight mismatch are likely to occur. These defects make the existing methods prone to risks such as redundant inspection paths, delayed emergency obstacle avoidance response, or over-budget energy consumption in complex industrial scenarios, seriously restricting the operation efficiency and reliability of intelligent inspection robots. Summary of the Invention

[0003] The purpose of the present invention is to provide an intelligent inspection robot path optimization method and system based on an edge inference model. The embodiments of the present invention are implemented as follows: In a first aspect, an embodiment of the present invention provides an intelligent inspection robot path optimization method based on an edge inference model, the method comprising: Obtaining an environmental feature topology map of a target inspection area and determining an initial edge inference model; the environmental feature topology map includes a plurality of path nodes and dynamic connection weights between the path nodes, and each path node is associated with a real-time environmental perception data stream; the initial edge inference model includes a multi-layer feature fusion module and a path decision module; Performing incremental training on the initial edge inference model through the real-time environmental perception data stream to generate a dynamic inference model adapted to the current environmental features; the incremental training includes adjusting the parameters of the multi-layer feature fusion module according to the spatio-temporal distribution differences of the environmental perception data stream; Perform priority scoring on each path node in the environmental feature topology map based on the dynamic inference model to generate an initial optimized path sequence; the priority scoring includes an obstacle density correlation factor and an energy consumption correlation factor between path nodes; Trigger the feedback calibration mechanism of the dynamic inference model according to the real-time updated environmental perception data stream, perform dynamic path node replacement on the initial optimized path sequence, and generate a final inspection path; the feedback calibration mechanism includes performing backpropagation compensation on the connection weights of path nodes; Control the intelligent inspection robot to perform inspection tasks according to the final inspection path, and continuously collect new environmental perception data streams during the inspection tasks to update the parameter set of the dynamic inference model.

[0004] In a second aspect, an embodiment of the present invention provides a computer system, including: One or more processors; A memory; One or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method described above is implemented.

[0005] The beneficial effects of the present invention at least include: The intelligent inspection robot path optimization method based on the edge inference model provided by the present invention realizes the intelligent generation and dynamic adjustment of the inspection path by obtaining the environmental feature topology map of the target inspection area and constructing an initial edge inference model, and combining real-time environmental perception data streams for dynamic model optimization and path decision-making. Specifically, the setting of dynamic connection weights in the environmental feature topology map can reflect the change in the passage difficulty of path nodes in real time, and combining the multi-layer feature fusion module to model the spatio-temporal differences of multi-modal environmental features effectively improves the accuracy of path node evaluation. By the priority scoring mechanism, an initial path sequence is generated by comprehensively considering the obstacle density and energy consumption factors, which can balance path safety and inspection efficiency. In addition, the feedback calibration mechanism dynamically adjusts the node connection weights through backpropagation compensation, realizing a fast response to sudden environmental changes and ensuring the robustness of path planning. Finally, through the parameter synchronization mechanism between the edge computing node and the local controller, the dynamic inference model can continuously adapt to the evolution of environmental features, forming a closed-loop optimized intelligent inspection decision-making system. This method not only improves the adaptability of path planning to complex dynamic environments, but also significantly reduces the energy consumption and task execution risks of robot inspections through multi-factor collaborative optimization.

[0006] In the following description, other features will be partly stated. When examining the following content and the accompanying drawings, those skilled in the art will partly discover these features, or can learn about these features through production or application. By practicing or using various aspects of the methods, tools, and combinations listed in the detailed examples described later, the features in the current application can be realized and obtained. Brief Description of the Drawings

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for describing the embodiments of the present invention will be briefly introduced below.

[0008] Figure 1 It is a flowchart of an intelligent inspection robot path optimization method based on an edge inference model provided by an embodiment of the present invention.

[0009] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. Detailed Embodiments

[0010] The embodiments of the present invention will be described below with reference to the accompanying drawings in the embodiments of the present invention. The terms used in the implementation part of the embodiments of the present invention are only used to explain the specific embodiments of the present invention, and are not intended to limit the present invention.

[0011] In the embodiments of the present invention, the execution subject of the intelligent inspection robot path optimization method based on the edge inference model is a computer system, including but not limited to servers, personal computers, laptops, tablets, smart phones, etc. The computer system can be embedded in the intelligent inspection robot or set in the background control center and communicatively connected to the intelligent inspection robot. The application scenarios of the intelligent inspection robot path optimization method based on the edge inference model provided by the embodiments of the present invention are not limited. In the subsequent embodiment introductions, the application scenarios where the present invention can be applied will be listed as much as possible.

[0012] The embodiments of the present invention provide an intelligent inspection robot path optimization method based on an edge inference model as Figure 1 shown, and the method includes: Step S100: Obtain the environmental feature topology map of the target inspection area and determine the initial edge inference model; the environmental feature topology map includes multiple path nodes and the dynamic connection weights between the path nodes, and each path node is associated with a real-time environmental perception data stream; the initial edge inference model includes a multi-layer feature fusion module and a path decision module.

[0013] In this step, the environmental feature topology map of the target inspection area is, for example, a digital regional structure map constructed by multi-dimensional environmental perception technology, and its core elements include path nodes, dynamic connection weights between path nodes, and real-time environmental perception data streams associated with nodes. Specifically, path nodes are discrete coordinate points formed after spatial grid division in the target inspection area. Each path node represents a physical location that the robot can pass or needs to detect. The coordinates of the path node are calculated through environmental point cloud data collected by a three-dimensional laser scanner or a multi-spectral sensor. The dynamic connection weight is a dynamic parameter used to quantify the difficulty of passage between adjacent path nodes. Its value is negatively correlated with the obstacle distribution density, ground slope change rate, and energy consumption estimate between path nodes. For example, when there is an obstacle between two path nodes, its dynamic connection weight will be dynamically adjusted down according to the real-time density of the obstacle. The real-time environmental perception data stream is the environmental feature data continuously collected at a fixed sampling frequency by the sensor network (including laser radar, infrared thermal imager, and sonar array) deployed in the inspection area. The real-time environmental perception data stream associated with each path node contains the temperature gradient distribution, obstacle geometric contour, and voiceprint fluctuation characteristics within a preset radius around the node. The initial edge inference model is a neural network model pre-trained in the cloud. It consists of a multi-layer feature fusion module and a path decision module: the multi-layer feature fusion module is responsible for cross-modal feature alignment and spatiotemporal correlation analysis of multimodal environmental perception data streams, such as fusing temperature features with obstacle contour features through a cross-modal attention mechanism; the path decision module generates a path node's priority score based on the fused feature tensor, and its decision threshold is initialized by the optimization objective function of the historical inspection task.

[0014] During the implementation process, for the construction of the environmental feature topology map, for example, first, a three-dimensional laser scanner carried by a drone is used to conduct full-area coverage scanning to generate an environmental point cloud dataset with centimeter-level accuracy. When performing spatial gridding processing on the point cloud data, a non-uniform grid division algorithm is adopted, and the grid density is automatically increased in areas with dense obstacles to ensure that the distribution density of path nodes matches the regional complexity. The projected area of obstacles, the mean ground reflectivity, and the standard deviation of the slope angle are calculated within each grid cell to generate a gridded environmental attribute matrix. Based on this matrix, the morphological erosion algorithm is used to eliminate narrow areas where the robot cannot pass, and the contour center points of the continuous passable areas are retained as candidate path nodes. Further, density clustering optimization is performed on the candidate path nodes, and redundant nodes with a distance less than the minimum turning diameter of the robot are merged to form a set of backbone path nodes. The dynamic connection weights between adjacent path nodes are calculated based on the weighted sum of the ground slope change rate and the projected area of obstacles, where the weight coefficient of the slope change rate is set to 0.6, and the weight coefficient of the projected area of obstacles is set to 0.4. Finally, the initial dynamic connection weight matrix is generated through normalization processing. The input layer dimension of the initial edge inference model is strictly aligned with the topological structure of the path node connection map. For example, if the environmental feature topology map contains N path nodes, the input layer is designed as an N×K-dimensional tensor (K is the number of feature channels of the single-node environmental perception data stream), and the spatial coordinates of the path nodes are mapped to a 128-dimensional high-dimensional vector through an embedding layer to encode the geographical location correlation of the nodes.

[0015] Step S200: Incrementally train the initial edge inference model through the real-time environmental perception data stream to generate a dynamic inference model adapted to the current environmental features; the incremental training includes adjusting the parameters of the multi-layer feature fusion module according to the spatio-temporal distribution differences of the environmental perception data stream.

[0016] In this step, the incremental training is, for example, a machine learning process of dynamically fine-tuning the model parameters through the real-time environmental perception data stream without resetting the existing knowledge of the initial edge inference model. The spatio-temporal distribution differences of the real-time environmental perception data stream are manifested as feature offsets of the multi-modal data collected by the sensors in the time dimension and the space dimension. For example, the periodic fluctuations of the temperature gradient distribution due to day-night alternation, or the spatial coordinate changes of the obstacle positions caused by the movement of dynamic objects. The parameter adjustment of the multi-layer feature fusion module needs to focus on solving the cross-modal feature alignment problem: in the time dimension, the dynamic time warping algorithm is used to eliminate the feature misalignment caused by the data acquisition delay of different sensors; in the space dimension, the channel attention mechanism is used to assign adaptive weights to the features of each modality. For example, a higher spatial resolution weight is assigned to the lidar point cloud data, while a higher temperature sensitivity weight is assigned to the infrared thermal imaging data.

[0017] In specific implementation, the incremental training process first extracts timestamp-aligned multi-modal feature segments from the real-time environment perception data stream. Each feature segment contains data for consecutive T sampling periods, where the value of T is inversely proportional to the robot's movement speed (for example, when the robot speed is 0.5 m / s, T is set to 10 seconds to ensure that the feature segment covers the environmental changes within a 5-meter moving distance). After inputting the feature segment into the multi-modal feature fusion module of the initial edge inference model, the correlation matrix of each modal feature is calculated through the cross-modal attention allocation unit. For example, if there is a high-temperature area in the current environment, the attention weight of the thermal radiation feature channel is automatically increased to more than 0.7 to enhance the model's sensitivity to the heat source distribution. The feature channel alignment unit further performs dimensionality reduction on the weighted feature tensor, eliminating redundant feature channels and retaining key spatio-temporal correlation features. The design of the loss function needs to consider both the path prediction accuracy and the model stability: the mean square error is used to measure the deviation between the path node priority score and the true passage efficiency, and an elastic weight consolidation regular term is introduced to prevent the catastrophic forgetting problem during the incremental training process. When the change rate of the loss function gradient is continuously lower than the preset threshold (such as 1e-5) for three consecutive iterations, it is determined that the model converges, the parameters of the multi-modal feature fusion module are frozen, and the optimized path decision module and the feature fusion module are integrated and output as a dynamic inference model.

[0018] Step S300: Based on the dynamic inference model, perform priority scoring on each path node in the environmental feature topology map to generate an initial optimized path sequence; the priority scoring includes the obstacle density correlation factor and the energy consumption correlation factor between path nodes.

[0019] The priority scoring is a comprehensive evaluation index of the path node by the dynamic inference model, and its core consists of the obstacle density correlation factor and the energy consumption correlation factor. The obstacle density correlation factor quantifies the distribution density and dynamic change trend of obstacles within a preset radius around the path node. Its calculation process, for example, includes: extracting the geometric features of the obstacle contour through lidar point cloud data, and using the fast Fourier transform to analyze the high-frequency component energy of the contour geometric features. The higher the energy value, the denser the obstacle distribution or the more frequent the movement. The energy consumption correlation factor predicts the energy consumption required for the robot to move from the current node to the adjacent node. Its calculation model integrates the ground slope angle, the ground friction coefficient, and the robot motor torque characteristics. For example, when the ground slope angle between path nodes exceeds 15 degrees, the energy consumption correlation factor increases according to an exponential function.

