A closed-loop path planning method and system for a water plant composite ring track inspection robot

Abstract

The application discloses a water plant composite annular track inspection robot closed loop path planning method and system, abstracts a water plant composite annular track network into a weighted directed connected graph, establishes a four-dimensional comprehensive weight model containing path length, passing energy consumption, obstacle density and historical failure rate; adopts a Chinese postman problem CPP Eulerization method to complete Eulerization of non-Euler track graph by minimum weight matching of odd nodes, generates a global closed loop Euler inspection loop through an improved Hierholzer layered variant algorithm; adopts a dynamic window method DWA to perform local obstacle avoidance and trajectory fine adjustment, and dynamically updates track weight based on multi-source sensor data. The application also integrates AI visual intelligent detection, multi-robot collaborative scheduling and closed loop integrity checking mechanism, realizes multi-region unmanned, closed loop and intelligent inspection operation of water plant pump house, filter tank, pipeline corridor and the like, effectively eliminates an inspection blind area, solves a multi-track connection deadlock problem, and significantly improves inspection integrity and working condition self-adaptive capability.

Description

Technical Field

[0001] This invention relates to the field of industrial inspection robot path planning technology, and in particular to a closed-loop path planning method and system for a water plant composite circular track inspection robot. Background Technology

[0002] Routine inspections of water plants (including waterworks, water purification plants, and wastewater treatment plants) are crucial for ensuring water supply safety and water quality compliance. Traditional manual inspection methods have inherent drawbacks such as high workload, low timeliness, and reliance on personnel experience for inspection quality. In recent years, track-mounted inspection robots have been gradually applied to areas such as water plant pump rooms, filter beds, chemical dosing rooms, and pipeline corridors, partially replacing manual labor in equipment status monitoring and environmental inspection tasks.

[0003] However, existing track inspection robot technology has the following main technical shortcomings: (1) Single-track independent cruise mode with fragmented track topology. Existing technical solutions typically use a single closed loop track, with each process zone track being independent of each other and without connecting branch structures. Robots in each area can only conduct inspections within their own zone, making it impossible to achieve multi-area cross-track linkage inspections across the entire plant. The path logic is a preset fixed trajectory or single-point round trip, which only supports fixed-point inspections in a single area, limiting the inspection coverage.

[0004] (2) Path deadlock is prone to occur when multiple inspection robots run on tracks in different process zones. At the track intersection nodes, there is a lack of unified right-of-way scheduling and mutual exclusion lock mechanism, which can easily lead to problems such as multi-robot track intersection conflicts, path deadlock and switching misalignment, resulting in inspection interruption or robot collision.

[0005] (3) Lack of global closed-loop path planning and integrity verification mechanism. Most existing technologies adopt preset fixed trajectories or simple reciprocating motion modes, without abstracting the entire plant track network into a graph theory model, and without using the Eulerian method of the Chinese Postman Problem (CPP) to generate a fully covered closed-loop circuit that traverses all track edges, thus failing to theoretically guarantee that there are no blind spots or omissions in the inspection.

[0006] (4) Poor adaptability to dynamic environments. The water plant inspection environment has complex dynamic characteristics such as high humidity, water mist, low light, water and dust accumulation on the tracks, and frequent entry and exit of temporary workers. The basic obstacle avoidance algorithm of the existing technology can only stop and avoid simple obstacles, and cannot adaptively adjust the passage path and inspection strategy according to the dynamic changes of the environment.

[0007] (5) Low level of intelligence and lack of linkage between AI visual inspection and planning. In the existing technical solutions, the visual inspection and path planning systems are independent of each other and have not formed a closed-loop linkage mechanism of "perception-planning-scheduling". They cannot automatically adjust the inspection path and inspection frequency according to the on-site equipment and environmental conditions.

[0008] Therefore, there is an urgent need for a closed-loop path planning system and method that can achieve full-area, blind-spot-free closed-loop inspection, solve multi-track switching deadlock, dynamically adapt to complex water plant operating conditions, and integrate AI visual inspection and path planning linkage, so as to improve the integrity, stability and intelligence level of water plant inspection. Summary of the Invention

[0009] This invention addresses the shortcomings of existing technologies by providing a closed-loop path planning method and system for a water plant composite circular track inspection robot.

[0010] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows: A closed-loop path planning method for a composite circular track inspection robot in a water plant includes the following steps: S1. Obtain the topology data of the composite ring track of the water plant, wherein the composite ring track includes a main ring track and at least one sub-ring track, and the main ring track and the sub-ring track are connected by electric switches to form a connectable composite ring topology network. S2. Abstract the composite ring track topology network into a weighted directed connected graph. ,in For a set of nodes, For the set of orbital edges, To calculate the comprehensive passage cost weights, a four-dimensional comprehensive weight model is performed on the weighted directed connected graph, wherein the four-dimensional comprehensive weights include a path length factor. Energy consumption factor for passage Obstacle density factor and historical failure rate factor The track edge The overall cost of passage is: in , , , The normalized balance coefficients and ; The path length factor is obtained by linearly normalizing the actual physical distance between two points on the track. The passage energy consumption factor is assigned in a graded manner based on the characteristics of the track structure. The obstacle density factor is assigned based on historical perception data. A historical failure rate factor that is quantitatively assigned based on historical failure logs from the water plant operation and maintenance platform; S3. Based on the weighted directed connected graph, the Chinese Postman Problem (CPP) Eulerianization method is used to perform minimum weight matching of odd-degree nodes on the non-Eulerian orbit graph to complete the Eulerianization, transforming the original graph into an Eulerian graph; the minimum weight matching uses a weighted method. The algorithm calculates the minimum weighted path cost between odd-degree nodes; S4. An improved Hierholzer hierarchical variant algorithm is used to perform hierarchical optimal traversal on the Eulerian graph to generate a global closed-loop Eulerian inspection loop that starts from a preset starting point, traverses all main loops and sub-loops of the composite circular track, and returns to the starting point. This loop is used as a global reference path. The improved Hierholzer hierarchical variant algorithm follows the hierarchical traversal constraint of "first fully traversing a single sub-loop and then returning to the main loop, traversing all sub-loops in sequence, and finally closing the main loop". S5. During the robot's cruise along the global reference path, the Dynamic Window Method (DWA) is used for local trajectory fine-tuning and real-time obstacle avoidance. The operating mode is automatically switched between the global planning layer and the local planning layer according to the preset level switching conditions. S6. Real-time acquisition of environmental perception data and robot status data during the inspection process, and dynamic updating of the obstacle density factor in the four-dimensional comprehensive weight. and historical failure rate factor This enables online iterative updates of orbital weights.

