Method and system for dynamically optimizing path of intelligent inspection robot

Abstract

The invention relates to the technical field of artificial intelligence and robots, discloses an intelligent inspection robot path dynamic optimization method and system, and aims to solve the problems of path interruption, low efficiency and resource imbalance caused by dependence on a static map, response lag, frequent replanning and ignoring of task semantic information in the prior art. According to the method, a multi-mode sensor is used for collecting environment data in real time and constructing a local occupation grid map, an initial path is generated in combination with a global topological map and an improved RRT * algorithm, and a path deviation index PII and an obstacle approximation rate OAR are calculated in real time. Through the technologies, low-delay and high-robustness path optimization in a dynamic environment is realized, the response speed is remarkably improved, the number of replanning times is reduced by 65%, the inspection path is shortened by 12.3%, the steering action is reduced by 21.5%, and the running stability and safety are ensured. The improvement strategies are consistent with research results of path planning and optimization of the intelligent inspection robot, and the method has a wide prospect in engineering application.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and robotics technology, specifically relating to a method and system for dynamic path optimization of an intelligent inspection robot. Background Technology

[0002] With the continuous improvement of intelligent manufacturing and industrial automation, intelligent robots are being used more and more widely in complex industrial environments. Especially in key scenarios such as equipment inspection, safety monitoring, and fault early warning, intelligent robots are playing a crucial role. For example, intelligent inspection robots can be equipped with high-precision sensors and cameras to perform autonomous inspections and tests in hazardous or hard-to-access environments, significantly improving inspection efficiency and accuracy.

[0003] As an important component of high-end equipment, intelligent inspection robots integrate multiple functional modules such as perception, decision-making, and execution. They can replace manual labor in completing periodic inspection tasks in unattended or high-risk environments, significantly improving the safety and efficiency of operation and maintenance. Among these capabilities, path planning, as one of the core abilities of autonomous robot movement, directly affects the completeness of inspection coverage, the timeliness of task execution, and the economy of energy consumption.

[0004] In the field of industrial inspection, path optimization technology in dynamic environments has become a key breakthrough for achieving efficient inspection. This technology aims to adjust the robot's path online based on real-time environmental information, changes in task priority, and fluctuations in the target area's state, adapting to constantly changing operational needs. Ideally, dynamic path optimization should balance global reachability and local responsiveness, ensuring full coverage while avoiding repetitive travel or path oscillations, thereby achieving a multi-objective balance between time, energy consumption, and inspection quality.

[0005] Existing technologies still have significant shortcomings in the dynamic optimization of intelligent inspection robot paths: First, most methods rely on preset static maps and fixed inspection sequences, making it difficult to cope with path interruptions caused by sudden obstacles or temporary restricted areas. Second, the response mechanism to environmental changes is lagging, lacking the ability to instantly analyze sensor data streams and quickly replan, causing robots to frequently stall or detour through inefficient paths under dynamic interference. Third, path update strategies often only consider the shortest geometric distance, ignoring task semantic information such as equipment heating status and fault risk level, resulting in missed inspections of critical areas or imbalanced resource allocation. Fourth, in multi-robot collaborative scenarios, there is a lack of a unified dynamic coordination mechanism, making it easy for path conflicts, overlapping inspection areas, or communication congestion to occur between robots, seriously affecting the overall system operating efficiency. These problems are particularly prominent in application scenarios with complex structures and extremely high safety requirements, such as large substations and petrochemical plants, and have become technical bottlenecks restricting the improvement of the practical application level of intelligent inspection systems. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies by providing a method and system for dynamic path optimization of intelligent inspection robots, which can effectively solve the problems in the background technology. In the field of robotic arms and robots in high-end equipment applications, especially in intelligent inspection tasks in complex industrial environments, traditional path planning methods generally rely on static maps and preset routes, which are difficult to adapt to the actual working conditions of frequent changes in on-site equipment layout, dynamic appearance of obstacles, and concurrent execution of multiple tasks. Such methods usually generate initial paths based on static graph search algorithms such as A* or Dijkstra, lacking the ability to effectively integrate real-time environmental perception data, causing the robot to frequently trigger emergency stops or detours during operation, seriously affecting inspection efficiency and system reliability. In addition, existing technologies have failed to construct a quantitative correlation model between the degree of environmental change and the path replanning trigger mechanism, resulting in excessively frequent replanning or delayed response, further exacerbating the technical contradiction between computational resource consumption and decreased motion continuity. This invention, by constructing a closed-loop control architecture that integrates multi-source environmental perception, dynamic risk assessment, and adaptive replanning decision-making, achieves real-time generation and smooth switching of highly robust, low-latency, and optimal energy-efficient paths for inspection robots in unstructured environments.