[0020] During the implementation process, the priority score of each path node in the dynamic inference model is generated through the following process: First, the real-time environment perception data stream associated with the path node is input into the multi-layer feature fusion module, and a 128-dimensional feature vector containing spatio-temporal correlation features is output. Then, the path decision module performs multi-scale convolution operations on the feature vector to extract local environment features under different receptive fields (for example, a 3×3 convolution kernel captures the distribution of adjacent obstacles, and a 5×5 convolution kernel identifies the terrain undulation at the regional level). The obstacle density correlation factor is obtained by weighted summation of the energy of the high-frequency components of the convolution feature map. The specific formula is: ODF = Σ(FFT(featuremap[:,k]) × wk), where wk is the weight coefficient of the k-th high-frequency channel, which is determined by training with historical data. The energy consumption correlation factor maps the convolution features to a preset energy consumption evaluation space through a fully connected layer, and its output value is linked to the robot motor power model to ensure that the error between the predicted value and the actual energy consumption is less than 5%. Finally, the priority score is calculated by the weighted sum of ODF and ECF, and the weight coefficients are dynamically adjusted according to the inspection task type (for example, the ODF weight is set to 0.7 in safety-priority tasks, and the ECF weight is set to 0.6 in energy-efficiency-priority tasks). The initial optimized path sequence is generated using an improved A* algorithm, which incorporates the priority score into the traditional path cost function to ensure that high-score nodes are preferentially included in the inspection path.

[0021] Step S400: Trigger the feedback calibration mechanism of the dynamic inference model according to the real-time updated environment perception data stream, perform dynamic path node replacement on the initial optimized path sequence, and generate the final inspection path; the feedback calibration mechanism includes backpropagation compensation for the connection weights of the path nodes.

[0022] The feedback calibration mechanism is a self-correction module for the dynamic inference model to cope with environmental mutations. Its triggering conditions include abnormal fluctuations in the environment perception data stream (such as a temperature gradient mutation exceeding 10°C / m or an instantaneous increase in obstacle density by 50%). When an abnormal fluctuation is detected, the system automatically intercepts the environmental feature subgraph within the current time window (including the abnormal area and the path nodes within 3 meters around it), and inputs it into the backpropagation compensation unit. The backpropagation compensation unit compensates and corrects the dynamic connection weights of the path nodes by calculating the error gradient between the current model prediction value and the actual passage efficiency. For example, if the actual passage time of a certain node far exceeds the predicted value due to a sudden obstacle, the compensation amount of its connection weight ΔW = η × (tactual - tpredicted), where η is the learning rate, and tactual and tpredicted are the actual and predicted passage times respectively.

[0023] The specific process of dynamic path node replacement is, for example: recalculate the path node priority scores of the affected area based on the corrected dynamic connection weights. For nodes with scores lower than the replacement threshold (such as a score drop exceeding 20%) in the initial optimized path sequence, the system retrieves alternative nodes from the environmental feature topology map. The selection of alternative nodes needs to meet three constraints: spatial adjacency, a difference in priority scores less than 5%, and an increase in energy consumption lower than 10%. After the replacement operation is completed, the path planning engine uses Dijkstra's algorithm to globally optimize the updated node sequence and eliminate path detours caused by local replacement. The output of the final inspection path needs to ensure that the total path score is increased by at least 15% compared to the initial path, and the total energy consumption does not exceed 85% of the upper limit of the robot's battery capacity.

[0024] Step S500: Control the intelligent inspection robot to perform inspection tasks according to the final inspection path, and continuously collect new environmental perception data streams during the inspection tasks to update the parameter set of the dynamic inference model.

[0025] In this step, exemplarily, the motion control of the intelligent inspection robot adopts a hierarchical decision-making architecture: the upper-layer path tracking controller decomposes the final inspection path into a series of waypoints, and the lower-layer motion controller adjusts the hub motor speed through the PID algorithm to ensure that the robot moves along a trajectory with a preset heading angle deviation less than 2 degrees. The new environmental perception data stream acquisition module integrates a multi-spectral sensor, an inertial measurement unit, and an odometer to synchronously acquire infrared thermal imaging sequences, lidar point cloud frames, and robot pose data at a frequency of 10Hz. The data preprocessing link performs spatio-temporal alignment on the multi-source data. For example, the IMU and odometer data are fused through an extended Kalman filter to eliminate the cumulative error of pose estimation.

[0026] The core of the parameter update process is the incremental feature space similarity comparison: calculate the Wasserstein distance between the newly acquired incremental spatio-temporal feature tensor and the benchmark feature tensor in the historical training dataset to quantify the distribution difference between the two. When the distance value exceeds the preset similarity threshold, the system activates the parameter update engine to separate the feature segments that deviate significantly from the historical distribution (such as emerging high-temperature regions or clusters of moving obstacles) from the incremental data. These abnormal feature segments extract key patterns through knowledge distillation technology and inject them into the online training loop of the path decision module. The parameter adjustment adopts a federated learning framework to ensure that the model synchronization error between multiple edge nodes is less than 1%. The updated parameter set needs to be tested by an offline verification unit. The verification criteria include path generation stability (the coincidence degree of the planned paths for 10 consecutive times is greater than 90%) and real-time performance (the single planning time is less than 200ms). Finally, the verified parameters are deployed to all online robots through the distributed communication protocol of the edge computing node to complete the closed-loop optimization.

[0027] As an implementation manner, in step S100, obtain the environmental feature topology map of the target inspection area and determine the initial edge inference model, which may specifically include: Step S110: Use a three-dimensional laser scanner carried by a drone deployed in the target inspection area to collect multi-angle environmental data and generate environmental point cloud data covering the entire area.

[0028] In this step, the three-dimensional laser scanner carried by the drone scans the target inspection area at multiple angles along a preset flight trajectory. The multi-angle scanning includes three modes: vertical downward scanning, 45-degree inclined scanning, and horizontal circumferential scanning to ensure complete coverage of complex terrains and three-dimensional obstacles. The environmental point cloud data consists of millions of three-dimensional coordinate points obtained by the lidar in each scanning cycle. Each coordinate point contains spatial position (X, Y, Z), reflection intensity, and timestamp information. For example, in the vertical downward scanning mode, the lidar emits pulses at a frequency of 10 Hz and calculates the distance of each point from the sensor through the time-of-flight (ToF) principle to generate high-precision terrain elevation data. In the horizontal circumferential scanning mode, the lidar rotates at 30-degree angular intervals to capture the side profiles of vertical structures such as columns and pipes. During the data collection process, the drone controls the positioning error within ±2 cm through real-time kinematic (RTK) technology and compensates for the point cloud distortion caused by changes in the flight attitude through an inertial measurement unit (IMU). The finally generated environmental point cloud dataset is stored through a spatial index structure (such as an octree) to support rapid retrieval of the point density distribution and reflection intensity statistical characteristics in any area.

[0029] Step S120: Perform spatial grid division on the environmental point cloud data, and statistically analyze the obstacle height distribution characteristics and ground reflection intensity characteristics in each grid unit to generate a grid-based environmental attribute matrix.

[0030] Exemplarily, an adaptive resolution strategy is adopted for spatial grid division, and the grid size is dynamically adjusted according to the point cloud density. For example, a coarse-grained grid of 1m×1m is used in flat areas, and a fine-grained grid of 0.2m×0.2m is switched to in areas with dense obstacles. Three types of key attributes are calculated within each grid cell: the obstacle height distribution feature is obtained by statistically calculating the difference between the maximum and minimum values of the Z coordinates of all points within the grid. For example, if the Z coordinate range of a certain grid is [1.2m, 2.5m], then the obstacle height distribution feature value is 1.3m; the ground reflection intensity feature is extracted from the mode of the reflection intensity values within the grid and is used to distinguish different material surfaces (such as metal reflection intensity > 80, asphalt reflection intensity ≈ 40); the ground slope feature is calculated by fitting the normal vector inclination angle of the three-dimensional plane equation of the points within the grid. The grid-based environmental attribute matrix is stored in the form of a two-dimensional array, and each element contains the above three types of features and the coordinates of the grid center point. For example, the matrix element M[i][j] = {obstacle height: 1.3m, reflection intensity: 45, slope angle: 5°, center coordinates: (x, y)}. This matrix provides a structured data basis for subsequent path node extraction and connection weight calculation.

[0031] Step S130: Identify the boundaries of the passable area based on the grid-based environmental attribute matrix, extract the position coordinates of the key path nodes, and mark the physical connection relationship between the path nodes.

[0032] Exemplarily, the identification of the passable area boundary depends on the binary processing of the grid-based environmental attribute matrix. For example, grids with an obstacle height lower than the robot chassis height (such as 0.5m) and a slope angle less than 15 degrees are marked as passable units, and the rest are marked as obstacle units. Holes within the passable area are eliminated through morphological closing operations, and the contour polygon of the continuous area is extracted using an edge detection algorithm. The position coordinates of the key path nodes are selected from the vertices and the center point of the contour polygon. For example, for a rectangular contour, four vertices and the center point are extracted as candidate path nodes. The determination of the physical connection relationship is based on the line visibility criterion: if all the grids passed by the connection line between two nodes are passable units, a two-way connection edge is established. During this process, the Bresenham algorithm is used to traverse the grids passed by the connection line to detect whether there is an obstacle unit blocking. For example, the connectivity between node A(10, 20) and node B(30, 40) needs to verify the attribute status of 20 intermediate grids. The finally generated physical connection relationship table records all valid connection edges and their geometric lengths for use in dynamic connection weight calculation.

[0033] As an implementation manner, step S130, identifying the boundaries of the passable area based on the grid-based environmental attribute matrix, extracting the position coordinates of the key path nodes, and marking the physical connection relationship between the path nodes, includes: Step S131: Binarize the gridded environmental attribute matrix. Mark the grids where the obstacle height is lower than the height of the robot chassis as passable units, and the rest as obstacle units.

[0034] Exemplarily, the binarization process dynamically sets a threshold based on the physical parameters of the robot. For example, when the height of the robot chassis is 0.5m, the obstacle height threshold is set to 0.5m + safety margin 0.1m = 0.6m. For each grid cell, if its obstacle height distribution eigenvalue ≤ 0.6m and the slope angle < 15 degrees, it is marked as a passable unit; otherwise, it is marked as an obstacle unit. For example, a grid with an obstacle height of 0.8m (exceeding the threshold) and a slope angle of 10 degrees is still classified as an obstacle unit; while a grid with a height of 0.4m and a slope angle of 12 degrees is marked as a passable unit. The processed binary matrix eliminates isolated noise points through an erosion operation: Use a 3×3 structuring element to erode the passable area and remove small connected regions with an area less than 0.5m² to ensure the continuity of the robot's travel path.

[0035] Step S132: Detect the contours of continuous regions in the passable units and extract the center points of each contour as the initial positions of the candidate path nodes.

[0036] Exemplarily, the detection of continuous region contours can adopt an edge tracking algorithm. For example, start scanning from the bottom-left grid. When an unmarked passable unit is found, start clockwise boundary tracking and record the vertex coordinates of the contour polygon. The center points of each contour are obtained by calculating the geometric median of the polygon vertices. For example, the four vertices of a rectangular contour are (10,20), (30,20), (30,40), (10,40), and its center point is calculated as (20,30). The initial position set of the candidate path nodes also needs to be spatially homogenized: Generate grid points at 2m intervals within large connected regions (area > 50m²) to ensure that the path node density matches the region scale.

[0037] Step S133: Perform density clustering analysis on the initial positions of the candidate path nodes, and merge adjacent path nodes with a spacing less than the minimum turning radius of the robot to form a set of backbone path nodes.

[0038] Exemplarily, density clustering can adopt the DBSCAN algorithm. For example, the parameters are set as follows: the neighborhood radius ε = the minimum turning radius of the robot × 1.2 (e.g., 0.6m × 1.2 = 0.72m), and the minimum number of samples MinPts = 3. If multiple nodes form a cluster within the ε neighborhood and the number of samples ≥ MinPts, then the cluster is merged into a backbone node, and the position is taken as the geometric center of the points within the cluster. For example, the coordinates of three candidate nodes are (10.1, 20.2), (10.3, 20.5), and (10.5, 20.3) respectively, and the cluster center is calculated as (10.3, 20.33). This operation eliminates redundant nodes and ensures that the path spacing meets the kinematic constraints of the robot.

[0039] Step S134: Select the path node with the maximum distance from the boundary of the obstacle unit in the set of backbone path nodes as the safe path node, and record the straight-line visible distance between each pair of safe path nodes.

[0040] Exemplarily, the safe path nodes can be selected according to the minimum safety distance criterion. For example, calculate the Euclidean distance from each backbone node to the boundary of the nearest obstacle unit, and select the nodes with the top 20% maximum distances as the safe path nodes. For example, if the distance of a certain node to the obstacle boundary is 3.5m, exceeding the distance values of 80% of the other nodes (all < 2m), then it is marked as a safe node. The straight-line visible distance is calculated by the ray casting algorithm: emit rays from the current safe node to other safe nodes, and record the length of the rays not blocked by the obstacle units. This distance value is used for risk assessment in subsequent path optimization. For example, when the visible distance between two nodes > 10m, it is preferentially included in the inspection path.