[0011] Furthermore, the dynamic update described in step S6 includes: Offline initialization update: During low-load periods each day, read historical database statistics on obstacle density, slippage frequency, communication packet loss rate, and equipment failure frequency for each track segment, and recalculate. and The quantized value is used to refresh the global topology weight table; Online real-time updates: When an anomaly is detected in a local section during the inspection, the speed of that track section is immediately increased locally. and The weights are modified only in the local weights of the current sub-ring. Coefficient adaptive adjustment: Automatically increases during the rainy season. The obstacle weighting coefficient is automatically increased when the track is prone to slippage in low temperatures during winter. Failure rate weighting coefficient.

[0012] Furthermore, the hierarchical switching conditions in step S5 include: The trigger condition for switching from the global planning layer to the local dynamic planning layer is that any of the following conditions are met: a) The lidar detects static or dynamic obstacles within a 3-meter radius of the direction of travel; b) Water accumulation or slippage on the track causes the robot's posture to deviate from the preset trajectory threshold by more than 5 degrees; c) Sudden changes in lighting or increased water mist concentration cause visual positioning deviation to exceed the limit; d) Received a safety alarm signal for temporary manual operations in the track section; The reset condition for switching back from the local dynamic programming layer to the global planning layer is that all of the following conditions are met: a) Obstacles completely disappear within the detection range, and the duration of continuous unobstructed access. Second; b) The robot's posture and positioning deviation have returned to within the allowable threshold range; c) The obstacle avoidance is complete, and the current track section's traffic status has returned to normal.

[0013] Furthermore, the closed-loop path planning method for the water plant composite circular track inspection robot also includes obstacle classification and graded handling steps: the detected obstacles are divided into three categories: static fixed obstacles, dynamic slow obstacles, and high-risk emergency obstacles; three-level priority handling rules are set: priority 0 corresponds to high-risk emergency obstacles, triggering immediate emergency braking; priority 1 corresponds to dynamic personnel obstacles, prioritizing deceleration and waiting while simultaneously uploading safety alarms; priority 2 corresponds to static road-occupying obstacles, initiating local path detour replanning.

[0014] Furthermore, the closed-loop path planning method for the water plant composite circular track inspection robot also includes a hierarchical path replanning step: When an obstacle is detected to enter the preset safe distance threshold, the trajectory deviates laterally from the global reference path by more than 0.4 meters, the vehicle attitude deviates by more than 5 degrees, or the environmental conditions change abruptly or the track weights change dynamically, the DWA algorithm is used to perform local incremental replanning. When a static obstacle persists for more than 10 minutes, the track topology changes, there is a periodic update, or the scheduling strategy is switched, the CPP Eulerization and weighted Dijkstra algorithm are re-executed to perform a global complete replanning.

[0015] Furthermore, the closed-loop path planning method for the water plant composite circular track inspection robot also includes a step of multi-source positioning data fusion using an extended Kalman filter (EKF): using the global absolute node coordinates of the track output when triggered by the RFID landmark as the global correction reference, using the relative displacement of the track travel output by the wheel odometer as the continuous position prediction input, and using the three-axis angular velocity and acceleration output by the IMU inertial measurement unit for attitude compensation; after fusion, the position error is controlled within a certain range throughout the entire process. Within this range, the positioning error of the turnout section .

[0016] Furthermore, the closed-loop path planning method for the water plant composite circular track inspection robot also includes a closed-loop integrity verification step: Set the trigger timing for four levels of verification: real-time point trigger, sub-ring segment trigger, global loop termination trigger, and background timed silent trigger; Calculate the node traversal completeness rate ;in This represents the number of RFID nodes or inspection points visited in this round. This represents the total number of nodes in the plant's topology. Calculate trackside coverage integrity ;in This represents the number of track edges that have been traversed in this round. This represents the total number of all physical track sections in the entire plant. Calculate the closed-loop degree of Euler circuits ;in This represents the actual closed-loop mileage traveled by the robot. This represents the total mileage of the theoretical CPP Euler circuit; when or or Not here When within the specified range, automatically trigger local backtracking and re-inspection based on the missing type. Sub-path replanning or global CPP replanning.

[0017] Furthermore, the closed-loop path planning method for the water plant composite circular track inspection robot also includes a multi-robot collaborative scheduling step: A dynamic task allocation algorithm based on cost greed and load balancing is adopted to distribute inspection tasks to multiple robots, allocating costs accordingly. ,in For the first Taiwan robot, For the first One inspection task For robots To the mission The orbital weighted distance, For the robot Execute the task The energy cost, Weighting the robot's current remaining battery power. This represents the current number of tasks the robot is currently handling. , , , These are the normalized balance coefficients; A switch track mutual exclusion lock and time-sharing occupation mechanism are adopted to allow only one robot to occupy the same track segment at any given time. Three levels of power thresholds are set for charging scheduling: a warning threshold of 35%, a forced return-to-charge threshold of 20%, and an emergency protection threshold of 10%. The robot automatically selects the station with the lowest charging cost for return-to-charge based on the weighted distance of the robot from the charging station and the current number of people waiting in the queue at the charging station.

[0018] 9. A closed-loop path planning system for a water plant composite circular track inspection robot, characterized in that it comprises: The track topology modeling module is used to acquire the topological structure data of the composite circular track in the water plant, and abstract the composite circular track topology network into a weighted directed connected graph. A four-dimensional comprehensive weight model is then performed, which includes a path length factor. Energy consumption factor for passage Obstacle density factor and historical failure rate factor ; The hierarchical closed-loop path planning module is used to Eulerianize the non-Eulerian trajectory graph based on the weighted directed connected graph using the CPP Eulerianization method, generate a global closed-loop Eulerian inspection loop using an improved Hierholzer hierarchical variant algorithm, and then use a weighted... The algorithm finds the globally optimal Eulerian circuit as the global reference path; The dynamic obstacle avoidance and replanning module is used to perform local trajectory fine-tuning and real-time obstacle avoidance using DWA, and automatically switches the operating mode according to preset level switching conditions. The environmental perception fusion module includes vehicle-mounted sensors and track-fixed sensors. The vehicle-mounted sensors include 2D LiDAR, ultrasonic sensors, depth industrial cameras, RFID readers, water mist / temperature and humidity composite sensors, and three-axis attitude and tilt sensors. The track-fixed sensors include wall-mounted depth cameras, water accumulation detection ultrasonic modules, and RFID landmarks. The AI ​​visual intelligent inspection module is equipped with an improved YOLOv8s lightweight target detection network, a ResNet18+ lightweight FPN semantic segmentation network, and a lightweight ResNet10 infrared thermal imaging classification network. It is used to perform equipment appearance defect detection, pool surface anomaly detection, and on-site personnel safety violation detection. The online weight update module is used to dynamically update the obstacle density factor in the four-dimensional comprehensive weight based on real-time collected environmental perception data and robot state data. and historical failure rate factor ; The multi-robot collaborative scheduling module is used to allocate inspection tasks to multiple robots using a dynamic task allocation algorithm based on cost greed and load balancing, and to coordinate track access rights using a branch track mutual exclusion lock and time-sharing occupancy mechanism. The closed-loop integrity verification module is used to set up a four-level verification trigger mechanism, calculate the node traversal integrity rate, track edge coverage integrity rate and Eulerian circuit closure degree, and automatically trigger supplementary inspection or global replanning when the verification fails.