[0007] To achieve the above objectives, the present invention provides the following technical solution: On one hand, an intelligent inspection robot path dynamic optimization system, comprising the following components: an environmental perception module, used to collect in real time spatial point cloud data, obstacle thermal feature information, and illumination condition parameters within the inspection area via a multimodal sensing array composed of lidar, depth camera, and infrared sensor, and to perform time synchronization and coordinate unification calibration on the raw data, outputting a structured three-dimensional environmental state representation; and a local topology modeling module, connected to the environmental perception module, used to construct a locally occupied grid map within a dynamically scrolling window centered on the robot's current position, voxelizing the point cloud data into a 0.1m × 0.1m grid. The system uses lattice cells to identify passable areas, temporary obstacles, and potential danger zones based on the difference between echo intensity and surface normal vectors, forming an incremental environment model with semantic labels. A global path guidance module loads a pre-built factory-level topology map, determines the target inspection point sequence based on task scheduling instructions, and generates an initial collision-free path curve from the start point to the end point using an improved RRT* algorithm. This path is represented as a series of discrete path nodes and their corresponding attitude angle sets. A dynamic risk assessment module connects the local topology modeling module and the global path guidance module, used to compare the matching degree between the local environment model and the global path prediction segment in real time, and calculate the path deviation index (PII). The system uses the Deviation Index (VDI) and the Obstacle Approach Rate (OAR) as weights to generate a Viability Fitness Score (VFS). A replanning decision module, connected to the dynamic risk assessment module, sets a path feasibility threshold VFS_th = 0.75. When VFS falls below this threshold for three consecutive sampling periods, a path replanning instruction is triggered, and a replanning request signal is sent to the path optimization engine. The path optimization engine receives the trigger signal from the replanning decision module, calls the Mixed Integer Quadratic Programming (MIQP) solver, and, while satisfying robot dynamics constraints, regenerates a smooth and safe alternative path with the objective functions of minimizing path length, cumulative change in steering angle, and distance from high-risk areas. A trajectory smoothing and tracking module, connected to the path optimization engine, uses B-spline interpolation to process the newly generated path, generating a continuously differentiable velocity-acceleration profile. This module employs a Model Predictive Control (MPC) algorithm to adjust the motor output torque in real time, ensuring the robot can stably travel along the optimized trajectory. The MPC algorithm improves the robot's motion accuracy, speed, and efficiency by predicting future states and optimizing control inputs, taking into account the robot's dynamics and environmental constraints. The communication coordination interface module enables high-speed data interaction between the aforementioned functional modules, employing the ROS 2 middleware framework to ensure the real-time performance and reliability of message transmission. Preferably, in the environmental perception module, the scanning frequency of the lidar is set to 10Hz, the frame rate of the depth camera is 30fps, and the system is equipped with a timestamp alignment unit. Hardware trigger pulses are used to achieve precise synchronization of multi-sensor data within a ±2ms time window, avoiding spatial misalignment errors caused by asynchronous acquisition. Furthermore, in the local topology modeling module, the voxelization process introduces an adaptive resolution mechanism. When a curve with a curvature greater than 0.05 rad / m is detected in the path ahead, the mesh accuracy of that area is automatically increased to 0.05 m × 0.05 m, while the remaining areas maintain the basic resolution. This reduces the overall computational load while ensuring the modeling accuracy of key areas. In addition, in the global path guidance module, the improved RRT* algorithm introduces a target bias strategy in the random sampling stage, sets the target point sampling probability to pbias=0.2, and adds a collision cone check step before connecting neighboring nodes, allowing only nodes with an included angle of less than 60° to expand edges, which significantly improves the convergence speed and path quality of the tree structure towards the target direction. Preferably, in the dynamic risk assessment module, the path deviation index (PII) is defined as the ratio of the distance from the current robot pose to the nearest path node to the local window radius, and the obstacle approach rate (OAR) estimates the projection component of the relative velocity of the obstacle ahead in the path tangential direction using Kalman filtering. Both are assigned weights. and This constitutes the overall score. ; Furthermore, the replanning decision module is equipped with a two-level triggering mechanism, with the initial warning state being... At this point, a lightweight path pre-calculation process is initiated in the background to generate a set of candidate paths in advance. When entering an emergency replanning state, i.e., VFS < 0.6, the optimal candidate path is directly selected from these paths for switching, shortening the response latency to within 300ms. Furthermore, in the path optimization engine, the objective function expression is: ,in This is the total path length. It is the sum of the absolute values ​​of all steering angle differences. The coefficient is the reciprocal integral term of the minimum distance from the path point to the boundary of the known high-temperature and high-pressure equipment. , , The results were determined through fitting of a large amount of measured data to ensure a balance between energy consumption, comfort, and safety. Preferably, in the trajectory smoothing and tracking module, the B-spline interpolation order is set to 3, the dynamic adjustment range of the control point spacing is 0.3m to 1.2m, and it is automatically adjusted according to the path curvature. The high curvature segment uses a smaller spacing to maintain shape fidelity, and the straight segment increases the spacing to reduce data redundancy. On the other hand, a method for dynamic path optimization of an intelligent inspection robot is provided. The specific steps of this method are as follows: Step S110: Using a multimodal sensor array deployed on the inspection robot body, real-time acquisition of 3D point cloud data, temperature distribution images, and visible light video streams of the working environment is achieved, completing the time synchronization and spatial registration of multi-source information, and constructing an environmental observation set under a unified coordinate system; Step S120: Based on the robot's current position, a dynamic perception window with a radius of 5m is established, and the synchronized point cloud data is divided into two categories: static background and dynamic objects. The contours of moving obstacles are extracted using the Euclidean clustering algorithm, and their motion trend vectors are labeled; Step S130: Combining a pre-stored factory topology map and the current inspection task list, a global path planning unit generates a baseline navigation path from the current point to the next target point. The path is in discrete... The path point sequence is stored in the shared memory area; Step S140: Real-time monitoring of the spatial consistency between the local environment model and the baseline path, calculation of the path deviation index (PII) and obstacle approach rate (OAR), and fusion to generate a path feasibility score (VFS); Step S150: Determining whether the VFS is continuously lower than the set threshold (VFS_th). If it is continuously lower, the path replanning process is activated; otherwise, the original path tracking continues; Step S160: Under the condition of triggering replanning, the mixed integer quadratic programming solver is called to comprehensively consider the three indicators of path length, turning cost, and safety distance to generate a new optimized path; Step S170: B-spline curve fitting is performed on the newly generated path to generate a continuous and differentiable reference trajectory, and the robot is driven to perform trajectory tracking through the model prediction controller to complete the path switching operation; Preferably, in step S120, the dynamic object classification uses a support vector machine (SVM) combined with motion feature vectors for discrimination. The feature vectors include four parameters: average velocity, acceleration variance, profile aspect ratio, and echo stability index, achieving a classification accuracy of over 98.7%. Furthermore, in step S140, the path feasibility score VFS is updated every 100ms, and the historical score sequence cache length is N=20, which is used to detect changes in the score trend. When a continuous downward trend is found to have exceeded 8 cycles, the path optimization engine resources are preloaded in advance to improve the agility of the system response. Furthermore, in step S160, the constraints of the mixed integer quadratic programming problem explicitly include the robot's maximum linear velocity vmax=1.2m / s, maximum angular velocity ωmax=0.6rad / s, minimum turning radius rmin=0.8m, and safe distance d_safe≥1.5m from the danger zone, ensuring that the generated path conforms to the actual motion capability limitations; Preferably, in step S170, the prediction time domain of the model predictive control algorithm is set to Tp=2.5s, the control time domain to Tc=1.0s, and the sampling period to Ts=50ms. Within each control period, the optimal control problem in the finite time domain is solved, and the differential drive command of the front and rear wheels is output to achieve high-precision trajectory tracking, with the lateral tracking error controlled within ±0.08m. Compared with the prior art, the present invention has the following beneficial effects: By constructing a closed-loop architecture of "perception-evaluation-decision-optimization", a rapid response to complex dynamic environments is achieved, and the average time for path replanning is reduced from 1.2s in traditional methods to 0.4s, improving the response speed by 3 times. The introduction of a path feasibility quantitative scoring mechanism effectively avoids frequent replanning caused by minor perturbations, reducing the number of replanning triggers by 65% ​​and significantly improving motion continuity and energy utilization efficiency. An alternative path is generated by using a multi-objective mixed integer optimization model. Under the premise of ensuring a safe distance, the average length of a single inspection path is shortened by 12.3%, the steering action is reduced by 21.5%, and the service life of the mechanical transmission system is effectively extended. By combining trajectory smoothing with model predictive control, a continuous transition of speed and direction is achieved during path switching, reducing the peak impact acceleration to below 0.3 m / s², thus improving the stability of robot operation and the quality of sensor data acquisition. The system adopts a modular design, supports flexible deployment on different models of inspection robot platforms, has strong adaptability, and has been successfully verified in various scenarios such as substations, chemical plants, and rail transit, and has broad engineering application prospects. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the overall technical architecture of the intelligent inspection robot path dynamic optimization method and system proposed in this invention; Figure 2 This is a schematic diagram illustrating the core principle framework of the dynamic risk assessment and adaptive replanning decision-making mechanism in this invention. Detailed Implementation