[0041] Step S136: Generate the physical connection relationship between path nodes according to the straight-line visible distance, and establish a two-way connection edge when there is a continuous passable unit between two path nodes and it is not blocked by the obstacle unit.

[0042] Exemplarily, the determination of the physical connection relationship comprehensively considers the straight-line visible distance and path passability. For example, for two safe nodes, if their straight-line visible distance ≤ the maximum allowable detection distance (e.g., 50m) and the grids passed by the connection line are all passable units, then a two-way connection edge is established. The attributes of the connection edge include geometric length, average slope angle, and maximum obstacle height difference. For example, the connection edge from node A to node B records a length of 15.3m, an average slope angle of 4 degrees, and a maximum height difference of 0.3m. This information provides the basic data for the real-time update of the dynamic connection weight.

[0043] Step S137: Assign an initial dynamic connection weight to each two-way connection edge, and the initial dynamic connection weight is dynamically adjusted according to the average ground reflection intensity of the grid cells passed by the connection edge.

[0044] Exemplarily, the calculation of the initial dynamic connection weight can introduce a ground reflection intensity factor. For example, the higher the average reflection intensity (e.g., ≥60), it indicates that the ground material is smoother (such as metal or tile), and the weight value increases accordingly; a low average reflection intensity (e.g., ≤30) may correspond to a rough ground (such as gravel), and the weight value is adjusted downward. The adjustment formula can be, for example: W = Wbase×(1 + 0.2×(Ravg−50) / 50), where Wbase is the base weight calculated in step S140, and Ravg is the average ground reflection intensity of the connection edge passing through the grid. For example, when Ravg = 70, W = 0.54×(1 + 0.2×20 / 50) = 0.54×1.08 = 0.58, which improves the passage priority.

[0045] Step S138: Combine the set of safe path nodes, physical connection relationships, and initial dynamic connection weights into an initial version of the environmental feature topology map for subsequent loading and parsing by the dynamic inference model.

[0046] Exemplarily, the initial version of the environmental feature topology map can be stored in JSON format, including a node list (including coordinates, safety level, sensor ID), an edge list (including start point, end point, weight, passage status), and global parameters (such as area size, coordinate system type). When the model is loaded, the JSON data is converted into a graph data structure by a parser, and a spatial index is constructed to accelerate the query of neighbor nodes. For example, querying all nodes within a radius of 5m of the node (20, 30) can be completed within 1ms. This initial topology map serves as a benchmark for incremental optimization and dynamically updates node attributes and connection weights according to real-time data during the inspection process.

[0047] Step S140: Calculate the initial dynamic connection weight according to the ground slope change rate and the obstacle projection overlap area between adjacent path nodes. The initial dynamic connection weight is negatively correlated with the ground slope change rate and negatively correlated with the obstacle projection overlap area.

[0048] Exemplarily, the calculation model of the initial dynamic connection weight integrates terrain complexity and obstacle distribution factors. For example, the ground slope change rate is calculated by the ratio of the elevation difference to the horizontal distance between path nodes. For example, if the horizontal distance between node A and node B is 5m and the elevation difference is 0.8m, then the slope change rate is arctan(0.8 / 5)≈9 degrees; the obstacle projection overlapping area refers to the intersection area of the projection area of the line connecting two nodes on the horizontal plane and the obstacle grid, which is realized by rasterizing the projection area and counting the number of overlapping grids. The weight calculation formula can be, for example: W = α×(1−slope change rate / maximum allowable slope)+β×(1−overlapping area / maximum allowable area), where α = 0.6, β = 0.4 are empirical coefficients, the maximum allowable slope is set to 20 degrees, and the maximum allowable area is set to 2m². For example, if the slope change rate of a connection edge is 12 degrees and the overlapping area is 0.5m², then W = 0.6×(1−12 / 20)+0.4×(1−0.5 / 2)=0.6×0.4+0.4×0.75 = 0.54. This weight value is normalized to the [0,1] interval, and the larger the value, the higher the path passing priority.

[0049] Step S150: Integrate the position coordinates, physical connection relationship, and initial dynamic connection weight of the key path nodes into the path node connection map of the environmental feature topology map.

[0050] Exemplarily, the path node connection map can be stored in a graph structure. The node attributes include three-dimensional coordinates, the type of the area it belongs to (such as corridor, equipment area), and the real-time sensor ID; the edge attributes include dynamic connection weight, physical connection status, and historical passing time records. The topological structure of the map is represented by both an adjacency matrix and an adjacency list: the adjacency matrix is used to quickly query the connection weight between any two nodes. For example, the matrix element Adj[i][j]=0.54 represents the weight value from node i to node j; the adjacency list records all reachable neighbor node lists of each node, supporting efficient traversal operations. In addition, a spatial index structure (such as an R-tree) is embedded in the map to achieve fast retrieval of nodes based on a geographical fence. For example, the operation time for finding all nodes within a radius of 5m is less than 10ms.

[0051] Step S160: Download the pre-trained general edge inference model from the cloud model library as the initial edge inference model, and encode the topological structure of the path node connection map into the input feature dimension of the initial edge inference model.

[0052] Exemplarily, the pre-trained general edge inference model can adopt a graph convolutional neural network (GCN) architecture. Its input layer is designed as an N×F-dimensional tensor, where N is the number of path nodes and F is the single-node feature dimension (including coordinates, connection weights, and environmental attributes). The topological structure encoding process includes: converting the node coordinates to a relative coordinate system (with the regional center as the origin) and normalizing them to the range of [−1,1]; the dynamic connection weights are processed through logarithmic transformation to alleviate the long-tail distribution problem. For example, if the original coordinates of a certain node are (105.3, 78.2) and the regional center is (100, 80), the encoded coordinates are (5.3, −1.8), and after normalization, they are (0.053, −0.018). The model input feature vector is composed of the coordinate encoding values, connection weights, and ground reflection intensity, with the dimension F = 5 (X, Y, W, reflection intensity, slope angle). The pre-trained model in the cloud model library is trained based on millions of simulation environment data and has the ability to transfer across scenarios.

[0053] Step S170: Map the path node position coordinates into a high-dimensional space vector through the embedding layer of the initial edge inference model.

[0054] Exemplarily, the embedding layer can adopt a 128-dimensional dense vector space, and the mapping function is E(x,y)=ReLU(We·[x,y]+be), where We is the trainable weight matrix and be is the bias term. Through this mapping, the geographical location correlation of path nodes is encoded as a geometric relationship in the vector space. For example, the cosine similarity between adjacent node vectors > 0.9, while the similarity between distant nodes approaches 0. For example, the cosine similarity between the embedding vectors of node A(0.1, 0.2) and its adjacent node B(0.12, 0.25) is 0.93, and the similarity with the distant node C(−0.5, 0.8) is 0.15. This high-dimensional encoding enables the model to capture the implicit patterns in the spatial topology (such as the symmetry of the circular path) and improve the generalization ability of path decision-making.

[0055] As an implementation, in step S200, incrementally train the initial edge inference model through the real-time environment perception data stream to generate a dynamic inference model adapted to the current environmental characteristics, including: Step S210: Extract multi-modal environmental feature vectors from the real-time environment perception data stream. The multi-modal environmental feature vectors include thermal radiation distribution features, voiceprint fluctuation features, and visual texture features.

[0056] In this step, for example, the real-time environment perception data stream is continuously collected by the multispectral sensor array carried by the intelligent inspection robot at a fixed sampling frequency, and the multimodal environment feature vector is a structured data representation after feature engineering processing. The thermal radiation distribution feature is captured by an infrared thermal imager at a frequency of 5Hz, and its value range is -20℃ to 100℃. Each thermal radiation distribution feature vector contains a temperature gradient distribution matrix in a 3m×3m area around the current path node. The matrix elements represent the temperature mean of a 0.1m×0.1m sub-area. For example, the thermal radiation distribution feature vector of a node is a 64-dimensional vector (8×8 grid), the temperature value of the central area is 45℃, and the edge area is 32℃. The voiceprint fluctuation feature is collected by an ultrasonic sensor array, and the energy spectrum density of the 0.1kHz to 40kHz frequency band is extracted by fast Fourier transform (FFT) to form a 128-dimensional frequency domain feature vector. For example, the detection of a sudden increase in energy in the 20kHz frequency band indicates the presence of metal structure vibration. Visual texture features are acquired by binocular cameras at a frame rate of 30fps. The local binary pattern (LBP) algorithm is used to calculate the texture complexity of the image block and generate a 256-dimensional texture histogram vector. For example, the LBP histogram of the concrete surface is densely distributed in the low grayscale range, while the metal surface has a significant peak in the high grayscale range. The time alignment of the multimodal environment feature vector is achieved through hardware synchronization signals to ensure that the timestamp deviation of thermal radiation, voiceprint and visual data is less than 10ms.

[0057] Step S220: input the multimodal environment feature vector into a multi-layer feature fusion module to generate a fused spatiotemporal feature tensor; the multi-layer feature fusion module includes a cross-modal attention allocation unit and a feature channel alignment unit.

[0058] Exemplarily, the cross-modal attention allocation unit of the multi-layer feature fusion module can adopt a multi-head attention mechanism to allocate dynamic weights to features of different modalities. Specifically, the thermal radiation distribution feature, the voiceprint fluctuation feature, and the visual texture feature are respectively used as the query (Query), key (Key), and value (Value) to input into the attention calculation layer. The formula can be, for example: Attention(Q, K, V) = softmax((QKT) / √d)V, where d = 64 is the scaling factor. For example, when a high-temperature area is detected, the attention weight of the thermal radiation feature is increased to 0.8, and the weights of the voiceprint and visual features are respectively decreased to 0.1 and 0.1 to enhance the sensitivity of the model to the heat source. The feature channel alignment unit solves the sampling rate difference of multi-modal data through the dynamic time warping (DTW) algorithm: aligns the 128-dimensional vector of the voiceprint fluctuation feature and the 256-dimensional vector of the visual texture feature on the time axis to generate a 256-dimensional vector of a unified length. The dimension of the fused spatio-temporal feature tensor is N×T×512 (N is the number of path nodes, T is the time window length), where the first 256 dimensions are the aligned voiceprint-visual joint features, and the last 256 dimensions are the extended representation of the thermal radiation feature after spatial interpolation. This tensor encodes the spatio-temporal evolution pattern of environmental features, such as the spatial correlation between the temperature gradient diffusion direction and the voiceprint energy propagation path.

[0059] Step S230: Calculate the gradient of the loss function of the initial marginal inference model according to the dimension distribution difference of the spatio-temporal feature tensor, and iteratively optimize the decision threshold of the path decision module based on the gradient direction.

[0060] Exemplarily, the loss function is designed as a weighted sum of the mean square error (MSE) and the elastic weight consolidation (EWC) regularization term: L = α×MSE(ypred, ytrue) + β×Σλi(θi−θ’i)2, where α = 0.7, β = 0.3 are the balance coefficients, θi are the model parameters, θ’i are the initial model parameters, and λi are the diagonal elements of the Fisher information matrix. The dimension distribution difference of the spatio-temporal feature tensor is quantified by calculating the KL divergence of each feature channel. For example, when the KL divergence between the thermal radiation channel and the voiceprint channel exceeds 0.5, the regularization term is strengthened. The gradient calculation adopts the backpropagation algorithm, and the optimization of the decision threshold of the path decision module is implemented through the adaptive moment estimation (Adam) algorithm. The learning rate is set to 3e−4, and the momentum parameters are β1 = 0.9, β2 = 0.999. For example, the initial decision threshold is 0.5, and after three iterations, it is adjusted to 0.48, increasing the sensitivity of the priority score in the obstacle-dense area by 12%.

[0061] Step S240: When the change rate of the loss function gradient is lower than the preset convergence threshold, freeze the parameters of the multi-layer feature fusion module, and combine the optimized path decision module and the multi-layer feature fusion module into a dynamic inference model.

[0062] Exemplarily, the preset convergence threshold can be set to 1e−5, for example. When the absolute value of the gradient change rate for five consecutive iterations is less than this threshold, it is determined that the model has converged. The parameter freezing operation is achieved by setting a gradient mask to prevent parameter updates in the cross-modal attention allocation unit and the feature channel alignment unit. The optimized path decision module and the frozen multi-layer feature fusion module are integrated through a tensor splicing layer to form a dynamic inference model. The input layer of this model receives real-time environmental perception data streams, and the output layer generates path node priority scores, with the inference latency controlled within 50 ms (NVIDIA Jetson AGX Xavier platform). For example, the inference accuracy of the initial edge inference model is 85%, which is improved to 92% after optimization, while the energy consumption is reduced by 18%.