[0019] Furthermore, the backbone network of the improved YOLOv8s lightweight target detection network adopts a lightweight C2f-Lite module instead of the standard C2f module, and introduces deformable convolutional DCN and multi-scale FPN feature fusion layers, as well as a pre-processing module for adaptive grayscale equalization; the electric switch configured in the composite circular track adopts an industrial waterproof electric switch, including a mechanical body composed of a movable core rail, guide arm, return spring and mechanical position limit block, and a DC deceleration waterproof electric push rod with built-in Hall position sensor as the drive unit. Encrypted RFID markers are pre-embedded 1 meter before and after the switch and used in conjunction with the built-in position sensor of the switch for double verification.

[0020] Compared with the prior art, the advantages of the present invention are as follows: First, it enables closed-loop inspection with no blind spots across the entire area, eliminating blind spots at the source.

[0021] This invention abstracts the composite circular track network of a water plant into a weighted directed connected graph. It employs the Chinese Postman Problem (CPP) Eulerianization method to perform minimum-weight matching of odd-degree nodes in the non-Eulerian track graph, transforming the original graph into an Eulerian graph. Then, an improved Hierholzer hierarchical variant algorithm is used to generate a global closed-loop Eulerian inspection loop that traverses all main and sub-loop track edges. This scheme guarantees the existence of a fully covered closed-loop loop traversing all track edges from a graph theory perspective, enabling the robot to complete a full-coverage closed-loop inspection of all process areas, track segments, and equipment locations throughout the entire plant in a single operation. This completely solves the inspection blind spots and missed inspection problems caused by the fragmentation and regional isolation of traditional single-track systems.

[0022] Second, solve the deadlock problem during multi-track switching and improve the stability of multi-robot collaborative operation.

[0023] This invention utilizes a branching mutual exclusion lock and a time-sharing occupancy mechanism to treat each branching node and intersecting sub-loop as an independent mutually exclusive resource. It employs a closed-loop control process of right-of-way application-response-release, allowing only one robot to occupy the same track segment at any given time. Simultaneously, the improved Hierholzer hierarchical variant algorithm follows a hierarchical traversal constraint of "completely traversing each sub-loop before returning to the main loop," reducing the frequency of branching switching. Combined with precise RFID positioning and verification, this invention fundamentally solves the technical problems of traditional multi-track connection conflicts, trajectory deadlock, and switching misalignment, significantly improving the stability and reliability of multi-robot cross-regional collaborative cruising.

[0024] Third, it dynamically adapts to complex water plant operating conditions, significantly improving environmental adaptability.

[0025] This invention establishes a four-dimensional comprehensive weight model including path length, energy consumption, obstacle density, and historical failure rate. It also sets up a dual-layer dynamic weight update mechanism with offline initialization and online real-time updates, enabling the track weights to adaptively iterate according to water plant operating conditions, seasonal environment, and equipment aging. Simultaneously, it employs a hierarchical planning architecture combining global CPP+Dijkstra and local DWA, setting explicit quantization threshold switching conditions. This allows the system to respond to dynamic disturbances such as track water accumulation, personnel intrusion, and temporary debris within milliseconds, adapting to complex scenarios in water plants such as high humidity, fog, low light, track water and dust accumulation, and temporary operational interference. This effectively solves the problems of poor adaptability, frequent jamming, and frequent shutdowns associated with traditional robots.

[0026] Fourth, the level of intelligence and precision in inspection has been significantly improved.

[0027] This invention integrates an AI visual intelligent detection module, equipped with an improved YOLOv8s lightweight target detection network, a ResNet18+ lightweight FPN semantic segmentation network, and a lightweight ResNet10 infrared thermal imaging classification network. This enables fully automated identification of equipment appearance defects, pool surface anomalies, and personnel violations, replacing manual visual inspection. More importantly, the visual inspection results directly participate in track weight iteration and replanning triggering, forming a closed-loop linkage mechanism of "perception-planning-scheduling." This allows the system to automatically adjust the inspection path and frequency based on the on-site equipment and environmental conditions, balancing inspection quality and efficiency.

[0028] Fifth, reduce operation and maintenance costs and achieve unmanned intelligent operation and maintenance.

[0029] This invention achieves 24 / 7 unmanned continuous inspection operations through a closed-loop inspection process, automatic alarms, tiered handling, intelligent re-inspection scheduling, and data iteration optimization, coupled with multi-robot collaborative scheduling and a three-level power threshold recharge scheduling strategy. The branch road mutual exclusion lock and time-sharing occupancy mechanism ensure the safe collaborative operation of multiple robots, while the closed-loop integrity verification mechanism automatically detects and repairs missed inspection areas, significantly reducing the frequency of manual inspections and the workload of re-inspections, lowering the probability of human error and missed inspections, and comprehensively improving the automation and intelligent operation and maintenance level of water plant inspections. It has extremely high engineering adaptability and feasibility. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1This is a flowchart of the closed-loop path planning method for the water plant composite circular track inspection robot in this embodiment of the invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example 1: A Closed-Loop Path Planning Method for a Water Plant Composite Circular Track Inspection Robot like Figure 1 As shown in the figure, this embodiment provides a closed-loop path planning method for a water plant composite circular track inspection robot. The method is executed by an AI group control intelligent control box deployed at the water plant site and includes the following steps.

[0034] Step S1: Obtain the composite circular track topology data.

[0035] Obtain the physical layout data of the composite circular track of the water plant. The composite circular track includes a closed main loop track and at least one sub-loop track, which are connected by motorized switches. In this embodiment, a water plant has a main loop track (approximately 1200 meters) and four process area sub-loop tracks, respectively connecting the sedimentation tank sub-loop (approximately 300 meters), the filter sub-loop (approximately 450 meters), the pump house sub-loop (approximately 200 meters), and the chemical dosing room sub-loop (approximately 150 meters). The motorized switches include Y-shaped unidirectional branching switches and cross-shaped interchange switches. The Y-shaped switches have two states: a main loop straight position and a sub-loop cut-in position. The cross-shaped interchange switches support independent reversing in four directions.

[0036] Step S2: Establish a four-dimensional comprehensive weight model.