[0009] Please refer to Figure 1 and Figure 2 To further illustrate the technical means and effects of the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0010] Example 1 In the typical unstructured industrial scenario of large substations, intelligent inspection robots need to perform periodic infrared thermography and partial discharge detection tasks in a dynamic environment composed of densely arranged high-voltage equipment, mobile maintenance vehicles, and temporary fences. In this scenario, equipment thermal states change frequently, and channel conditions are unpredictable, making it difficult for traditional path planning strategies based on static A* algorithms to maintain continuous and efficient operation. To address this problem, this embodiment constructs a closed-loop system architecture that integrates multi-source perception, local modeling, global guidance, and adaptive replanning to achieve highly robust path optimization.

[0011] After system startup, the environmental perception module, using a multimodal sensor array consisting of a 16-line LiDAR, an RGB-D depth camera, and a long-wave infrared thermal imager mounted on the robot's top gimbal, begins synchronously acquiring environmental information. The LiDAR emits a laser beam at a frequency of 10Hz to acquire spatial 3D point cloud data, generating approximately 32,000 valid points per scan. The depth camera operates at 30fps, outputting a depth map and color image with a resolution of 640×480. The infrared sensor captures the surface temperature distribution of equipment within the field of view, forming a grayscale mapping thermal map with a temperature measurement range of -20℃ to +150℃ and an accuracy of ±2℃. All three types of sensors are equipped with hardware timestamp units, with a unified trigger signal provided by a pulse generator on the main control board, completing data acquisition synchronization within a ±2ms time window to avoid spatial misalignment caused by movement. All raw data undergoes coordinate transformation by the FPGA preprocessing unit, uniformly converting it to a right-handed Cartesian coordinate system with the robot chassis center as the origin, and is encapsulated into ROS 2 message packets with time tags, transmitted to the local topology modeling module via gigabit Ethernet.