[0063] Step S250: Establish a mapping relationship between the output layer of the dynamic inference model and the path nodes of the environmental feature topology map to generate a path node priority scoring rule library.

[0064] Exemplarily, the mapping relationship can be achieved through a spatial index hash table. For example, the hash key is the normalized value of the path node coordinates (longitude, latitude, altitude), and the hash value is the priority score and the associated feature index. The path node priority scoring rule library contains two types of core factors: The obstacle density correlation factor is calculated by the weighted sum of the energies of the high-frequency components of the spatio-temporal feature tensor, and the weight coefficient is determined to be 0.6 through training with historical data; the energy consumption correlation factor fuses the predicted value of the motor torque and the terrain slope angle, and the weight coefficient is 0.4. For example, if the obstacle density correlation factor of node A is 0.75 and the energy consumption correlation factor is 0.35, then the priority score = 0.6×0.75 + 0.4×0.35 = 0.59. The scoring rule library is stored in JSON format and supports a dynamic update mechanism. When the structure of the environmental feature topology map changes, the rule weights are adjusted in real time through incremental learning to ensure the consistency of the path planning strategy with the physical environment.

[0065] As an implementation, in step S300, based on the dynamic inference model, priority scores are assigned to each path node in the environmental feature topology map to generate an initial optimized path sequence, including: Step S310: Traverse all path nodes in the environmental feature topology map to obtain the multi-modal environmental feature vectors corresponding to the real-time environmental perception data streams of each path node.

[0066] In this step, the real-time environmental perception data stream of the path nodes is continuously collected by the multi-sensor fusion system carried by the intelligent inspection robot, and the multimodal environmental feature vector is a quantitative representation of the environmental state in the preset physical space around each path node. Specifically, the real-time environmental perception data stream associated with each path node includes the following three types of features: thermal radiation distribution features are captured by an infrared thermal imager at a sampling frequency of 5 Hz to generate an 8×8 grid temperature gradient matrix. Each grid unit corresponds to a physical area of 0.5 m×0.5 m, and the temperature resolution is accurate to 0.1°C. For example, the temperature of the central area of the thermal radiation distribution feature vector of a path node is 52.3°C, and that of the edge area is 28.7°C; voiceprint fluctuation features are collected by an ultrasonic sensor array, and the energy spectrum density of the 0.1 kHz to 40 kHz frequency band is extracted by fast Fourier transform to form a 128-dimensional frequency domain feature vector. For example, the detection of an energy peak in the 15 kHz frequency band indicates the presence of mechanical friction vibration; visual texture features acquire image data through a binocular camera at a frame rate of 30 fps, and the texture complexity histogram of the image block is calculated using a local binary pattern algorithm to generate a 256-dimensional feature vector. For example, the texture histogram of the rusted metal surface is densely distributed in the low grayscale range, while the smooth ceramic surface has a significant peak in the high grayscale range. The time synchronization of multimodal environmental feature vectors is achieved through hardware trigger signals to ensure that the acquisition timestamp deviation of thermal radiation, voiceprint and visual data is less than 5ms to eliminate data fusion errors caused by timing misalignment. When traversing the environmental feature topology map, the system accesses each path node in spatial index order and extracts the multimodal environmental feature vector within a radius of 3m to form an input data matrix of dimension N×(64+128+256) (N is the total number of path nodes) for subsequent feature fusion module processing.

[0067] Step S320: Input the multimodal environment feature vector into the multi-layer feature fusion module of the dynamic reasoning model, and output the spatiotemporal feature tensor of each path node.

[0068] Exemplarily, the cross-modal attention allocation unit of the multi-layer feature fusion module adopts a multi-head attention mechanism to allocate dynamic weights to different modal features. Specifically, the thermal radiation distribution feature, the voiceprint fluctuation feature, and the visual texture feature are respectively input into the attention calculation layer as the query vector, the key vector, and the value vector. For example, when a certain path node detects a high-temperature anomaly (the central temperature of the thermal radiation feature exceeds 60°C), the attention mechanism raises the weight of the thermal radiation feature to 0.75, and the weights of the voiceprint and visual features are reduced to 0.15 and 0.10 respectively, so as to enhance the sensitivity of the model to the heat source distribution. The feature channel alignment unit eliminates the sampling rate differences of multi-modal data through the dynamic time warping algorithm: aligns the 128-dimensional vector of the voiceprint fluctuation feature with the 256-dimensional vector of the visual texture feature on the time axis to generate a 256-dimensional vector of a unified length. The dimension of the fused spatio-temporal feature tensor is N×T×512 (T is the time window length), where the first 256 dimensions are the aligned voiceprint-visual joint features, and the last 256 dimensions are the representation of the thermal radiation feature expanded by bilinear interpolation. For example, within the time window T = 10 seconds, the spatio-temporal feature tensor of a certain path node contains the joint encoding of the temperature diffusion rate, the voiceprint energy propagation direction, and the texture change trend, which is used to characterize the environmental evolution law of the node in the spatio-temporal dimension.

[0069] Step S330: Calculate the obstacle density correlation factor of each path node according to the energy distribution spectrum of the spatio-temporal feature tensor; wherein, the obstacle density correlation factor is positively correlated with the amplitude of the high-frequency component of the energy distribution spectrum.

[0070] Exemplarily, the energy distribution spectrum of the spatio-temporal feature tensor analyzes the frequency domain characteristics of each feature channel through the fast Fourier transform, and the amplitude of the high-frequency component reflects the mutation frequency of the environmental characteristics. Specifically, perform a per-channel FFT transform on the 512-dimensional spatio-temporal feature tensor, extract the energy amplitude in the frequency band from 1 Hz to 10 Hz, and the obstacle density correlation factor is calculated as: ; where the weight coefficient wk is determined by training with historical data, and the determination threshold of the high-frequency component is 5 Hz. For example, when there are moving obstacles (such as temporarily stacked equipment) around a certain path node, the energy amplitude of its voiceprint channel in the 8 Hz frequency band increases significantly, resulting in the obstacle density correlation factor jumping from the reference value of 0.45 to 0.82. This factor is linearly positively correlated with the dynamic density of obstacles. Experimental data shows that for every 0.1 increase in the obstacle density correlation factor, the traffic risk assessment value of the path node increases by 12%.

[0071] Step S340: Perform convolution processing on the spatio-temporal feature tensor of each path node through the path decision module to generate an energy consumption correlation factor; wherein, the energy consumption correlation factor is negatively correlated with the activation degree of the feature channels after convolution processing.

[0072] Exemplarily, the convolution processing of the path decision module adopts a multi-scale convolution kernel structure. The 3×3 convolution kernel is used in the first layer to extract local environment features (such as small-range slope changes), and the 5×5 convolution kernel is used in the second layer to capture regional terrain features (such as continuous uphill areas). The activation degree of the feature channels is quantified by the ReLU function, and the calculation formula is, for example: ; where H×W is the spatial dimension of the feature map. The energy consumption correlation factor is generated by inverse mapping according to the activation degree: ; For example, when the terrain complexity of a certain path node is relatively high (the activation degree of the feature map reaches 0.85), its energy consumption correlation factor drops to 0.15, indicating that the robot needs to consume more electric energy when passing through this node. The energy consumption model incorporates the predicted value of the motor torque. When the slope angle exceeds 10 degrees, the convolution kernel automatically enhances the feature response in the corresponding area to ensure that the energy consumption correlation factor accurately reflects the actual energy consumption level.

[0073] Step S350: Synthesize the weighted sum of the obstacle density correlation factor and the energy consumption correlation factor to generate the priority score of each path node, and generate the initial optimized path sequence in descending order of the priority score.

[0074] The calculation model of the priority score is, for example: ; where the weight coefficients α and β are dynamically adjusted according to the task type: in the safety-first mode (such as chemical plant inspection), α is set to 0.7 and β is set to 0.3; in the energy efficiency-first mode (such as warehouse routine inspection), α is set to 0.4 and β is set to 0.6. For example, the obstacle density correlation factor of a certain path node is 0.82 and the energy consumption correlation factor is 0.35. The priority score in the safety-first mode is 0.7×0.82 + 0.3×0.35 = 0.68, while in the energy efficiency-first mode it is 0.4×0.82 + 0.6×0.35 = 0.53. The initial optimized path sequence is generated using an improved A* algorithm, which incorporates the priority score into the traditional path cost function, and its heuristic function is designed as: ; where g(n) is the actual cost from the starting point to the current node n, h(n) is the estimated cost from the current node to the end point, and λ is the priority score adjustment coefficient (default value 1.2). When λ > 1, the algorithm tends to select high-score nodes. For example, in an obstacle-dense area, the algorithm automatically bypasses low-score nodes and generates an optimized path sequence with a total score improvement of more than 18%. The final initial optimized path sequence is stored in a linked list structure, and each node records its three-dimensional coordinates, priority score, and connection weight with adjacent nodes, supporting real-time dynamic adjustment and incremental update.

[0075] As an implementation mode, step S400 triggers a feedback calibration mechanism of a dynamic reasoning model according to the real-time updated environment perception data stream to perform dynamic path node replacement on the initial optimized path sequence, for example, including: Step S410: monitor abnormal fluctuation signals in the real-time environment perception data stream, and activate the feedback calibration mechanism when the fluctuation amplitude exceeds a preset safety threshold.

[0076] In this step, the feedback calibration mechanism is a self-correction module of the dynamic reasoning model to cope with sudden changes in the environment. Its real-time guarantee is achieved through the streaming data processing architecture of the edge computing node. The real-time updated environmental perception data stream is continuously collected by the multispectral sensor array carried by the intelligent inspection robot with millisecond delay. The data stream contains three types of core information: infrared thermal imaging sequence, lidar point cloud frame and high-definition video stream. The triggering conditions of the feedback calibration mechanism include but are not limited to the following abnormal fluctuation signals: the temperature gradient change rate of the thermal radiation distribution characteristics exceeds 10℃ / m in adjacent sampling periods, the energy of the voiceprint fluctuation characteristics in a specific frequency band (such as 20kHz-40kHz) increases by more than 50%, or the spatial continuity of the visual texture features is broken due to the addition of obstacles and exceeds the preset threshold. For example, when the lidar detects an obstacle with a height exceeding the robot chassis safety margin 3m in front of a path node, the system immediately activates the feedback calibration mechanism, interrupts the current path execution process and starts the dynamic path node replacement process.

[0077] The monitoring of abnormal fluctuation signals is based on the statistical feature analysis within the sliding time window: the system uses 1 second as the time window length to calculate the mean, variance and kurtosis indicators of the three types of data streams: thermal radiation, voiceprint and vision. The setting of the preset safety threshold is based on historical operation data and physical constraints: the threshold of the thermal radiation temperature gradient change rate is set to 10℃ / m (based on the tolerance limit of the robot's heat dissipation system), the threshold of the voiceprint energy mutation is set to 3 times the standard deviation of the benchmark value (according to the equipment vibration safety specification), and the threshold of the continuity break of the visual texture is set to the structural similarity index (SSIM) of adjacent image blocks less than 0.6. For example, when the temperature gradient of the thermal radiation distribution characteristics of a path node jumps from 5℃ / m to 15℃ / m (exceeding the threshold by 50%) within two consecutive sampling cycles, the system determines it as a thermal anomaly event and immediately triggers the feedback calibration mechanism. The calculation delay of the monitoring algorithm is controlled within 5ms to ensure the real-time response capability to sudden environmental changes.

[0078] Step S420: extracting the environmental feature subgraph corresponding to the abnormal fluctuation signal, and inputting the environmental feature subgraph into the back propagation compensation unit of the dynamic reasoning model.

[0079] Exemplarily, the environmental feature sub-graph is a local area topological structure intercepted from the environmental feature topological graph, and its coverage range is a spherical area (three-dimensional scene) or a rectangular area (two-dimensional scene) with a radius of 5m centered on the abnormal fluctuation signal source point. The process of extracting the sub-graph includes: quickly locating the set of affected path nodes through a spatial index hash table, and retrieving the associated multi-modal environmental feature vectors and connection weight data. For example, when an abnormal fluctuation signal occurs at the coordinates (X = 102.4, Y = 75.6, Z = 1.2), the system extracts 12 path nodes and their connecting edges within a radius of 5m from this point, and generates an environmental feature sub-graph tensor with a dimension of 12×512 (12 nodes, each node with 512-dimensional features). The backpropagation compensation unit adopts a lightweight neural network architecture. Its input layer receives the environmental feature sub-graph tensor, and the output layer generates the weight offset matrix of the path decision module. The core function of this unit is to calculate the compensation amount of the model parameters by comparing the prediction errors between the abnormal state and the historical normal state.