[0037] The composite ring track topology network is abstracted as a weighted directed connected graph. Among them, the node set This includes passive RFID landmark nodes, switch nodes, and equipment inspection points deployed along the entire track; edge collection For the track segments between adjacent nodes; The weight of the overall passage cost for each track.

[0038] arbitrary track edge The overall toll cost is calculated using the following formula: in, All coefficients are greater than 0. The quantification rules for each weighting factor are as follows: Path length factor Linear normalization based on the actual physical distance between two points on the track.

[0039] Passage energy consumption factor The values ​​are assigned according to the characteristics of the track structure: 0.1 for straight tracks, 0.3 for gentle curves, 0.6 for steep slopes / sharp curves, and 0.9 for electric turnout switching sections.

[0040] Obstacle density factor Based on historical perception data, the following grading values ​​are assigned: 0.1 for unobstructed clean corridors, 0.4 for areas with occasional personnel operations, 0.7 for areas prone to obstruction due to long-term water mist, and 1.0 for areas with frequent temporary occupation of the road.

[0041] Historical failure rate factor Based on the historical fault logs of the water plant operation and maintenance platform, the fault-free historical track is 0.1, the occasional communication packet loss section is 0.4, the section where slippage / water accumulation faults have occurred is 0.7, and the high-risk section with multiple faults is 1.0.

[0042] The initial weighting coefficients are set to , , , .

[0043] Step S3: CPP Eulerization.

[0044] Identify weighted directed connected graphs Find all vertices of odd degree in the graph. Using the minimum weighted perfect matching algorithm, calculate the minimum weighted path cost between these odd-degree vertices and add corresponding virtual edges to the original graph, transforming it into an Eulerian graph. For example, if there are two vertices with odd degree in the graph. and The minimum weighted path cost between them is Then in and Add a cost between them Virtual edges.

[0045] Step S4: Improve the Hierholzer hierarchical variant algorithm to generate global closed-loop Euler circuits.

[0046] In Euler The following algorithm is executed: 1. Initialization: Create the stack , set the preset starting point (Charging station RFID node) Push onto the stack. Clear path cache. Mark all track edges as "unvisited".

[0047] 2. Hierarchical Optimal Cyclic Expansion: When If the stack is not empty, pop the top node. Get The set of all unvisited adjacent edges .like If not empty, then: By overall weight Sort in ascending order, and prioritize the edge with the smallest weight.

[0048] Follow the hierarchical rules: if you are currently in the main loop, prioritize the branch path edge leading to the untraversed sub-loop; if you are currently inside the sub-loop, prioritize the closed loop path edge within the sub-loop.

[0049] Under the condition of equal weight, priority should be given to the track side containing RFID inspection nodes to ensure that all equipment points are passed through.

[0050] Select adjacent nodes , will the side Mark as "visited" Push it onto the top of the stack.

[0051] 3. Backtracking: If If empty, will Pop from the stack and store. .

[0052] 4. Generate a closed loop: After the loop ends, Reverse order, to get from Start, traverse all main loop and sub-loop edges and return. The entire closed-loop Euler inspection circuit.

[0053] Step S5: Weighting Find the globally optimal path.

[0054] Using the global closed-loop Eulerian circuit generated in step S4 as a constraint, a weighted average is applied. The algorithm in the original weighted directed connected graph The globally optimal Eulerian circuit is solved to serve as the robot's global reference path. The weighted Dijkstra's distance update formula is: in From the starting point to the node The weighted total cost, For nodes to adjacent nodes The overall weight of the track.

[0055] Step S6: DWA local obstacle avoidance and level switching.

[0056] During the robot's navigation along the global reference path, the Dynamic Window Method (DWA) is used for local trajectory fine-tuning and real-time obstacle avoidance. The evaluation function of DWA is: in To maintain a safe distance between the robot and obstacles on the track, For the smoothing term of driving speed, This represents the offset between the actual trajectory and the global reference path. , , The balance coefficient is selected by the system. Minimum speed As a local control output.

[0057] The conditions for switching levels are as follows: Default state: The global planning layer is always running.

[0058] Conditions for switching to a local layer (switching occurs if any one of these conditions is met): (a) The lidar detects a static or dynamic obstacle within 3 meters of the direction of travel; (b) Water accumulation or slippage on the track causes the robot's attitude to deviate from the preset trajectory threshold by more than 5 degrees; (c) Sudden changes in lighting or increased water mist concentration cause visual positioning deviation to exceed the limit; (d) Received a safety alarm signal for temporary manual operations in the track section.

[0059] Conditions for switching back to the global layer (reset only if all conditions are met): (a) Obstacles completely disappear within the detection range, with no obstruction for a continuous duration. Second; (b) The robot's posture and positioning deviation return to within the allowable threshold range; (c) The obstacle avoidance is completed and the current track section is back to normal traffic conditions.

[0060] Step S6-1: Obstacle classification and graded handling.

[0061] Based on the multi-sensor fusion recognition results, obstacles within the water plant track are divided into three categories and assigned three priority levels: Static fixed obstacles (priority 2): These include maintenance tools, pipe supports, abandoned debris, temporary fencing, and sections of track flooded with water. Identification features include no displacement in point cloud detection for three consecutive rounds, and no personnel features detected by the depth camera. Handling strategy: Initiate local path rerouting and replanning.

[0062] Dynamic slow-moving obstacles (priority 1): These include factory workers, temporary mobile small equipment, and floating debris. Identification features include continuous small-amplitude displacement of the point cloud and human silhouette recognition by a depth camera. Handling strategy: Prioritize slowing down and waiting, while simultaneously uploading a safety alarm.

[0063] High-risk emergency obstacles (priority 0): These include suddenly falling components, rapidly intruding personnel, large-scale water accumulation and landslides, and sudden falling foreign objects from the track. Identification characteristics include obstacles being less than the emergency braking threshold from the robot, or objects intruding into the track at high speed. Handling strategy: Immediately trigger emergency braking, cut off the walking drive output, and lock the track switches.

[0064] Step S7: Dynamic weight update.

[0065] Set up a two-level weight update mechanism: Offline initialization update (daily): Every morning, read the historical database, calculate the obstacle density, slippage frequency, communication packet loss rate, and equipment failure frequency for each track segment, and recalculate. and Refresh the global topology weight table.

[0066] Online real-time updates (milliseconds): During the inspection process, when lidar, attitude sensors, etc. detect anomalies in a local section, the speed of that track segment is immediately increased locally. and Weights. The weight update formula is: in This is the historical smoothing coefficient.

[0067] Coefficient adaptive adjustment: Automatically increases during the rainy season Automatic adjustment in winter .

[0068] In addition, this method includes the following steps that are executed in parallel: Step S8: Multi-source fusion positioning.