[0012] The local topology modeling module receives the synchronized point cloud stream and establishes a dynamic scrolling window with a radius of 5m centered on the robot's current position, dividing the entire point cloud space into a 0.1m × 0.1m two-dimensional occupied grid map. During the systematization process, the system first performs noise reduction filtering on the point cloud, using a statistical outlier removal method to eliminate noise points with a distance from the mean exceeding two standard deviations. Subsequently, for each grid cell, the consistency index between its internal point cloud density, average echo intensity, and surface normal vector is calculated. If the point cloud density within a grid cell is greater than a set threshold (8 points / cell) and the absolute value of the Z component of the normal vector is less than 0.3, it is determined to be a passable area on the ground; if the echo intensity is significantly higher than the background mean (1.5 standard deviations) and has vertical distribution characteristics, it is marked as a static obstacle; for point clusters whose positions shift within multiple consecutive sampling periods, Euclidean clustering algorithm is used for segmentation, extracting their boundary contours and fitting minimum bounding boxes, combined with an SVM classifier to determine their motion attributes. The feature vector used for SVM classification includes: the target's average moving speed, acceleration variance, aspect ratio (dimensionless), and laser echo stability index (defined as the reciprocal of the coefficient of variation of the echo intensity of the same target between two adjacent frames). Field testing verified that the classification model achieved an accuracy of 98.7% in identifying personnel, vehicles, and floating objects, with a false alarm rate of less than 1.2%. The system further estimates the motion trend vector of dynamic obstacles based on a Kalman filter, including position, velocity, and acceleration state variables, predicting their trajectory envelope over the next 3 seconds for risk assessment.

[0013] Meanwhile, the global path guidance module loads a pre-built plant-level topology map, which stores the geographical coordinates, equipment type, and priority weights of key inspection nodes (such as transformers, switchgear, and surge arresters) in XML format. Based on the task scheduling instructions, the sequence of target points for this inspection is determined, for example, sequentially visiting the T1 main transformer, G2 bay, and C3 capacitor bank. On this basis, an improved RRT* algorithm is called to generate an initial collision-free path from the current robot pose to the next target point. This algorithm introduces a target bias strategy during random sampling in the configuration space: each sampling has an 80% probability of uniformly selecting new nodes from free space, and a 20% probability of directly selecting the target point as the sampling direction, thereby accelerating the convergence of the search tree to the target. In the neighborhood node connection stage, a collision cone check mechanism is added—edge connections are only allowed when the angle between the connection between the candidate parent node and the new node and the tangent of the existing path is less than 60°, preventing drastic path reversals. Furthermore, in each rewiring optimization, a heuristic pruning strategy is used to remove redundant nodes, ensuring that the final output path node sequence satisfies both geometric optimality and good curvature continuity. The generated baseline path is stored in a shared memory area as discrete path points. Each point contains three-dimensional coordinates (x, y, z), heading angle θ, and suggested driving speed v. The update cycle is once every 5 seconds or triggered when the mission changes.

[0014] The dynamic risk assessment module continuously reads local occupancy grid map and global path prediction segment data, performing real-time matching degree analysis. Specifically, the system calculates a path feasibility score every 100ms. First, it calculates the path deviation index. Project the robot's current precise pose (obtained through IMU and wheel velocity fusion localization) onto the nearest path node, calculate its lateral deviation d, and then divide it by the local sensing window radius R = 5m to obtain... When the robot closely follows the path, the obstacle proximity index (PII) approaches 0; if it deviates completely from the path, the PII approaches 1. Next, the obstacle approach rate (OAR) is calculated: All dynamic obstacles within the preceding 3m path segment are traversed. Using the relative velocity vector output by the Kalman filter, its projection component (vtangent) in the path tangent direction is extracted, and the maximum positive value is taken as the OAR input. This value reflects the urgency of the approaching obstacle. Finally, the two indicators are combined to generate a path feasibility score. Its expression is: Among them, the weighting coefficient , The method was determined through regression fitting based on a large amount of field test data, ensuring a balance between path deviation and obstacle threat. The value ranges from [0,1], with higher values ​​indicating a safer and more feasible current path. The system maintains a historical score cache queue of length N=20 for trend analysis. Whenever a new score enters, it checks for a continuous downward trend. If the VFS decreases monotonically for more than 8 consecutive periods, it is determined that the environment is deteriorating significantly, and the background resource preloading process is activated in advance to prepare for possible replanning.

[0015] The replanning decision module receives the VFS output and performs a two-stage trigger judgment. The path feasibility threshold VFS_th is set to 0.75. When... When the system enters a primary warning state, it does not interrupt the current path tracking but initiates a lightweight path pre-computation subroutine: within a certain range (e.g., ±1.5m) near the current path, it generates several candidate alternative paths using the Fast Path Method (FM) and caches their node sequences and cost estimates. These paths only consider static obstacle avoidance, ignoring dynamic constraints, and the computation time is controlled within 150ms. When VFS < 0.6, the system immediately switches to an emergency replanning state, sending a high-priority replanning request signal to the path optimization engine, and simultaneously selecting the lowest-cost path from the candidate set as a transitional path until a completely new optimized path is generated.