[0080] Step S430: Calculate the weight offset of the path decision module in the dynamic inference model through the backpropagation compensation unit, and adjust the scoring coefficient of the priority scoring rule library according to the weight offset.

[0081] The calculation of the weight offset is based on the contrast loss function: ; where η = 0.001 is the compensation learning rate, and L is the mean square error between the priority score prediction value and the actual traffic efficiency. For example, when the actual travel time of a certain path node is extended by 8 seconds compared with the predicted value due to a sudden obstacle, the backpropagation compensation unit calculates that its connection weight needs to be reduced by 0.15 to reflect the increased traffic risk. The adjustment of the priority scoring rule library adopts a dynamic coefficient mapping mechanism: the weight coefficient α of the obstacle density correlation factor is adjusted positively according to the offset (Δα = 0.1×ΔW), while the weight coefficient β of the energy consumption correlation factor is adjusted inversely (Δβ = -0.05×ΔW). The adjusted scoring coefficient takes effect immediately. For example, the original priority scoring formula 0.7×ODF + 0.3×ECF becomes 0.75×ODF + 0.25×ECF after compensation, strengthening the sensitivity to the obstacle density.

[0082] Step S440: Recalculate the priority scores of the affected path nodes according to the adjusted scoring coefficient, and replace the path nodes in the initial optimized path sequence whose scores are lower than the replacement threshold.

[0083] Exemplarily, the rescoring range of the affected path nodes includes the abnormal fluctuation signal source point and all nodes within 3 hops upstream and downstream thereof. The replacement threshold is set to 70% of the original priority score. For example, if the original score of a certain node is 0.68 and it is reduced to 0.45 (lower than 0.68×0.7 = 0.476) after adjustment, then a replacement operation is triggered. The selection criteria for the replacement node include, for example: the spatial deviation distance between the alternative node and the current path is less than 2m, the priority score is higher than 120% of the adjusted score of the original node, and the increase in energy consumption does not exceed 15%. The replacement process adopts a doubly linked list operation: the low-score node is removed from the initial optimized path sequence, the alternative node is inserted, and the connection weight is updated. For example, node A with a score reduced from 0.45 to 0.32 is replaced by adjacent node B (score 0.58), and at the same time, the connection weight between node B and the predecessor node C is increased from 0.65 to 0.72 to reflect the improvement in terrain safety.

[0084] Step S450: Reorder the replaced path nodes according to the connection weight to generate a final inspection path including redundant obstacle avoidance path nodes.

[0085] Exemplarily, the introduction of redundant obstacle avoidance path nodes can be achieved through a path branching mechanism: 1-2 standby connection edges are preset at key nodes, and their priority scores are 10%-20% lower than that of the main path, but they remain in a passable state in real time. The reordering algorithm adopts an improved Dijkstra algorithm, and a redundancy factor is introduced into the path total score optimization target: ; where λ = 0.3 is the redundancy weight coefficient. For example, the main path includes the node sequence A→B→C (total score 2.15), and the redundant path A→D→C (total score 1.98) is used as an alternative. The output format of the final inspection path is a weighted directed graph, and each node records the switching condition between the main path and the redundant path (such as automatically switching when the connection weight drops by more than 25%). The path data is synchronized to all associated robots through the distributed storage protocol of the edge computing node to ensure the consistency of path planning during multi-robot collaborative operations.

[0086] As an implementation, in step S500, control the intelligent inspection robot to perform the inspection task according to the final inspection path, and continuously collect new environmental perception data streams during the inspection task to update the parameter set of the dynamic inference model, including: Step S510: When the intelligent inspection robot moves along the final inspection path, collect a new environmental perception data stream through the multi-spectral sensor carried. The new environmental perception data stream includes an infrared thermal imaging sequence, a lidar point cloud frame, and a video stream.

[0087] In this step, the motion control of the intelligent inspection robot and the model update form a closed-loop optimization system. When the robot moves along the final inspection path, the multi-spectral sensor array carried by it continuously collects environmental data at a high sampling frequency. The new environmental perception data stream includes three types of core data: infrared thermal imaging sequences, lidar point cloud frames, and video streams. The infrared thermal imaging sequence captures the temperature distribution at a frequency of 30 Hz through an uncooled microbolometer. Each pixel corresponds to an area of 0.1 m × 0.1 m, and the temperature resolution reaches 0.1 °C. The lidar point cloud frame is generated by a 16-line rotating scanner at a frequency of 20 Hz, with a point cloud density of 300 points / m² and an accuracy of ±2 cm. The video stream captures 1080P resolution images at a frame rate of 60 fps through a global shutter CMOS sensor. During the data acquisition process, the hardware synchronization module ensures that the acquisition timestamp deviation of the three types of data is less than 5 ms. For example, at the timestamp T = 2023-08-20 14:00:00.000, the infrared thermal imaging sequence records the surface temperature of a certain device as 65.3 °C, the lidar detects its contour coordinates as (X = 102.4, Y = 75.6, Z = 1.2), and the video stream captures the rotation state of the cooling fan of the device.

[0088] The deployment position of the multi-spectral sensor array is strictly matched with the robot kinematic model: the infrared thermal imager is installed on the top pan-tilt of the robot, and the vertical pitch angle adjustment range is ±30°, ensuring full coverage scanning of the device surface; the lidar is fixed on the forward bracket of the robot, with a horizontal field of view of 360° and a vertical field of view of 30°; the binocular cameras are symmetrically installed on both sides of the robot, with a baseline distance of 0.2 m and a focal length of 4 mm, supporting stereo vision ranging. The storage of the new environmental perception data stream adopts the circular buffer structure of the edge computing node. The buffer capacity is the data volume of the last 5 minutes (9000 frames of infrared thermal imaging sequence, 6000 frames of lidar point cloud frame, 18000 frames of video stream). When the buffer reaches the capacity limit, the old data is overwritten according to the first-in-first-out rule.

[0089] Step S520: Extract the temperature gradient distribution feature from the infrared thermal imaging sequence, extract the obstacle contour geometric feature from the lidar point cloud frame, extract the dynamic object motion trajectory feature from the video stream, and splice the three types of features into an incremental multi-modal environmental feature vector after aligning them according to the timestamp.

[0090] Exemplarily, the extraction of the temperature gradient distribution feature can be achieved by calculating the temperature difference matrix of adjacent pixel points. For example, for an infrared thermal imaging frame with a resolution of 80×60, a gradient matrix of 79×59 is generated, and each element value is the absolute value of the temperature difference between adjacent pixels. For example, if the temperature of a pixel point (i,j) is 52°C and the right pixel (i,j+1) is 56°C, then the value of the gradient matrix element (i,j) is 4°C. The geometric features of the obstacle contour are extracted from the lidar point cloud, and the RANSAC algorithm is used to fit the plane and calculate the vertex coordinates of the convex hull polygon. For example, the contour vertex sequence of an obstacle is {(102.4,75.6,1.2), (102.6,75.8,1.3), (102.5,75.7,1.1)}. The dynamic object motion trajectory feature is extracted from the video stream by the optical flow method. For a video frame with a resolution of 1920×1080, the Lucas-Kanade algorithm is used to calculate the displacement vector of feature points between adjacent frames. For example, the displacement vectors of the motion trajectory of a cooling fan blade in three consecutive frames are (+2px, +1px), (+3px, -1px). Timestamp alignment is achieved by an interpolation algorithm: based on the timestamp of the lidar point cloud frame, linear interpolation is performed on the infrared thermal imaging sequence and the video stream data to ensure the spatial consistency of the three types of features at the same time point. The dimension of the spliced incremental multi-modal environmental feature vector is 1024 (256 for temperature gradient + 512 for obstacle contour + 256 for motion trajectory).

[0091] Step S530: Input the incremental multi-modal environmental feature vector into the multi-layer feature fusion module of the dynamic inference model to generate an incremental spatio-temporal feature tensor, and perform a feature space similarity comparison between the incremental spatio-temporal feature tensor and the reference spatio-temporal feature tensor in the historical training dataset.

[0092] Exemplarily, the cross-modal attention allocation unit of the multi-layer feature fusion module adopts a three-head attention mechanism to allocate dynamic weights to temperature, obstacle, and motion features. For example, when a sudden increase in the temperature gradient is detected in a certain area, the attention weight of the temperature feature is increased to 0.7, and the weights of the obstacle and motion features are reduced to 0.2 and 0.1 respectively. The generation process of the incremental spatio-temporal feature tensor includes: compressing the 1024-dimensional input vector to 512 dimensions through a fully connected layer, and then extracting the temporal correlation pattern through a temporal convolutional layer, and outputting a spatio-temporal feature tensor with a dimension of T×512 (T = 10-second time window). The feature space similarity comparison is calculated based on the Wasserstein distance: ; where P is the incremental feature distribution, Q is the reference feature distribution, γ represents the joint probability distribution, Γ(P,Q) is the set of all joint distributions γ that satisfy the marginal distributions of P and Q, and x and y are sample points from distributions P and Q. When the distance value exceeds the preset similarity threshold (default 0.15), it is determined that a significant change has occurred in the current environmental feature.

[0093] Step S531: Retrieve a subset of benchmark spatio-temporal feature tensors corresponding to the current inspection path area from the historical training dataset. The subset of benchmark spatio-temporal feature tensors contains standardized environmental feature samples collected during multiple historical inspection cycles.

[0094] Exemplarily, the historical training dataset can adopt a spatio-temporal four-dimensional index structure (longitude, latitude, altitude, time), and the retrieval range is limited to the area within a radius of 5m of the current path node and the historical data of the most recent 30 days. The preprocessing of the standardized environmental feature samples includes: normalizing the temperature gradient feature to the interval of [0, 100] °C, converting the obstacle contour coordinates to a relative coordinate system (with the area center as the origin), and normalizing the movement trajectory speed to the interval of [0, 10] m / s. For example, for area R-1024, the subset of benchmark spatio-temporal feature tensors contains 120 samples, and each sample is a 512-dimensional feature vector within a 10-second time window.

[0095] Step S532: Decompose the incremental spatio-temporal feature tensor into a temperature gradient channel, an obstacle contour channel, and a movement trajectory channel according to the feature channel dimension, and calculate the distribution similarity index between each channel and the corresponding channel in the subset of benchmark spatio-temporal feature tensors within each channel.

[0096] Exemplarily, the channel decomposition is implemented through a masking operation. For example, the temperature gradient channel extracts the first 256 dimensions of the incremental spatio-temporal feature tensor, the obstacle contour channel extracts the middle 256 dimensions, and the movement trajectory channel extracts the last 256 dimensions. The distribution similarity index is calculated using the Jensen-Shannon divergence: ; where M = (P + Q) / 2, and DKL is the Kullback-Leibler divergence. For example, the similarity index of the temperature gradient channel is 0.12 (lower than the threshold of 0.18), the obstacle contour channel is 0.20, and the movement trajectory channel is 0.25.

[0097] Step S533: Apply a first weight coefficient to the distribution similarity index of the temperature gradient channel, a second weight coefficient to the distribution similarity index of the obstacle contour channel, and a third weight coefficient to the distribution similarity index of the movement trajectory channel, where the first weight coefficient is greater than the second weight coefficient and the second weight coefficient is greater than the third weight coefficient.

[0098] Exemplarily, the weight coefficients can be set according to the task safety level. For example, the first weight coefficient of the temperature gradient channel is 0.6 (because temperature anomalies may cause equipment failures), the second weight coefficient of the obstacle contour channel is 0.3, and the third weight coefficient of the movement trajectory channel is 0.1. The weighted similarity index is calculated as: S = 0.6×0.12 + 0.3×0.20 + 0.1×0.25 = 0.157.

[0099] Step S534: Normalize and sum the weighted three-channel distribution similarity metrics to generate an overall feature space similarity comparison result.

[0100] For example, the normalization process maps the weighted similarity metric to the interval [0, 1]: , where Smin = 0.1 (historical minimum similarity), Smax = 0.3 (historical maximum similarity). If the current S = 0.157, the normalization result is (0.157 - 0.1) / (0.3 - 0.1) = 0.285.