[0069] Extended Kalman Filter (EKF) is used for multi-source localization fusion. System state vector. The prediction phase utilizes odometer and IMU data for state recursion; the update phase uses RFID absolute coordinates as observations for correction upon RFID landmark triggering. After fusion, the overall position error is controlled within... Within, the branch road section .

[0070] Step S9: AI visual intelligent detection and linkage.

[0071] It incorporates three lightweight deep learning models: an improved YOLOv8s (object detection), ResNet18+ lightweight FPN (semantic segmentation), and a lightweight ResNet10 (infrared classification). Detection results are categorized into Level I-IV alarms. Level I alarms trigger emergency braking, Level II alarms trigger deceleration cruise and generate a work order, Level III alarms are simply recorded, and Level IV alarms are retained as evidence. Visual detection results directly participate in track weight iteration and replanning triggering.

[0072] Step S10: Multi-robot collaborative scheduling.

[0073] A dynamic task allocation algorithm based on cost-based greedy and load balancing is adopted. Allocation cost. Track access rights are implemented using a switchway mutual exclusion lock + time-sharing occupancy mechanism. Charging scheduling is set with three levels of power thresholds: 35% (early warning), 20% (forced recharging), and 10% (emergency protection).

[0074] Step S11: Closed-loop integrity verification.

[0075] Configure the trigger timing for four levels of verification: real-time point trigger, sub-ring segment trigger, global loop termination trigger, and background timed silent trigger. Verification indicators include: Node traversal completeness , must ; Track edge coverage completeness , must ; Euler loop closed-loop degree , must be between.

[0076] If the verification fails, a supplementary patrol or global replanning will be automatically triggered.

[0077] Example 2: A Closed-Loop Path Planning System for a Water Plant Composite Circular Track Inspection Robot This embodiment provides a closed-loop path planning system for a water plant composite circular track inspection robot, used to implement the closed-loop path planning method for the water plant composite circular track inspection robot described in Embodiment 1. The system includes the following modules, and each module works collaboratively through the communication bus inside the AI ​​group control intelligent control box.

[0078] 1. Track Topology Modeling Module To obtain the topological data of the composite circular track in the water plant, the topological network of the composite circular track is abstracted into a weighted directed connected graph. It also performs four-dimensional comprehensive weight modeling. This module stores the coordinates of all track nodes, edge connection relationships, and weight parameters for the entire plant.

[0079] 2. Hierarchical Closed-Loop Path Planning Module This is used to perform steps S3 to S5 in Example 1. Specifically, it includes: CPP Eulerian Unit: Identifies odd-degree nodes, performs minimum weight perfect matching, and generates an Eulerian graph.

[0080] Improved Hierholzer algorithm unit: Generate a global closed-loop Euler inspection circuit according to the hierarchical constraint of "first fully traversing a single sub-loop and then returning to the main loop, traversing all sub-loops in turn, and finally closing the main loop".

[0081] Weighted Dijkstra Solver Unit: The globally optimal Eulerian circuit obtained by the weighted Dijkstra algorithm is used as the global reference path.

[0082] 3. Dynamic obstacle avoidance and replanning module This is used to perform steps S6 and S6-1 in Example 1. It includes: DWA Local Planning Unit: Samples in velocity space and outputs the optimal velocity command through an evaluation function.

[0083] Layer switching control unit: Automatically switches between global and local layers based on preset quantization thresholds (obstacle distance 3 meters, attitude deviation 5 degrees, etc.).

[0084] Obstacle classification and handling unit: Based on the sensor fusion results, obstacles are classified into three categories: static, dynamic, and high-risk, and the corresponding handling strategies are executed according to priority 0 / 1 / 2.

[0085] Replanning Management Unit: When the local replanning trigger condition is met ( When the DWA local replanning is initiated (e.g., when a new obstacle enters the safe distance), the CPP Eulerization and weighted Dijkstra algorithm are called again when the global replanning trigger conditions are met (e.g., the obstacle persists for more than 10 minutes, topology changes).

[0086] 4. Environmental Perception Fusion Module It consists of three parts: onboard sensors, track-line fixed sensors, and electric switch hardware. (a) Vehicle-mounted sensors A 2D lidar (range 0.1-10m, 360° / 180° selectable) is installed in the center of the front of the robot body.

[0087] Two ultrasonic sensors (range 5-80cm) are symmetrically arranged on the left and right sides to cover the blind spots near the left and right boundaries of the track.

[0088] A depth industrial camera (with infrared illumination) is mounted on the central waterproof bracket, with the lens angled downwards at 45°, simultaneously capturing images of the track surface and workers in the corridor.

[0089] An RFID reader is installed at the bottom, and an RFID marker is pre-embedded on the bottom surface of the track.

[0090] One set of water mist / temperature and humidity composite sensors is fixed on each of the left and right outer shells.

[0091] The internal control compartment integrates a three-axis attitude and tilt sensor.

[0092] (b) Sensors fixed along the track Wall-mounted depth cameras are installed every 10 meters in high-risk sections such as branch roads, filter corridors, and chemical dosing rooms.

[0093] Ultrasonic water accumulation detection modules are installed on the sides of the tracks in areas prone to water accumulation, such as pump rooms and filter beds.

[0094] An RFID marker is added to each turnout location for ring switching positioning calibration.

[0095] (c) Hardware structure of electric switch All electric turnouts are equipped with industrial waterproof electric switch units, directly driven by the relay terminals of the turnout linkage output compartment of the group control AI control box. The entire mechanical structure is IP65 waterproof. The main mechanical component includes a movable core rail, a guide arm, a return spring, and a mechanical position limit block; the arm drives the core rail to switch left and right travel directions, and locks mechanically when in position.

[0096] Drive unit: DC geared waterproof electric push rod with built-in Hall position sensor; the push rod extends / retracts to correspond to two passage tracks, and outputs a switch feedback signal when in position.

[0097] Positioning and sensing components: Encrypted RFID markers are pre-embedded 1m before and after the switch, and the switch status is double-verified in conjunction with the built-in position sensor of the switch.

[0098] Electrical connection method: The switch drive control line and position feedback signal line are uniformly connected to the lower-level linkage control interface group of the control box; the relay output compartment inside the control box provides passive switching quantity, outputs high level to drive the push rod to move, and at the same time collects the position feedback signal to form a closed loop control.

[0099] (d) Positioning fusion unit Extended Kalman filter (EKF) is used to fuse RFID absolute coordinates, odometer relative displacement, and IMU attitude data to output high-precision location information. The RFID landmark deployment density is as follows: 3m spacing in conventional straight track sections, 1m spacing at intersections, turns, and key inspection points, and 0.8m spacing in charging station areas.