[0016] After receiving the trigger signal, the path optimization engine constructs a mixed-integer quadratic programming (MIQP) problem model to solve for a new safe path. The objective function is defined as: in The total length of the path to be optimized (in meters). It is the sum of the absolute values ​​of the differences in turning angles between all adjacent nodes on the path (unit: radians). This is the inverse of the minimum distance from the path point to the boundary of the known high-temperature, high-pressure equipment, integrated along the path, i.e.: Here Location on the path The actual distance to the nearest danger zone. A mandatory safety interval threshold is used to prevent integral divergence. The coefficients are determined by fitting measured data from multiple operating conditions. (Focusing on shorter paths) (Suppressing frequent turning) (Enhanced safety avoidance). The constraints explicitly include limitations on the robot's own motion capabilities: maximum linear velocity vmax = 1.2 m / s, maximum angular velocity ωmax = 0.6 rad / s, minimum turning radius rmin = 0.8 m, and a buffer distance of d ≥ 0.5 m from dynamic obstacles. The MIQP solver uses the IPOPT nonlinear optimization library combined with the branch and bound method to handle integer variables. The inputs are the current robot pose, target point coordinates, local obstacle distribution, and semantic labels of the danger zone. The output is a new sequence of path nodes, with the node spacing initially set to 1.0 m and subsequently adjusted according to curvature.

[0017] The trajectory smoothing and tracking module receives new path nodes and performs B-spline curve fitting. A cubic B-spline (order k=3) is selected to ensure the position, velocity, and acceleration of the generated trajectory are continuously differentiable. The control point selection strategy is as follows: in straight sections or low-curvature regions (curvature κ < 0.03 rad / m), the control point spacing is set to 1.2 m to reduce data redundancy and maintain trajectory smoothness; in curved regions (κ ≥ 0.03 rad / m), the control point spacing is automatically reduced to 0.3 m to 0.6 m to ensure the fidelity of the trajectory shape, which is particularly important in industrial robot trajectory planning. After fitting, a parametric curve C(u), u∈[0,1], is generated and discretized into reference trajectory points with equal time intervals, with a time step Ts=50 ms. This trajectory point sequence is sent to the Model Predictive Controller (MPC) for trajectory tracking. The MPC is set with a prediction time Tp = 2.5s (50 steps) and a control time Tc = 1.0s (20 steps). Within each control cycle, a finite-time optimal control problem is solved, with the front and rear wheel differential drive commands as the optimization variable. The state equation is based on a bicycle model, considering tire lateral deviation and motor response delay. The cost function includes tracking error, control increment, and terminal state penalty terms. The controller outputs a PWM duty cycle signal to the motor driver to adjust the left and right wheel speeds, achieving high-precision path following. Real-world data shows that the lateral tracking error is stably controlled within ±0.08m, and the peak acceleration spike during path switching is less than 0.3m / s², ensuring the stability of the onboard sensor platform.

[0018] The communication coordination interface module adopts the ROS 2 middleware framework and implements high-speed asynchronous communication between various functional modules based on the DDS (Data Distribution Service) protocol. All modules run as independent nodes on the Ubuntu 20.04 operating system, publishing and subscribing to messages through a topic mechanism. Critical data flows such as point clouds, path nodes, and control commands are configured with QoS policies: reliability is set to RELIABLE, historical depth is set to KEEP_LAST_N=5, and lifecycle management is enabled with automatic discovery and fault recovery mechanisms. Cross-process communication latency is controlled within 10ms, and end-to-end system response latency (from perception to execution) does not exceed 300ms, meeting real-time requirements.

[0019] Throughout the operation, the system forms a closed-loop feedback loop of "perception → modeling → evaluation → decision-making → optimization → execution". For example, when the robot is moving along the predetermined path towards the T1 main transformer, a parked maintenance forklift suddenly appears ahead, causing a large area of ​​dynamic obstacles to appear in the local occupied grid. At this time, the path deviation index (PII) rapidly rises to 0.45, the obstacle approach rate (OAR) reaches 0.9 m / s, and the velocity-speed-free (VFS) drops to 0.68 and continues to decline. The system first enters a primary warning state, and the background pre-generates two detour paths: one detours to the left, crossing the equipment gap; the other detours to the right, passing through the open passage. Because the right-hand path is lower in cost and farther from the high-temperature area, it is selected as the candidate optimal. As the VFS value further decreases to 0.55, the system triggers an emergency replanning mechanism. The path optimization engine responds quickly, calling the hybrid quadratic integer programming (MIQP) solver to fine-tune the existing path candidate schemes. In this process, dynamic constraints are added to ensure that the generated alternative paths not only meet real-time requirements but also take into account the robot's dynamic characteristics, thereby improving the efficiency and safety of path planning. The trajectory smoothing module performs B-spline interpolation on the path, and the MPC controller smoothly switches the control target. The robot travels along the new path at a speed of 0.8 m / s without any sudden stops or collisions, and successfully completes the task.