[0101] Step S535: When the overall feature space similarity comparison result is lower than the preset similarity threshold, mark the current incremental spatio-temporal feature tensor as an environmental feature mutation sample and trigger the priority scoring rule iteration process of the dynamic inference model.

[0102] Exemplarily, the preset similarity threshold is 0.15 (configurable parameter). When Snorm = 0.285 > 0.15, the iteration is not triggered; if Snorm = 0.12, it is marked as a mutation sample. The metadata record of the mutation sample includes timestamp, geographical location, sensor readings, and the original feature vector, and the storage format is Protocol Buffers binary encoding, with a single sample size of approximately 2KB. After the priority scoring rule iteration process is started, the model enters the online learning mode and pauses the real-time path planning task to ensure data consistency.

[0103] Step S536: Add the environmental feature mutation sample to the abnormal sample library of the historical training dataset, and recalculate the cluster center coordinates of the benchmark spatio-temporal feature tensor subset to expand the feature space coverage.

[0104] Exemplarily, the abnormal sample library can use the Faiss index to accelerate the retrieval of similar samples, and the new samples update the cluster center of the benchmark feature through the K-means++ algorithm. For example, there are 3 cluster centers in the original benchmark feature subset, and after adding the new mutation sample, it is recalculated to 4 centers, where the fourth center corresponds to the temperature anomaly pattern. After the feature space coverage is expanded, the variance of the benchmark feature distribution increases by 15%, improving the model's adaptability to rare events.

[0105] Step S537: Adjust the decision boundary of the preset similarity threshold according to the expanded feature space coverage.

[0106] Exemplarily, the similarity threshold is adjusted to a dynamic parameter. For example, , for example, when 500 samples are newly added to the abnormal sample library, the threshold is increased to 0.15 × 1.5 = 0.225, reducing the false alarm rate. The threshold update period is once every 24 hours, synchronized with the model version release period.

[0107] Step S540: When the feature space similarity comparison result is lower than the preset similarity threshold, activate the parameter update engine, and separate abnormal feature segments that are significantly different from the current environmental feature topology map from the incremental spatio-temporal feature tensor.

[0108] Exemplarily, the parameter update engine can adopt a sliding window segmentation algorithm to extract continuous time segments from the incremental spatio-temporal feature tensor. For example, within a 10-second time window, if the feature similarity from the 3rd to the 5th second continuously falls below the threshold, then separate this 3-second segment as an abnormal feature segment. The dimension of the abnormal segment is 3 × 512 (3 seconds × 512 dimensions / second), and it is labeled as the "abnormal high-temperature equipment" category.

[0109] Step S550: Inject the abnormal feature segment into the path decision module of the dynamic inference model for forward propagation calculation, and obtain the prediction path node priority score error at the output layer of the path decision module.

[0110] Exemplarily, forward propagation calculation can use the Monte Carlo Dropout method (Dropout rate = 0.2) to evaluate the model uncertainty. For the abnormal feature segment, the average priority score output by the path decision module is 0.35, while the actual traffic efficiency evaluation value is 0.18, resulting in an error Δ = 0.17. The error calculation uses the absolute percentage error (APE): .

[0111] Step S560: According to the prediction path node priority score error, reversely adjust the cross-modal attention allocation weights of the multi-layer feature fusion module, and synchronously correct the convolution kernel parameters related to the energy consumption correlation factor in the path decision module.

[0112] Exemplarily, backpropagation can adopt the elastic weight consolidation (EWC) algorithm to constrain the adjustment amplitude of important parameters (identified by the Fisher information matrix). The update formula for the cross-modal attention weight is: ; where η = 0.001 is the learning rate, λ = 0.5 is the elastic coefficient, and F is the Fisher information diagonal matrix. The correction amount of the convolution kernel parameters of the energy consumption correlation factor is calculated by gradient descent. For example, the weight matrix of the 3 × 3 convolution kernel is adjusted from [0.2, -0.1, 0.3] to [0.18, -0.08, 0.28].

[0113] Step S570: Combine the adjusted cross-modal attention allocation weights and the corrected convolutional kernel parameters into an updated parameter set, and synchronize the updated parameter set to the local decision controller of the intelligent inspection robot through the distributed communication protocol of the edge computing path node.

[0114] Exemplarily, the parameter set can adopt differential coding compression technology to only transmit the change amount to reduce bandwidth occupancy. For example, the original parameter file size is 10MB, and the differential update package is only 120KB. The distributed communication protocol is implemented based on MQTT, the QoS level is set to 1 (at least once delivery), and the retransmission timeout is 200ms. The synchronization process adopts a two-phase commit protocol to ensure the parameter consistency of multiple robot nodes.

[0115] Step S580: Load the updated parameter set through the local decision controller to overwrite the original parameters, so that the subsequent path optimization task can dynamically generate a priority scoring rule based on the latest environmental features.

[0116] Exemplarily, the parameter loading process includes integrity verification (SHA-256 hash verification) and a rollback mechanism: if the new parameters cause abnormal model inference (such as outputting NaN values), automatically switch to the previous stable version. When the average scoring error of 10 consecutive path planning tasks of the updated model during verification inspections in a restricted area is less than 5%, it is marked as a stable version. The dynamic generation period of the priority scoring rule is shortened to 30% of the original system, realizing near-real-time environmental adaptability.

[0117] As an implementation, after step S500, controlling the intelligent inspection robot to perform the inspection task according to the final inspection path and continuously collecting new environmental perception data streams during the inspection task to update the parameter set of the dynamic inference model, the method further includes: Step S600: Input the updated parameter set of the dynamic inference model into the offline verification unit to generate a simulated inspection path sequence matching the historical environmental feature topology map Exemplarily, the offline verification unit is a virtual simulation environment constructed based on the historical environmental feature topology map. Its core function is to reproduce the operating conditions of the real scenario by loading historical data. The updated dynamic inference model parameter set includes the adjusted cross-modal attention allocation weights, the corrected convolutional kernel parameters, and the latest version of the priority scoring rule library. The generation process of the simulated patrol path sequence is as follows: First, extract the environmental feature topology map (including path node coordinates, dynamic connection weights, and historical perception data) corresponding to the current task area from the historical version library, and then load the updated parameter set into the dynamic inference model to perform the full-area path planning task in the simulation environment. For example, for the historical topology map of a chemical plant (including 120 path nodes), the offline verification unit generates 10 simulated path sequences, each path covering all key equipment detection points, and introduces random noise (such as a ±5% connection weight perturbation) through Monte Carlo simulation to test the model robustness. The generated simulated path sequences are stored with double indexing of timestamp and geographical coordinates, supporting the comparison and analysis with the historical actual paths.

[0118] Step S700: Calculate the path optimization confidence of the updated parameter set by comparing the actual execution efficiency differences between the simulated patrol path sequence and the final patrol path. Exemplarily, the actual execution efficiency differences are quantified from three dimensions: time efficiency, energy consumption, and safety. The time efficiency difference is the ratio of the total execution time of the simulated path to the final patrol path. The energy consumption difference is calculated through the integral of the motor torque and the battery discharge curve. The safety difference is evaluated based on the standard deviation of the path node priority scores. The calculation formula for the path optimization confidence can be, for example: ; where D1, D2, and D3 represent the percentage differences in time, energy, and safety respectively, and the weight coefficients are w1 = 0.5, w2 = 0.3, and w3 = 0.2. For example, if the time efficiency of the simulated path is 8% higher than the actual path (D1 = -0.08), but the energy consumption increases by 12% (D2 = 0.12), and the safety score drops by 5% (D3 = 0.05), then the confidence is 1 - (0.5×0.08 + 0.3×0.12 + 0.2×0.05) = 0.892. The preset confidence threshold is 0.9. When the calculation result is lower than this value, it is determined that there is a potential risk in the parameter update.

[0119] Step S800: When the path optimization confidence is lower than the preset confidence threshold, activate the rollback mechanism and extract the parameter set of the previous stable version of the dynamic inference model from the historical version library of the edge computing node. Exemplarily, the rollback mechanism can be implemented through a version control system. The historical version library stores the stable parameter sets of each time indexed by timestamps and hash values. For example, when the current updated version is V2.3 and the confidence level is 0.892 (lower than 0.9), the system retrieves the hash value (such as a1b2c3d4) of the most recent stable version V2.2 and downloads the corresponding parameter file from the distributed storage nodes. During the rollback process, the edge computing nodes pause the processing of new tasks, clear the current model cache, and load the old version parameter set. The version compatibility verification module ensures that the parameter set matches the topological structure of the current environmental feature topology map. If changes in the number of nodes or connection relationships are detected (such as the deletion of path nodes due to newly added obstacles), the topology adaptation algorithm is triggered to adjust the parameter dimensions through interpolation or pruning operations.

[0120] Step S900: Perform feature-level difference analysis on the parameter set of the previous stable version and the updated parameter set to locate the abnormal weight distribution area that causes the confidence level to decrease. Exemplarily, the feature-level difference analysis can adopt gradient saliency mapping technology to calculate the output gradient difference of the parameters of the two versions under the same input data. The specific method is, for example: select the key path nodes in the historical environmental feature topology map (such as the nodes near high-risk devices), input the corresponding multi-modal environmental feature vectors into the old and new models, and calculate the output layer gradients of the path decision module respectively. By comparing the gradient amplitude distributions, locate the network layer where the parameter changes have the greatest impact on the priority score. For example, the updated convolutional kernel generates an abnormally high gradient (150% increase compared to the old version) in the third layer, indicating that the parameter adjustment in this layer causes a deviation in the calculation of the energy consumption correlation factor. Further, the weight distribution difference is displayed through feature visualization technology (such as t-SNE dimensionality reduction) to identify the abnormal areas in the parameter space that deviate from the cluster center.

[0121] Step S1000: Generate a parameter correction mask based on the abnormal weight distribution area and apply the parameter correction mask to the updated parameter set to filter out the abnormal weight values. Exemplarily, the parameter correction mask is, for example, a binary matrix whose dimension is exactly the same as the dynamic inference model parameter set. The value of the matrix element being 1 indicates that the parameter is allowed to be retained, and the value being 0 indicates that it needs to be replaced with the old version value. The mask generation rules include, for example: 1) Set 0 for the parameter positions where the gradient difference exceeds the threshold (such as 50%); 2) Set 0 for the parameters that deviate more than 2σ from the old version cluster center in feature visualization. For example, in the weight matrix of a certain convolutional kernel, the gradient differences of 3 weight values are 62%, 45%, and 78% respectively, then the corresponding positions in the mask are set to 0, 1, and 0. The correction operation is performed bit by bit: ; where Denotes element-wise multiplication. This process can eliminate the influence of abnormal weights while preserving effective parameter updates.

[0122] Step S1100: Reload the corrected parameter set into the dynamic inference model and initiate a verification patrol task within a restricted area to test the stability of path generation Exemplarily, the restricted area is defined as an area with a high incidence of historical failures or an area significantly affected by parameter anomalies. For example, 5 path nodes in a certain storage tank area once failed to avoid obstacles due to over-adjustment of parameters. The verification patrol task performs full-coverage detection in this area, and the task indicators include: 1) The path node coverage rate needs to reach 100%; 2) The fluctuation range of the priority score is less than ±10%; 3) The real-time path replanning response time is less than 2 seconds. During the test process, the system records the consistency index (such as path coincidence degree > 85%) between the path sequence and the old version and the execution efficiency data. If a sudden change in the score is detected (such as the score of a certain node drops from 0.6 to 0.3 suddenly), the task is immediately interrupted and an exception log is recorded.

[0123] Step S1200: When the path execution efficiency of the verification patrol task reaches the preset efficiency standard, mark the corrected parameter set as the stable version and synchronize it to all associated edge computing nodes Exemplarily, the preset efficiency standard includes, for example: 1) The time efficiency difference ≤ 5%; 2) The energy consumption difference ≤ 8%; 3) The safety score difference ≤ 3%. When the test results meet the above conditions, the system generates a globally unique version identifier (such as V2.3.1) for the parameter set, and through a distributed synchronization algorithm based on the Paxos protocol, shards and transmits the parameter file to all associated edge computing nodes. The synchronization process includes data integrity verification (CRC32 checksum comparison) and conflict resolution mechanisms (such as the version with the latest timestamp takes precedence). After synchronization is completed, each node updates its local model loading list and sets the new version as the default parameter set.