[0100] 5. AI Visual Intelligent Inspection Module Equipped with three lightweight deep learning models, inference is performed on the local NPU of the AI ​​control box: (a) Model Architecture An improved YOLOv8s object detection network: The backbone uses a lightweight C2f-Lite module, introduces deformable convolutional DCN and multi-scale FPN feature fusion layers, and includes a pre-processing module for adaptive grayscale equalization. It is used to detect personnel violations (missing safety helmets, reflective clothing, or safety shoes), foreign objects on equipment, and floating debris on the pool surface.

[0101] ResNet18+ lightweight FPN semantic segmentation network: used for pixel-level segmentation of areas such as water seepage marks in pipes, cracks in pool walls, and clumps on the pool surface.

[0102] Lightweight ResNet10 infrared thermal imaging classification network: used for identifying overheating defects in equipment such as motors and terminal blocks.

[0103] (b) Source and annotation method of training dataset The datasets used for training the above models came from three sources: 1. On-site data collection at the plant site (70%): Based on the implementation of this project at multiple waterworks and water purification plants, the composite track inspection sites were used to take on-site photos using a track robot with a depth camera and an infrared thermal imager, covering the entire scene of sedimentation tanks, filters, pump rooms, dosing rooms, and corridors. The photos included images under various working conditions such as sunny days, heavy fog, rainy season water accumulation, nighttime supplemental lighting, and backlighting and reflection, with a total of approximately 120,000 images collected.

[0104] 2. Publicly available industry open-source datasets (accounting for 20%): including publicly available water equipment defect datasets, publicly available PPE safety protection identification datasets for industrial plants, and abnormal image datasets of sewage treatment ponds, used to expand rare defect samples.

[0105] 3. Image enhancement to generate expanded samples (accounting for 10%): Online enhancement algorithms such as Mosaic, random fog overlay, raindrop simulation, and brightness contrast perturbation are used, while GAN is used to generate simulated samples of water mist, water accumulation, and low light.

[0106] Annotation methods: For object detection, the LabelImg tool is used to output VOC / YOLO standard bounding box annotation files; for semantic segmentation, the LabelMe tool is used to output pixel-level polygon mask JSON annotations; for infrared classification, the entire thermal image is directly annotated with binary classification folders. All images are uniformly scaled to [size missing]. The standard input size is used to divide the training set, validation set, and test set in a 7:2:1 ratio.

[0107] (c) Procedure for handling test results The detection results are output as a structured JSON message, containing information such as alarm level (I-IV), defect type, confidence level, and location coordinates. Level I alarms (red) trigger emergency braking, turnout locking, and on-site audible and visual alarms; Level II alarms (orange) trigger deceleration cruise and generate a work order, requiring verification within 2 hours; Level III alarms (yellow) are only recorded, increasing inspection frequency; Level IV alarms (blue) are retained as evidence. The detection results directly participate in track weight updates and replanning triggers.

[0108] 6. Online Weight Update Module This is used to perform step S7 in Example 1. It includes: Offline update unit: Reads the historical database every morning and recalculates. and Refresh the global weight table.

[0109] Online update unit: Receives sensor data in real time during inspection and adjusts the values ​​of abnormal sections accordingly. and .

[0110] Coefficient adaptive adjustment unit: automatically adjusts according to the season. and .

[0111] 7. Multi-robot collaborative scheduling module This is used to perform step S10 in Example 1. It includes: Task allocation unit: A dynamic task allocation algorithm based on cost greed and load balancing is adopted to allocate inspection tasks and re-inspection tasks to multiple robots.

[0112] Track access right scheduling unit: Each branch node and intersection sub-ring is regarded as a mutually exclusive resource. The access right application-response-release mechanism is adopted, and only one robot is allowed to occupy the same track segment at the same time.

[0113] Charging management unit: Sets three power thresholds (35% / 20% / 10%), and automatically selects the optimal station for recharging based on the weighted distance of the robot from the charging station and the number of people queuing at the charging station.

[0114] 8. Closed-loop integrity verification module To perform step S11 in Example 1. Includes: Verification trigger unit: Set up a four-level triggering mechanism: real-time point triggering, sub-ring segment triggering, global loop termination triggering, and background timed silent triggering.

[0115] Verification calculation unit: Calculate the node traversal completeness rate Track edge coverage integrity R_E, Euler loop closure degree .

[0116] Supplementary patrol scheduling unit: When the verification fails, it automatically triggers local backtracking supplementary patrol and local supplementary patrol based on the missing type. Sub-path replanning or global CPP replanning.

[0117] Example 3: The entire closed-loop inspection process of a water plant This embodiment demonstrates the complete workflow of the system and method of the present invention using a specific application scenario.

[0118] A waterworks deployed four inspection robots (numbered R1-R4). The composite circular track configuration consisted of one main ring (1200m) and four sub-rings (sedimentation tank sub-ring 300m, filter sub-ring 450m, pump house sub-ring 200m, and chemical dosing room sub-ring 150m). Approximately 800 RFID tags were deployed along the track, with a spacing of 1m in the densely spaced sections and 3m in the regular sections.

[0119] Phase 1: System Initialization and Global Path Generation Every day at 2:00 AM, the AI-controlled group control box reads historical data from the water plant's operation and maintenance platform and performs offline weight updates. It then calls the CPP Eulerianization module to process the entire plant's trajectory graph, identifying four odd-degree nodes. By adding two virtual edges through minimum-weight perfect matching, the graph is transformed into an Eulerian graph. Next, an improved Hierholzer hierarchical variant algorithm is executed to generate a path from the charging station... Start, sequentially traverse the main loop → filter sub-loop → main loop → sedimentation tank loop → main loop → pump house loop → main loop → chemical dosing room sub-loop → main loop → return. A fully closed-loop Euler circuit, with a total length of approximately 2300m. Weighted Based on this, the globally optimal path is calculated and distributed to all robots.

[0120] Phase Two: Robot R1 performs sub-loop inspection of the filter tank. At 9:00 AM, robot R1 began its first full-area inspection of the day. When it reached 1 meter before the Y-shaped junction S1 connecting the main ring and the filter sub-ring, the RFID reader read the encrypted landmark, and R1 sent a junction access request to the control box. The control box checked the S1 resource lock status and found it to be idle, issued authorization, locked S1, and simultaneously drove the electric push rod to switch the core rail to the sub-ring entry position. After the Hall sensor returned a positioning signal, the control box released the request, and R1 entered the filter sub-ring.