[0020] Example 2 Unlike the first embodiment, which focuses on path optimization for a single robot in a substation scenario, this embodiment focuses on a multi-robot collaborative inspection scenario in a chemical plant, emphasizing the resolution of path conflicts and resource competition among multiple robots. In such environments, multiple inspection robots of the same type simultaneously execute tasks with different priorities. Without a unified coordination mechanism, path intersections, deadlocks, or duplicate coverage can easily occur. To address this, this embodiment adds a distributed negotiation module to the original system architecture and reconstructs the global path guidance and replanning decision logic to achieve decentralized dynamic path coordination.

[0021] The system still includes basic components such as an environmental perception module, a local topology modeling module, and a trajectory smoothing and tracking module, and its technical implementation is basically the same as in Embodiment 1. The main difference lies in the functional expansion of the global path guidance module and the replanning decision module. In this embodiment, the global path guidance module not only loads the local plant topology map, but also needs to access the global task allocation table published by the central task server. This table records the ID, current position, target sequence, task urgency level (divided into four levels: P0-P3), and estimated completion time of all online robots. When a robot completes the detection of the current target point, it requests the next task from the server, and the server assigns a new target point according to the task load balancing strategy.

[0022] The key improvement lies in the introduction of the concept of a "virtual repulsion field" for path conflict prediction. Each robot constructs a dynamic repulsion region around its local path prediction segment. This region is a strip-shaped space with its width dynamically adjusted according to the task level, centered on the path: 2.0m for P0 level, 1.8m for P1 level, 1.5m for P2 level, and 1.2m for P3 level. This repulsion region information is periodically broadcast to neighboring robots by the communication coordination interface module (communication radius set to 50m), using UDP multicast to reduce network overhead. After receiving the repulsion field data from other robots, each robot overlays it onto its own local occupancy grid map, forming an enhanced environment model that includes "social obstacles."

[0023] The dynamic risk assessment module has upgraded its scoring mechanism. In addition to the existing Path Deviation Index (PII) and Obstacle Approach Rate (OAR), a new "Conflict Risk Index (CRI)" has been added. CRI is defined as the ratio of the length of the spatial intersection between the current robot's path and the repulsion field of any other robot (lcross) to the length of its own path (lp), i.e., CRI = lcross / lp. If there is no intersection, CRI = 0; if the paths completely overlap, CRI = 1. This index is updated every 100ms and participates in the comprehensive score calculation. At this time, the Path Feasibility Score (VFS) is revised as follows: in This reflects a moderate level of concern for multi-machine conflicts, avoiding triggering replanning due to slight overlap. When VFS is below the threshold of 0.75 and CRI is greater than 0.3, it is judged as a high-conflict-risk state, and the negotiation mechanism is initiated first rather than direct replanning.

[0024] The replanning decision module has added a distributed negotiation submodule. When a high risk of conflict is detected, the system first attempts to resolve the issue through a lightweight negotiation protocol. The specific process is as follows: the initiating robot sends a negotiation request to the conflicting robot, including its own task level, remaining path, estimated transit time, and acceptable detour cost limit. The opposing robot assesses whether to yield based on its local state: if its task level is higher or the detour cost is too high, it replies with a refusal; otherwise, it accepts and promises to adjust its path. The negotiation process can last for a maximum of 3 rounds. If no agreement is reached within the time limit, each robot initiates its own independent replanning.

[0025] The path optimization engine adjusts the objective function weight strategy in multi-robot scenarios. For robots with higher task levels (P0 / P1), the path length weight α is reduced to 0.8, while the safety and conflict avoidance term γ is increased to 1.8 to ensure their priority passage. For robots with lower task levels (P2 / P3), the turning cost β is increased to 1.0 to encourage them to actively avoid obstacles. Furthermore, during MIQP solving, the predicted trajectory of the other robot is incorporated as a soft constraint, allowing for brief intrusions but imposing penalties to improve path flexibility.

[0026] In actual operation, when two robots approach the same utility tunnel intersection from the north and south sides respectively, they each detect overlapping repulsion fields, causing the CRI to rise to 0.4 and the VFS to drop to 0.72. The system enters a primary warning state and initiates negotiation. The task level of the robot on the south side is P1, and that of the robot on the north side is P2. The latter actively yields and replans its path to detour through the backup passage on the east side. The entire process requires no human intervention, avoiding potential deadlock. In an experiment, when 10 robots were operating simultaneously, by adopting advanced path planning and conflict detection technologies, the occurrence rate of path conflicts was significantly reduced from 23% using traditional methods to 4.1%, while the standard deviation of task completion time decreased by 37%. This demonstrates that effective path coordination and obstacle avoidance strategies can significantly improve the overall inspection efficiency and system throughput of multi-robot systems.

[0027] Example 3 This embodiment targets nighttime inspection scenarios in rail transit depots, focusing on solving the challenges of low light, strong reflection interference, and detection of small obstacles. It prioritizes improvements to the data fusion strategy of the environmental perception module and the semantic recognition capabilities of the local topology modeling module. Compared to the previous two embodiments, this embodiment exhibits substantial technical differences in sensor configuration, data processing flow, and risk assessment.