[0124] Step S1300: Update the backup record of the historical version library according to the synchronized stable version parameter set and set parameter update constraint conditions for the next incremental training process Exemplarily, the historical version library can adopt a cold and hot data hierarchical storage strategy: the latest 3 versions are retained in the SSD high-speed storage layer, and the early versions are archived to a distributed file system (such as HDFS). The backup record contains the metadata of the parameter set (version number, hash value, update time) and performance metrics (confidence level, test pass rate). The parameter update constraint conditions are implemented through regularization terms, including: 1) restricting the adjustment range of the convolutional kernel weights not to exceed ±20% of the old version; 2) forcing the temperature channel coefficient of the cross-modal attention weights to be not less than 0.5; 3) prohibiting the modification of historical highly sensitive parameters (such as the weight coefficient α of the priority scoring formula). These constraint conditions are encoded as additional terms of the incremental training loss function to ensure that subsequent updates are carried out within a controllable range.

[0125] As an implementation manner, the method provided by the embodiment of the present invention may further include the following steps: Step S1400: During the process of the intelligent inspection robot performing the inspection task, the energy reserve status and device health indicators are monitored in real time. When the energy reserve status is lower than the preset energy reserve status threshold or the device health indicator exceeds the preset device health indicator threshold, an emergency path replanning signal is triggered.

[0126] In this step, the energy reserve status is collected in real time through the battery management system of the intelligent inspection robot, and the current remaining power is represented as a percentage, and the estimated remaining battery life based on the load power is calculated. The device health indicators include parameters such as motor temperature, sensor accuracy, and communication module bit error rate. Among them, the motor temperature is sampled at a frequency of 1Hz through a thermocouple sensor, the sensor accuracy is periodically detected through a self-calibration module, and the communication bit error rate is calculated from the CRC check failure rate of the wireless module. The preset energy reserve status threshold is dynamically adjusted according to the current task type of the robot: it is set to 20% in the regular inspection mode and 30% in the high-speed movement mode; among the preset device health indicator thresholds, the motor temperature threshold is set to 75°C, the sensor accuracy deviation threshold is set to ±5%, and the communication bit error rate threshold is set to 1e-4. When any parameter exceeds the threshold, for example, the battery power drops to 18% or the motor temperature reaches 78°C, the system immediately generates an emergency path replanning signal, which contains a fault type code (such as E01 indicating low power and E02 indicating motor overheating) and the current geographical location coordinates (such as the longitude and latitude values in the WGS84 coordinate system). The transmission of the emergency signal adopts the highest priority interruption mechanism to ensure that it is delivered to the path planning control core within 5ms.

[0127] Step S1500: Interrupt the data collection process of the current inspection task according to the emergency path replanning signal and activate the fast response sub-module of the dynamic inference model.

[0128] Exemplarily, the interruption of the data acquisition process can be achieved through hardware-level signal shielding. For example, the power supply circuit of the multispectral sensor array is turned off, the lidar motor rotation is paused, and the camera lens is switched to a mechanically locked state to prevent data contamination. The fast response sub-module is a lightweight version of the dynamic inference model. The number of neural network layers is reduced to 30% of the original model, and 8-bit integer quantization compression technology is adopted, which increases the inference speed to 3 times that of the original system. The activation process includes loading a pre-compiled emergency inference kernel (about 15MB) into the FPGA accelerator and allocating an independent memory space (256MB) to prevent resource contention. For example, when an E01-class signal is received, the sub-module immediately releases the computing resources occupied by the current path planning thread and completes the initialization within 50ms, ready to receive the subset data of the environmental feature topology map.

[0129] Step S1600: Extract a subset of the environmental feature topology map of the current robot position through the fast response sub-module. The subset of the environmental feature topology map includes all path nodes within a preset radius centered on the current position and their dynamic connection weights.

[0130] Exemplarily, the preset radius can be dynamically calculated according to the remaining energy reserve status. For example, when the remaining battery power is 15%, the radius is set to the product of the current moving speed of the robot (0.5m / s) and the estimated endurance time (18 minutes), that is, 0.5×1080 = 540 meters; when the device health indicator is abnormal (such as too high motor temperature), the radius is reduced to a safe moving distance (such as 200 meters). The extraction of the subset of the environmental feature topology map is implemented through a spatial range filter: centered on the current position coordinates of the robot (such as X = 1024.56, Y = 768.32), a three-dimensional sphere query condition (radius R = 540 meters) is constructed, and 87 path nodes are retrieved from the octree index structure of the global topology map. Each node contains the latest value of the dynamically updated connection weight (such as the weight from node A to node B is reduced from 0.75 to 0.68) and the timestamp of the associated sensor data (ensuring that the data timeliness is within 30 seconds).

[0131] Step S1700: Topologically match the subset of the environmental feature topology map with the pre-stored location information of the charging station or maintenance point to generate a node connection sequence of multiple candidate emergency paths.

[0132] Exemplarily, the pre-stored location information of the charging station can be stored in GeoJSON format, including the coordinates of the charging pile, the interface type (such as CCS / CHAdeMO), and the real-time available status. The topological matching algorithm adopts an improved A* algorithm. For example, the heuristic function can be designed as: ; Where d(n) is the Euclidean distance from the current node to the target, w(n) is the mean of the dynamic connection weights, and α = 0.7, β = 0.3 are the balance coefficients. For example, during the matching process, 3 candidate paths are generated: Path 1 contains the node sequence A→C→D→charging station (total weight 2.15), Path 2 is A→E→F→charging station (total weight 1.98), and Path 3 is A→G→charging station (total weight 2.34). The node connection sequence of each path is recorded as a doubly linked list structure, supporting fast insertion and deletion operations.

[0133] Step S1800: The path decision module based on the dynamic inference model performs priority scoring on the node connection sequences of each candidate emergency path, calculates the comprehensive path risk coefficient and the estimated energy consumption.

[0134] Exemplarily, the comprehensive path risk coefficient is calculated by weighting the obstacle density correlation factor, the connection weight volatility, and the historical failure rate: ; where σw is the standard deviation of the path connection weights, and Hf is the number of failures of the path nodes in the past 24 hours. The estimated energy consumption is calculated through the motor torque integration model: ; where θi is the slope angle between nodes, vi is the preset moving speed (0.5 m / s), and k1 = 0.8, k2 = 0.2 are calibration coefficients. For example, for Path 1, the risk coefficient R = 0.62 and the estimated energy consumption E = 8500 J; for Path 2, R = 0.55 and E = 8200 J.

[0135] Step S1900: Select the candidate emergency path with the lowest comprehensive path risk coefficient and the smallest estimated energy consumption as the target emergency path, and adjust the dynamic connection weights of the nodes in the target emergency path to reduce the execution frequency of unnecessary detection tasks.

[0136] Exemplarily, the selection strategy can adopt the Pareto optimal criterion. For example, when both the R and E of Path 2 are better than those of other paths, it is directly selected; if there is a conflict (such as Path 1 has a lower R but a higher E), it is selected according to the principle of giving priority to the risk coefficient. The adjustment of the dynamic connection weights is for task execution nodes (such as equipment detection points or data collection points), and their weights are reduced by a factor of 0.7. For example, the original weight of a certain detection node is 0.65, and after adjustment, it is 0.45, reducing the selection priority of this node in path optimization by 31%.

[0137] As an implementation, in step S1900, adjusting the dynamic connection weights of the nodes in the target emergency path to reduce the execution frequency of unnecessary detection tasks includes: Step S1910: Identify task execution nodes related to device status monitoring or environmental data collection from the node connection sequence of the target emergency path.

[0138] Exemplarily, the identification of task execution nodes can be based on a preset POI (Point of Interest) tag system. For example, in the environmental feature topology graph, the device detection node is marked as Type = 1, and the data collection node is marked as Type = 2. By traversing the path node linked list, extract all nodes where Type ∈ {1, 2}. For example, node E (Type = 1) and node F (Type = 2) in path 2.

[0139] Step S1920: According to the priority level of the emergency path replanning signal, proportionally reduce the dynamic connection weight of the task execution nodes, so that the dynamic inference model reduces the selection probability of the task execution nodes during the path optimization process.

[0140] Exemplarily, the priority level can be divided into three levels: Level 1 (severe failure) with a reduction coefficient of 0.5, Level 2 (moderate failure) of 0.7, and Level 3 (minor failure) of 0.9. For example, for a Level 1 signal, the weight of node E is reduced from 0.68 to 0.34, and the weight of node F is reduced from 0.72 to 0.36. The reduced weights are immediately updated to the in-memory copy of the environmental feature topology graph subset.

[0141] Step S1930: Inject the reduced dynamic connection weights into the environmental feature topology graph subset, and recalculate the node priority scores through the feedback calibration mechanism of the dynamic inference model.

[0142] Exemplarily, the weight injection operation can adopt an atomic write method to ensure data consistency in a multi-threaded environment. After the feedback calibration mechanism is activated, the path decision module recalculates the priority scores of the affected nodes within 50 ms. For example, the score of node E is reduced from 0.75 to 0.48, and the score of node F is reduced from 0.82 to 0.53, resulting in the total score of path 2 being reduced from 1.98 to 1.65.

[0143] Step S1940: Generate a streamlined emergency path by removing unnecessary task nodes based on the recalculated node priority scores, and add redundant obstacle avoidance connection edges to the key navigation nodes in the streamlined emergency path.

[0144] Exemplarily, the strategy for adding redundant obstacle avoidance connection edges is as follows: retrieve backup path nodes within 3 meters around key nodes (such as turning points or narrow passage nodes), and establish connection edges with a 20% increased weight. For example, the original connection edge weight from node G to the charging station is 0.75, and a new redundant edge G→H→charging station is added with a weight set to 0.90. The streamlined path node sequence changes from A→E→F→charging station to A→H→charging station, and the path length is shortened by 18%.

[0145] Step S1950: Continuously monitor the remaining energy reserve status and the changing trend of device health indicators during the movement of the intelligent inspection robot, and dynamically adjust the weight allocation ratio of redundant obstacle avoidance connection edges.

[0146] Exemplarily, the dynamic adjustment algorithm can adopt proportional-integral (PI) control. For example, when the remaining battery power drops by 1%, the weight of the redundant edge is increased by 0.5%; when the motor temperature exceeds 70°C, the weight is increased by 1%. For example, during movement, the battery power drops from 15% to 12%, and the weight of the redundant edge of node H increases from 0.90 to 0.93; if the motor temperature rises to 72°C, it further increases to 0.95.

[0147] Step S1960: When the intelligent inspection robot arrives at the charging station or maintenance point, restore the dynamic connection weight configuration of the original task execution nodes, and upload the emergency path execution log to the distributed storage system of the edge computing node.

[0148] Exemplarily, the weight restoration operation can be achieved through a rollback transaction. For example, extract the latest weight value before the failure from the version library to overwrite the temporarily adjusted value. The emergency log includes path trajectories, energy consumption data, and fault codes, which are compressed and stored in the Avro binary format (about 120KB per single log), and encrypted and transmitted to the edge node through the HTTPS protocol. The log metadata records the timestamp, robot ID, and processing duration (such as this emergency task taking 8 minutes and 32 seconds).

[0149] Step S2000: Control the intelligent inspection robot to move along the target emergency path to the charging station or maintenance point, and reload the latest environmental feature topology map after completing energy replenishment or equipment maintenance to resume the original inspection task process.

[0150] Exemplarily, the movement control can adopt a PID trajectory tracking algorithm, with the lateral deviation controlled within ±0.1 m and the heading angle deviation less than 2 degrees. After charging is completed or the device is repaired, the system downloads the updated environmental feature topology map (such as newly added obstacle nodes or adjusted connection weights) from the edge node and verifies the data integrity through the checksum. When resuming the inspection tour, the initial path planning is regenerated based on the latest topology map to ensure full synchronization with the current environmental state. For example, if a newly added obstacle makes the original path node C inaccessible, the system automatically plans a detour path C’→D’→E, and the priority score is updated to 0.78.

[0151] An embodiment of the present invention provides a computer system, such as Figure 2 As shown, the computer system 100 includes: a processor 101 and a memory 103. Among them, the processor 101 and the memory 103 are connected, such as connected through a bus 102. Optionally, the computer system 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation to the embodiments of the present invention.

[0152] An embodiment of the present invention provides a computer system. The computer system in the embodiment of the present invention includes: one or more processors; a memory; one or more computer programs, where one or more computer programs are stored in the memory and are configured to be executed by one or more processors. When the one or more programs are executed by the processor, the above-mentioned intelligent inspection robot path optimization method based on the edge inference model provided by the embodiment of the present invention is implemented.