[0121] Phase 3: AI Visual Inspection and Collaborative Processing When R1 travels to RFID marker point number 27 on the filter sub-loop, the depth industrial camera detects an abnormal flange leakage at the outlet valve during local inference in the NPU, with a detection confidence level of 0.92. The AI ​​control box determines this as a Level II critical alarm (orange) and executes the following linked actions: 1. Reduce speed to 50% of standard speed (from 0.8m / s to 0.4m / s); 2. Obstacle density factor of the current track segment An upward movement of 0.3 triggers a local DWA replanning. 3. Capture 2 high-definition images + 5 seconds of short video, store locally and upload to the cloud to generate a defect work order; 4. After R1 completes the inspection of the current sub-ring, a fixed-point re-inspection sub-task is automatically generated, which returns to the fault location for multi-angle secondary imaging confirmation.

[0122] Phase Four: Obstacle Avoidance and Level Switching R1 continued traveling along the filter sub-loop. The lidar detected a worker (dynamic obstacle, priority 1) without reflective clothing within 3 meters ahead. The system immediately switched from the global planning layer to the local DWA layer, the robot decelerated to 0.2 m / s and issued a voice prompt. Simultaneously, the depth camera recognized the worker's lack of reflective clothing, triggering a Level II alarm, and captured and uploaded an image. After the worker left, and the unobstructed period reached 2 seconds, the system automatically switched back to the global planning layer, and R1 resumed its cruising speed.

[0123] Phase 5: Multi-robot collaboration and charging scheduling At this time, robot R2 is performing a ring inspection around the pump house, R3 is performing a ring inspection around the sedimentation tank, and R4 is on standby at the charging station. After R1 completes its ring inspection around the filter tank, its battery level drops to 38% (below the warning threshold of 35%), and the control box activates the charging cost formula. The system calculates that the nearest charging station, c1, has a queue count of 0, minimizing the cost. Therefore, it schedules R1 to go to c1 for recharging. At the same time, the control box assigns the filter sub-ring re-inspection task to the standby R4, and R4 enters the filter sub-ring along the global path to perform the re-inspection.

[0124] Phase Six: Closed-Loop Integrity Verification and Closure After R1 completed charging, it continued with the remaining inspection tasks. At 3:00 PM, R1 completed the final sub-loop inspection of the entire plant and returned to its starting point. This triggers a global loop termination check. The control box calculates the node traversal completeness rate. (All 800 RFID markers were read), trackside coverage completeness rate (Travel records are available for all track sections), Euler loop closed-loop degree (exist (Within the specified range). Verification passed, and this full-area closed-loop inspection was successfully completed. The control box generated a daily inspection coverage report and uploaded it to the smart water plant operation and maintenance platform.

[0125] The above three embodiments fully disclose the method, system and application process of the present invention. Those skilled in the art can implement the technical solution of the present invention based on the above description to realize intelligent closed-loop inspection of the water plant composite circular track with no blind spots, no deadlock and high adaptability.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A closed-loop path planning method for a water plant composite circular track inspection robot, characterized in that, Includes the following steps: S1. Obtain the topology data of the composite ring track of the water plant, wherein the composite ring track includes a main ring track and at least one sub-ring track, and the main ring track and the sub-ring track are connected by electric switches to form a connectable composite ring topology network. S2. Abstract the composite ring track topology network into a weighted directed connected graph. ,in For a set of nodes, For the set of orbital edges, To calculate the comprehensive passage cost weights, a four-dimensional comprehensive weight model is performed on the weighted directed connected graph, wherein the four-dimensional comprehensive weights include a path length factor. Energy consumption factor for passage Obstacle density factor and historical failure rate factor The track edge The overall cost of passage is: in , , , The normalized balance coefficients and ; The path length factor is obtained by linearly normalizing the actual physical distance between two points on the track. The passage energy consumption factor is assigned in a graded manner based on the characteristics of the track structure. The obstacle density factor is assigned based on historical perception data. A historical failure rate factor that is quantitatively assigned based on historical failure logs from the water plant operation and maintenance platform; S3. Based on the weighted directed connected graph, the Chinese Postman Problem (CPP) Eulerianization method is used to perform minimum weight matching of odd-degree nodes on the non-Eulerian orbit graph to complete the Eulerianization, transforming the original graph into an Eulerian graph; the minimum weight matching uses a weighted method. The algorithm calculates the minimum weighted path cost between odd-degree nodes; S4. An improved Hierholzer hierarchical variant algorithm is used to perform hierarchical optimal traversal on the Eulerian graph to generate a global closed-loop Eulerian inspection loop that starts from a preset starting point, traverses all main loops and sub-loops of the composite circular track, and returns to the starting point. This loop is used as a global reference path. The improved Hierholzer hierarchical variant algorithm follows the hierarchical traversal constraint of "first fully traversing a single sub-loop and then returning to the main loop, traversing all sub-loops in sequence, and finally closing the main loop". S5. During the robot's cruise along the global reference path, the Dynamic Window Method (DWA) is used for local trajectory fine-tuning and real-time obstacle avoidance. The operating mode is automatically switched between the global planning layer and the local planning layer according to the preset level switching conditions. S6. Real-time acquisition of environmental perception data and robot status data during the inspection process, and dynamic updating of the obstacle density factor in the four-dimensional comprehensive weight. and historical failure rate factor This enables online iterative updates of orbital weights.

2. The closed-loop path planning method for a water plant composite circular track inspection robot according to claim 1, characterized in that, The dynamic update mentioned in step S6 includes: Offline initialization update: During low-load periods each day, read historical database statistics on obstacle density, slippage frequency, communication packet loss rate, and equipment failure frequency for each track segment, and recalculate. and The quantized value is used to refresh the global topology weight table; Online real-time updates: When an anomaly is detected in a local section during the inspection, the speed of that track section is immediately increased locally. and The weights are modified only in the local weights of the current sub-ring. Coefficient adaptive adjustment: Automatically increases during the rainy season. The obstacle weighting coefficient is automatically increased when the track is prone to slippage in low temperatures during winter. Failure rate weighting coefficient.

3. The closed-loop path planning method for a water plant composite circular track inspection robot according to claim 1, characterized in that, The hierarchical switching conditions mentioned in step S5 include: The trigger condition for switching from the global planning layer to the local dynamic planning layer is that any of the following conditions are met: a) The lidar detected a static or dynamic obstacle within a 3-meter radius of the direction of travel; b) Water accumulation or slippage on the track causes the robot's posture to deviate from the preset trajectory threshold by more than 5 degrees; c) Sudden changes in lighting or increased water mist concentration cause visual positioning deviation to exceed the limit; d) Received a safety alarm signal for temporary manual operations in the track section; The reset condition for switching back from the local dynamic programming layer to the global planning layer is that all of the following conditions are met: a) Obstacles completely disappear within the detection range, and the duration of continuous unobstructed access. Second; b) The robot's posture and positioning deviation have returned to within the allowable threshold range; c) The obstacle avoidance is complete, and the current track section's traffic status has returned to normal.