[0028] The environmental perception module employs a heterogeneous sensor combination: in addition to a conventional 16-line lidar, a 77GHz millimeter-wave radar is added, featuring an angular resolution of 1.2° and a detection range of up to 100m. Compared to the 24GHz radar, the 77GHz radar has stronger water vapor penetration capabilities in adverse weather conditions, such as rain and fog, effectively detecting metal parts left behind on the track or areas of accumulated water. Furthermore, a high dynamic range visible light camera (HDR mode, 120dB signal-to-noise ratio) can still provide clear images even in environments with complex lighting and strong contrast between light and dark. Data from the three types of sensors is still synchronously acquired via hardware triggering, with time alignment accuracy controlled within ±3ms.

[0029] The data fusion strategy employs a three-level cascaded architecture. The first level is spatial alignment: a calibration matrix maps millimeter-wave point clouds and image pixels to the laser point cloud coordinate system. The second level is feature-level fusion: geometric edge features of the laser point cloud, Doppler velocity features of the millimeter-wave radar, and texture gradient features of the image are extracted and concatenated into a joint feature vector. The third level is decision-level fusion: a random forest classifier is trained to discriminate each spatial entity, with categories including "normal track area," "rail surface foreign object," "water accumulation and reflective strip," "personnel activity," and "equipment box door open." The classifier input contains 12-dimensional features: point cloud density, height variance, echo intensity gradient, Doppler mean, velocity standard deviation, radar RCS cross-section, image brightness mean, contrast, edge density, color histogram offset, motion consistency ratio, and time persistence factor. The model trained using multi-sensor fusion technology performs excellently on a real vehicle depot dataset, achieving a recall rate of 96.4%, especially in identifying small foreign objects with a diameter less than 10 cm, where its performance significantly outperforms single-sensor solutions.

[0030] The local topology modeling module has correspondingly enhanced semantic annotation capabilities. When generating the occupied grid map, it not only marks the occupancy status but also adds semantic label fields. For example, obstacles such as "rail surface foreign objects" are assigned a high-risk level (Level 3); "equipment box door open" is marked as medium-risk (Level 2); and "personnel activity" is determined based on their movement trends to determine whether they have entered a safe zone. Semantic information directly affects subsequent risk assessments.

[0031] The dynamic risk assessment module introduces a "Semantic Risk Weighting Factor (WSR)". In the original VFS formula, the terms (1−PII) and exp(−OAR) are multiplied by WSR, and its value is dynamically adjusted according to the obstacle type: WSR = 1.0 for ordinary static obstacles, WSR = 1.3 for small foreign objects, WSR = 1.2 for water accumulation areas, WSR = 1.5 for personnel approaching, and WSR = 1.4 for abnormal equipment conditions. This mechanism makes the system more sensitive to specific industrial hazards. For example, when a wrench is detected lying on the track, even if its physical size is small and its OAR is low, WSR = 1.3 may still cause the VFS to quickly drop below the threshold, triggering replanning and detour, and reporting an alarm.

[0032] The global path guidance module also considers the semantic risk map when generating the initial path. The plant topology map pre-stores "high-risk inspection sections," such as switch areas and oil-prone areas. The improved RRT* algorithm avoids sampling within a 1.5m radius of these areas unless the task mandates entry. A semantic risk integral term is added to the path cost function to guide the robot to prioritize low-risk routes.

[0033] In practical applications, during nighttime inspections, the robot detected oil stains and reflections on the track surface using deep learning technology, which the lidar misidentified as an open passage. However, millimeter-wave radar detected an abnormal dielectric constant, and the HDR camera identified specular reflection characteristics. Combining these sensor data and using fusion classification technology, the area was ultimately confirmed as a "water-reflective zone." The system immediately increased the risk level of the area, with the VFS plummeting from 0.82 to 0.58, triggering a replanning. The new path guided the robot around the maintenance passage, and an automatic "track slippery risk" warning was reported. By employing multimodal perception technology combined with semantic-driven path optimization, potential slippage accidents were effectively avoided, significantly improving the safety of autonomous driving.

[0034] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention are within the scope of the present invention.

Claims

1. A method for dynamic path optimization of an intelligent inspection robot, characterized in that, include: The environmental perception data of the inspection area is collected in real time by a multimodal sensor array. The environmental perception data includes spatial point cloud data, thermal feature information of obstacles, and visible light images. Based on environmental perception data, a local occupancy grid map is constructed with the robot's current position as the center, and passable areas, temporary obstacles and potential danger zones are identified according to the difference between echo intensity and surface normal vector, forming an incremental environmental model with semantic labels. By combining the pre-stored global topology map and task scheduling instructions, a baseline navigation path is generated from the current point to the next target point; The incremental environmental model and the predicted segments of the baseline navigation path are compared in real time to calculate the path deviation index and obstacle approach rate, and a path feasibility score is generated based on the weighted fusion of the two. When the path feasibility score remains below a preset threshold, the path replanning process is triggered. In the path replanning process, a new optimized path is generated with the objective function of minimizing the path length, the cumulative change in turning angle, and the degree of proximity to high-risk areas. The new optimized path is smoothed, and the robot is driven to perform trajectory tracking through a model predictive control algorithm.