[0153] An embodiment of the present invention provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program runs on the processor, the processor can execute the corresponding content in the foregoing method embodiments.

[0154] It should be understood that although the steps in the flowchart of the accompanying drawings are shown in sequence according to the arrows, these steps do not necessarily have to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps does not have a strict order limit, and they can be executed in other orders. Moreover, at least some of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily have to be executed at the same moment, but can be executed at different moments. Their execution order does not necessarily have to be sequential, but can be executed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

[0155] The above are only some embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. An intelligent inspection robot path optimization method based on an edge inference model, characterized in that The method includes: Obtaining an environmental feature topology map of a target inspection area and determining an initial edge inference model; the environmental feature topology map includes multiple path nodes and dynamic connection weights between the path nodes, and each path node is associated with a real-time environmental perception data stream; the initial edge inference model includes a multi-layer feature fusion module and a path decision module; Performing incremental training on the initial edge inference model through the real-time environmental perception data stream to generate a dynamic inference model adapted to the current environmental features; the incremental training includes adjusting the parameters of the multi-layer feature fusion module according to the spatio-temporal distribution differences of the environmental perception data stream; Based on the dynamic inference model, performing priority scoring on each path node in the environmental feature topology map to generate an initial optimized path sequence; the priority scoring includes an obstacle density correlation factor and an energy consumption correlation factor between path nodes; Triggering a feedback calibration mechanism of the dynamic inference model according to the real-time updated environmental perception data stream, performing dynamic path node replacement on the initial optimized path sequence to generate a final inspection path; the feedback calibration mechanism includes performing backpropagation compensation on the connection weights of the path nodes; Controlling the intelligent inspection robot to perform an inspection task according to the final inspection path, and continuously collecting new environmental perception data streams during the inspection task to update the parameter set of the dynamic inference model.

2. The method according to claim 1, wherein The performing incremental training on the initial edge inference model through the real-time environmental perception data stream to generate a dynamic inference model adapted to the current environmental features includes: Extracting a multi-modal environmental feature vector from the real-time environmental perception data stream, where the multi-modal environmental feature vector includes a thermal radiation distribution feature, a voiceprint fluctuation feature, and a visual texture feature; Inputting the multi-modal environmental feature vector into the multi-layer feature fusion module to generate a fused spatio-temporal feature tensor; the multi-layer feature fusion module includes a cross-modal attention allocation unit and a feature channel alignment unit; According to the dimensional distribution differences of the spatio-temporal feature tensor, calculating the loss function gradient of the initial edge inference model, and iteratively optimizing the decision threshold of the path decision module based on the gradient direction; When the change rate of the loss function gradient is lower than a preset convergence threshold, freezing the parameters of the multi-layer feature fusion module, and combining the optimized path decision module with the multi-layer feature fusion module into a dynamic inference model; Establishing a mapping relationship between the output layer of the dynamic inference model and the path nodes of the environmental feature topology map to generate a path node priority scoring rule library.

3. The method according to claim 2, wherein The performing priority scoring on each path node in the environmental feature topology map based on the dynamic inference model to generate an initial optimized path sequence includes: Traversing all path nodes in the environmental feature topology map to obtain the multi-modal environmental feature vector corresponding to the real-time environmental perception data stream of each path node; Inputting the multi-modal environmental feature vector into the multi-layer feature fusion module of the dynamic inference model to output the spatio-temporal feature tensor of each path node; Calculate the obstacle density correlation factor for each path node according to the energy distribution spectrum of the spatio-temporal feature tensor; wherein, the obstacle density correlation factor is positively correlated with the amplitude of the high-frequency component of the energy distribution spectrum; Perform convolution processing on the spatio-temporal feature tensor of each path node through the path decision module to generate an energy consumption correlation factor; wherein, the energy consumption correlation factor is negatively correlated with the activation degree of the feature channels after convolution processing; Generate the priority score of each path node by synthesizing the weighted sum of the obstacle density correlation factor and the energy consumption correlation factor, and generate an initial optimized path sequence in descending order of the priority score.

4. The method according to claim 3, characterized in that, The feedback calibration mechanism of the dynamic inference model is triggered according to the real-time updated environmental perception data stream, and dynamic path node replacement is performed on the initial optimized path sequence, including: Monitor the abnormal fluctuation signal in the real-time environmental perception data stream, and activate the feedback calibration mechanism when the fluctuation amplitude exceeds the preset safety threshold; Extract the environmental feature subgraph corresponding to the abnormal fluctuation signal, and input the environmental feature subgraph into the backpropagation compensation unit of the dynamic inference model; Calculate the weight offset of the path decision module in the dynamic inference model through the backpropagation compensation unit, and adjust the scoring coefficient of the priority scoring rule library according to the weight offset; Recalculate the priority score of the affected path nodes according to the adjusted scoring coefficient, and replace the path nodes in the initial optimized path sequence with scores lower than the replacement threshold; Reorder the replaced path nodes according to the connection weights to generate a final inspection path including redundant obstacle avoidance path nodes.

5. The method according to claim 4, wherein Control the intelligent inspection robot to execute the inspection task according to the final inspection path, and continuously collect new environmental perception data streams during the inspection task to update the parameter set of the dynamic inference model, including: When the intelligent inspection robot moves along the final inspection path, collect a new environmental perception data stream through the equipped multi-spectral sensor, and the new environmental perception data stream includes an infrared thermal imaging sequence, a lidar point cloud frame, and a video stream; Extract the temperature gradient distribution feature from the infrared thermal imaging sequence, extract the obstacle contour geometric feature from the lidar point cloud frame, extract the dynamic object motion trajectory feature from the video stream, and splice the three types of features into an incremental multi-modal environmental feature vector after aligning them by timestamp; Input the incremental multi-modal environmental feature vector into the multi-layer feature fusion module of the dynamic inference model to generate an incremental spatio-temporal feature tensor, and perform feature space similarity comparison between the incremental spatio-temporal feature tensor and the benchmark spatio-temporal feature tensor in the historical training dataset; When the feature space similarity comparison result is lower than the preset similarity threshold, activate the parameter update engine, and separate the abnormal feature segment that is significantly different from the current environmental feature topology map from the incremental spatio-temporal feature tensor; Inject the abnormal feature segment into the path decision module of the dynamic inference model for forward propagation calculation to obtain the prediction path node priority score error of the path decision module at the output layer; Backwardly adjust the cross-modal attention allocation weights of the multi-layer feature fusion module according to the prediction path node priority scoring error, and synchronously correct the convolution kernel parameters related to the energy consumption correlation factor in the path decision module; Combine the adjusted cross-modal attention allocation weights and the corrected convolution kernel parameters into an updated parameter set, and synchronize the updated parameter set to the local decision controller of the intelligent inspection robot through the distributed communication protocol of the edge computing path node; Load the updated parameter set by the local decision controller to overwrite the original parameters, so that subsequent path optimization tasks can dynamically generate priority scoring rules based on the latest environmental features.

6. The method according to claim 5, wherein The comparison of the feature space similarity between the incremental spatio-temporal feature tensor and the benchmark spatio-temporal feature tensor in the historical training dataset includes: Retrieve a subset of the benchmark spatio-temporal feature tensor corresponding to the current inspection path area from the historical training dataset, and the benchmark spatio-temporal feature tensor subset contains standardized environmental feature samples collected during multiple historical inspection cycles; Decompose the incremental spatio-temporal feature tensor into a temperature gradient channel, an obstacle contour channel, and a motion trajectory channel according to the feature channel dimension, and calculate the distribution similarity index of each channel with the corresponding channel in the benchmark spatio-temporal feature tensor subset; Apply a first weight coefficient to the distribution similarity index of the temperature gradient channel, a second weight coefficient to the distribution similarity index of the obstacle contour channel, and a third weight coefficient to the distribution similarity index of the motion trajectory channel, where the first weight coefficient is greater than the second weight coefficient and the second weight coefficient is greater than the third weight coefficient; Normalize and sum the weighted distribution similarity indexes of the three channels to generate an overall feature space similarity comparison result; When the overall feature space similarity comparison result is lower than the preset similarity threshold, mark the current incremental spatio-temporal feature tensor as an environmental feature mutation sample, and trigger the priority scoring rule iteration process of the dynamic inference model; Add the environmental feature mutation sample to the abnormal sample library of the historical training dataset, and recalculate the clustering center coordinates of the benchmark spatio-temporal feature tensor subset to expand the feature space coverage range; Adjust the decision boundary of the preset similarity threshold according to the expanded feature space coverage range.

7. The method according to claim 1, characterized in that The obtaining of the environmental feature topology map of the target inspection area and the determination of the initial edge inference model include: Collect multi-angle environmental data through a 3D laser scanner carried by a drone deployed in the target inspection area to generate environmental point cloud data covering the entire area; Perform spatial grid division on the environmental point cloud data, and count the obstacle height distribution characteristics and ground reflection intensity characteristics in each grid unit to generate a grid environmental attribute matrix; Identify the passable area boundary based on the grid environmental attribute matrix, extract the position coordinates of the key path nodes and mark the physical connection relationship between the path nodes; Calculate the initial dynamic connection weight according to the ground slope change rate and the obstacle projection overlapping area between adjacent path nodes, where the initial dynamic connection weight is negatively correlated with the ground slope change rate and negatively correlated with the obstacle projection overlapping area; Integrate the position coordinates, physical connection relationship, and initial dynamic connection weight of the key path nodes into the path node connection map of the environmental feature topology map; Download the pre-trained general edge inference model from the cloud model library as the initial edge inference model, and encode the topological structure of the path node connection map into the input feature dimension of the initial edge inference model; Map the path node position coordinates to a high-dimensional space vector through the embedding layer of the initial edge inference model.

8. The method according to claim 7, wherein The method for identifying the passable area boundary based on the grid environmental attribute matrix, extracting the position coordinates of the key path nodes, and marking the physical connection relationship between the path nodes includes: Perform binary processing on the grid environmental attribute matrix, mark the grids with obstacle height lower than the robot chassis height as passable units, and the rest as obstacle units; Detect the continuous area contour in the passable units, and extract the center point of each contour as the initial position of the candidate path node; Perform density clustering analysis on the initial positions of the candidate path nodes, and merge the adjacent path nodes with a distance less than the minimum turning radius of the robot to form a set of backbone path nodes; Select the path node with the maximum distance from the boundary of the obstacle unit in the set of backbone path nodes as the safe path node, and record the straight-line visible distance between each safe path node; Generate the physical connection relationship between the path nodes according to the straight-line visible distance, and establish a bidirectional connection edge when there are continuous passable units between two path nodes and they are not blocked by obstacle units; Assign an initial dynamic connection weight to each bidirectional connection edge, and the initial dynamic connection weight is dynamically adjusted according to the average ground reflection intensity of the grid cells passed by the connection edge; Combine the set of safe path nodes, physical connection relationship, and initial dynamic connection weight into the initial version of the environmental feature topology map.

9. The method according to claim 1, wherein After controlling the intelligent inspection robot to perform the inspection task according to the final inspection path and continuously collecting new environmental perception data streams during the inspection task to update the parameter set of the dynamic inference model, the method further includes: Input the updated parameter set of the dynamic inference model into the offline verification unit to generate a simulated inspection path sequence matching the historical environmental feature topology map; Calculate the path optimization confidence of the updated parameter set by comparing the actual execution efficiency difference between the simulated inspection path sequence and the final inspection path; When the path optimization confidence is lower than the preset confidence threshold, activate the rollback mechanism and extract the parameter set of the previous stable version of the dynamic inference model from the historical version library of the edge computing node; Perform feature-level difference analysis on the parameter set of the previous stable version and the updated parameter set to locate the abnormal weight distribution area that causes the confidence to decrease; Generate a parameter correction mask based on the abnormal weight distribution region, and apply the parameter correction mask to the updated parameter set to filter out abnormal weight values; Reload the corrected parameter set into the dynamic inference model, and start a verification patrol task within a restricted area to test the stability of path generation; When the path execution efficiency of the verification patrol task reaches the preset efficiency standard, mark the corrected parameter set as the stable version and synchronize it to all associated edge computing nodes; Update the backup record of the historical version library according to the synchronized stable version parameter set, and set parameter update constraint conditions for the next incremental training process.

10. A computer system, characterized in that, including: One or more processors; A memory; One or more computer programs; Wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processor, the method according to any one of claims 1 to 9 is implemented.

Images