4. The closed-loop path planning method for a water plant composite circular track inspection robot according to claim 1, characterized in that, It also includes obstacle classification and graded handling steps: the detected obstacles are divided into three categories: static fixed obstacles, dynamic slow obstacles, and high-risk emergency obstacles; and three priority handling rules are set: priority 0 corresponds to high-risk emergency obstacles, triggering immediate emergency braking; priority 1 corresponds to dynamic personnel obstacles, prioritizing deceleration and waiting while simultaneously uploading safety alarms; and priority 2 corresponds to static road-occupying obstacles, initiating local path detour replanning.

5. The closed-loop path planning method according to claim 1, characterized in that, It also includes a hierarchical path replanning step: When an obstacle is detected to enter the preset safe distance threshold, the trajectory deviates laterally from the global reference path by more than 0.4 meters, the vehicle attitude deviates by more than 5 degrees, or the environmental conditions change abruptly or the track weights change dynamically, the DWA algorithm is used to perform local incremental replanning. When a static obstacle persists for more than 10 minutes, the track topology changes, there is a periodic update, or the scheduling strategy is switched, the CPP Eulerization and weighted Dijkstra algorithm are re-executed to perform a global complete replanning.

6. The closed-loop path planning method for a water plant composite circular track inspection robot according to claim 1, characterized in that, It also includes a step of multi-source positioning data fusion using an extended Kalman filter (EKF): using the global absolute node coordinates of the track output when the RFID landmark is triggered as the global correction reference, using the relative displacement of the track travel output by the wheel odometer as the continuous position prediction input, and using the three-axis angular velocity and acceleration output by the IMU inertial measurement unit for attitude compensation; after fusion, the position error is controlled within a certain range throughout the entire process. Within this range, the positioning error of the turnout section .

7. The closed-loop path planning method for a water plant composite circular track inspection robot according to claim 1, characterized in that, It also includes a closed-loop integrity verification step: Set the trigger timing for four levels of verification: real-time point trigger, sub-ring segment trigger, global loop termination trigger, and background timed silent trigger; Calculate the node traversal completeness rate ;in This represents the number of RFID nodes or inspection points visited in this round. This represents the total number of nodes in the plant's topology. Calculate trackside coverage integrity rate ;in This represents the number of track edges that have been traversed in this round. This represents the total number of all physical track sections in the entire plant. Calculate the closed-loop degree of Euler circuits ;in This represents the actual closed-loop mileage traveled by the robot. This represents the total mileage of the theoretical CPP Euler circuit; when or or Not here When within the specified range, automatically trigger local backtracking and re-inspection based on the missing type. Sub-path replanning or global CPP replanning.

8. The closed-loop path planning method for a water plant composite circular track inspection robot according to claim 1, characterized in that, It also includes multi-robot collaborative scheduling steps: A dynamic task allocation algorithm based on cost greed and load balancing is adopted to distribute inspection tasks to multiple robots, allocating costs accordingly. ,in For the first Taiwan robot, For the first One inspection task For robots To the mission The orbital weighted distance, For robots Execute the task The energy cost, The weight of the robot's current remaining battery power. This represents the current number of tasks the robot is currently handling. , , , These are the normalized balance coefficients; A switch track mutual exclusion lock and time-sharing occupation mechanism are adopted to allow only one robot to occupy the same track segment at any given time. Three levels of power thresholds are set for charging scheduling: a warning threshold of 35%, a forced return-to-charge threshold of 20%, and an emergency protection threshold of 10%. The robot automatically selects the station with the lowest charging cost for return-to-charge based on the weighted distance of the robot from the charging station and the current number of people waiting in the queue at the charging station.

9. A closed-loop path planning system for a water plant composite circular track inspection robot, characterized in that, include: The track topology modeling module is used to acquire the topological structure data of the composite circular track in the water plant, and abstract the composite circular track topology network into a weighted directed connected graph. A four-dimensional comprehensive weight model is then performed, which includes a path length factor. Energy consumption factor for passage Obstacle density factor and historical failure rate factor ; The hierarchical closed-loop path planning module is used to Eulerianize the non-Eulerian trajectory graph based on the weighted directed connected graph using the CPP Eulerianization method, generate a global closed-loop Eulerian inspection loop using an improved Hierholzer hierarchical variant algorithm, and then use a weighted... The algorithm finds the globally optimal Eulerian circuit as the global reference path; The dynamic obstacle avoidance and replanning module is used to perform local trajectory fine-tuning and real-time obstacle avoidance using DWA, and automatically switches the operating mode according to preset level switching conditions. The environmental perception fusion module includes vehicle-mounted sensors and track-fixed sensors. The vehicle-mounted sensors include 2D LiDAR, ultrasonic sensors, depth industrial cameras, RFID readers, water mist / temperature and humidity composite sensors, and three-axis attitude and tilt sensors. The track-fixed sensors include wall-mounted depth cameras, water accumulation detection ultrasonic modules, and RFID landmarks. The AI ​​visual intelligent inspection module is equipped with an improved YOLOv8s lightweight target detection network, a ResNet18+ lightweight FPN semantic segmentation network, and a lightweight ResNet10 infrared thermal imaging classification network. It is used to perform equipment appearance defect detection, pool surface anomaly detection, and on-site personnel safety violation detection. The online weight update module is used to dynamically update the obstacle density factor in the four-dimensional comprehensive weight based on real-time collected environmental perception data and robot state data. and historical failure rate factor ; The multi-robot collaborative scheduling module is used to allocate inspection tasks to multiple robots using a dynamic task allocation algorithm based on cost greed and load balancing, and to coordinate track access rights using a branch track mutual exclusion lock and time-sharing occupancy mechanism. The closed-loop integrity verification module is used to set up a four-level verification trigger mechanism, calculate the node traversal integrity rate, track edge coverage integrity rate and Eulerian circuit closure degree, and automatically trigger supplementary inspection or global replanning when the verification fails.

10. The closed-loop path planning system for a water plant composite circular track inspection robot according to claim 9, characterized in that, The backbone network of the improved YOLOv8s lightweight target detection network uses a lightweight C2f-Lite module to replace the standard C2f module, and introduces deformable convolutional DCN and multi-scale FPN feature fusion layers, as well as a pre-processing module for adaptive grayscale equalization. The electric switch configured in the composite circular track adopts an industrial waterproof electric switch, which includes a mechanical body composed of a movable core rail, guide arm, return spring and mechanical position limit block, and a DC deceleration waterproof electric push rod with built-in Hall position sensor as the drive unit. The encrypted RFID markers are pre-embedded 1 meter before and after the switch and used with the built-in position sensor of the switch for double verification.

Images