2. The method for dynamic path optimization of intelligent inspection robots according to claim 1, characterized in that, Environmental perception data of the inspection area is collected in real time through a multimodal sensor array, including: Raw data is collected synchronously using a sensor array consisting of lidar, depth camera and infrared sensor; The original data is time-stamp aligned and coordinate system calibrated to output a structured three-dimensional environmental state representation as the environmental perception data.

3. The method for dynamic path optimization of intelligent inspection robots according to claim 1, characterized in that, The calculation of the path deviation index and obstacle approach rate includes: Project the robot's current pose onto the nearest path node on the reference navigation path, and calculate the ratio of its lateral deviation to the radius of the local perception window as the path deviation index; The projection component of the relative velocity of the dynamic obstacle ahead in the tangential direction of the path is estimated using Kalman filtering, and is used as the obstacle approximation rate.

4. The method for dynamic path optimization of intelligent inspection robots according to claim 1, characterized in that, When the path feasibility score remains below a preset threshold, a path replanning process is triggered, including: A two-level triggering mechanism is set up so that when the primary warning threshold is met, the background lightweight path pre-calculation process is started to generate a candidate path set. When the emergency replanning threshold is met, the optimal path is selected from the candidate path set for switching, or the path optimization engine is directly invoked to generate a brand new optimized path.

5. The method for dynamic path optimization of intelligent inspection robots according to claim 1, characterized in that, Using minimizing path length, cumulative change in steering angle, and proximity to high-risk areas as objective functions, a new optimized path is generated, including: Construct a mixed-integer quadratic programming problem model with constraints including the robot's maximum linear velocity, maximum angular velocity, minimum turning radius, and safe distance from the danger zone. The solver is invoked to solve the mixed-integer quadratic programming problem model, and the output path node sequence that satisfies all constraints is used as the new optimized path.

6. The method for dynamic path optimization of intelligent inspection robots according to claim 1, characterized in that, The new optimized path undergoes trajectory smoothing, including: B-spline curve fitting is performed on the path node sequence of the new optimized path to generate a continuously differentiable reference trajectory. Among them, the spacing between control points of the B-spline is automatically adjusted according to the curvature of the path. A smaller spacing is used in high curvature segments to maintain shape fidelity, and the spacing is increased in straight segments to reduce data redundancy.

7. The method for dynamic path optimization of intelligent inspection robots according to claim 1, characterized in that, In multi-robot collaborative scenarios, it also includes: It receives virtual repulsive force field information from neighboring robots and overlays it onto its own local occupancy grid map; The ratio of the spatial intersection length of the current path and the virtual repulsive force field to the length of its own path is calculated as a conflict risk index; The conflict risk index is incorporated into the calculation of the path feasibility score to prioritize resolving path conflicts through negotiation mechanisms.

8. The method for dynamic path optimization of intelligent inspection robots according to claim 7, characterized in that, Incorporating the conflict risk index into the calculation of the route feasibility score includes: Add a weighted term related to the conflict risk index to the path feasibility score; When the path feasibility score is lower than the preset threshold and the conflict risk index is higher than the conflict determination threshold, a distributed negotiation protocol is initiated instead of directly triggering path replanning.

9. A dynamic path optimization system for an intelligent inspection robot, characterized in that, include: The environmental perception module is used to collect environmental perception data of the inspection area in real time through a multimodal sensor group. The environmental perception data includes spatial point cloud data, obstacle thermal feature information and visible light images. The local topology modeling module, connected to the environment perception module, is used to construct a local occupancy grid map centered on the robot's current position based on the environment perception data, and to identify passable areas, temporary obstacles and potential danger zones according to the difference between echo intensity and surface normal vector, forming an incremental environment model with semantic labels. The global path guidance module is used to combine a pre-stored global topology map with task scheduling instructions to generate a baseline navigation path from the current point to the next target point; The dynamic risk assessment module connects the local topology modeling module and the global path guidance module. It is used to compare the predicted segments of the incremental environment model and the baseline navigation path in real time, calculate the path deviation index and obstacle approach rate, and generate a path feasibility score based on the weighted fusion of the two. The replanning decision module, connected to the dynamic risk assessment module, is used to trigger the path replanning process when the path feasibility score is continuously lower than a preset threshold. The path optimization engine is used in the path replanning process to generate new optimized paths with the objective functions of minimizing path length, cumulative change in turning angle, and proximity to high-risk areas. The trajectory smoothing and tracking module is connected to the path optimization engine and is used to smooth the trajectory of the new optimized path and drive the robot to perform trajectory tracking through the model predictive control algorithm.

10. The intelligent inspection robot path dynamic optimization system according to claim 9, characterized in that, It also includes a communication coordination interface module for enabling high-speed data interaction between various functional modules; in multi-robot collaborative scenarios, the communication coordination interface module is also used to broadcast the virtual repulsion field information of the local robot and receive the virtual repulsion field information of neighboring robots; the dynamic risk assessment module is also used to calculate the ratio of the spatial intersection length of the current path and the virtual repulsion field to its own path length as a conflict risk index, and incorporate the conflict risk index into the calculation of the path feasibility score.

Images