A path planning method for an all-terrain autonomous inspection robot for high-voltage transmission and distribution lines

Abstract

This paper discloses a path planning method for an all-terrain autonomous inspection robot for high-voltage transmission and distribution lines. This method constructs a comprehensive risk map reflecting the "dynamic risk" distribution within the distribution lines by combining exponential decay superposition with Markov prediction, providing accurate risk information for subsequent planning. It then uses a heuristic greedy-genetic algorithm hybrid strategy to iteratively optimize the task sequence and output the optimal inspection scheduling sequence. Finally, an adaptive task feedback mechanism based on a multimodal perception fusion model and Bayesian state estimation continuously and dynamically corrects the optimal inspection scheduling sequence, significantly improving inspection efficiency, accuracy, and reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of inspection robots, and in particular to a path planning method for an all-terrain autonomous inspection robot for high-voltage power transmission and distribution lines. Background Art

[0002] Distribution lines are key electrical infrastructure within urban power distribution systems, fulfilling crucial functions such as voltage transformation, power distribution, and protection. To ensure stable power system operation, equipment within these lines requires regular inspections. Routine checks include temperature, voltage, current, equipment operating status indicators, and instrument readings.

[0003] Traditionally, power distribution line inspections rely on manual patrols. Due to the confined spaces, complex passageways, and densely packed equipment, manual inspections are inefficient and pose safety risks during operations at night, during high temperatures, and during thunderstorms. Furthermore, manual recording of equipment information is subject to high subjectivity, a high rate of missed inspections, and difficulty tracing data, making it difficult to meet the demands of refined, data-based, and automated management.

[0004] In view of this, intelligent inspection robots are gradually being applied to the field of distribution line inspection. Currently, there are two main technical routes for intelligent inspection robots:

[0005] (1) Fixed-path navigation based on laser SLAM or QR code positioning. Laser SLAM navigation uses laser radar to scan and construct an internal environment map of the distribution line for self-positioning and navigation. The robot usually patrols along a preset fixed path and reaches the target equipment points along the way to collect images or detect status. QR code positioning, on the other hand, is a robot that scans pre-placed QR code points for positioning and navigation, and performs equipment detection and image collection along a preset route. The above navigation method completes various tasks in a fixed order and cannot be flexibly adjusted according to the equipment fault priority, the urgency of the inspection task, or environmental changes. This scheduling method results in low inspection efficiency and is unable to respond to emergencies or conduct key inspections of certain equipment in a short period of time.

[0006] (2) Dynamic path navigation integrating visual SLAM and multi-sensor fusion. The robot obtains environmental data of the distribution line through multiple sensors such as lidar, camera, and IMU, and uses visual SLAM technology for map construction and self-localization. In this process, visual SLAM uses the camera to capture environmental image information and estimates the pose through image feature matching; at the same time, the lidar provides distance data, and the IMU helps eliminate errors caused by posture changes. However, this dynamic path navigation solution requires large computing resources and poor real-time performance. This technology relies on the data fusion of multiple sensors such as visual SLAM, lidar, and IMU, and processing this data requires strong computing power. Summary of the Invention

[0007] In response to the above problems, the present invention proposes a path planning method for an all-terrain autonomous inspection robot for high-voltage transmission and distribution lines, which mainly solves at least one problem raised by the background technology.

[0008] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0009] A path planning method for an all-terrain autonomous inspection robot for high-voltage transmission and distribution lines, comprising:

[0010] Based on the static prior map, an immediate risk assessment is performed on static obstacles within the target distribution line, and the risk of dynamic obstacles is predicted using a Markov chain. These are defined as immediate risk and predicted risk, respectively. The immediate risk and predicted risk are superimposed through exponential decay to generate a comprehensive risk map.

[0011] A model graph is generated based on the target distribution line. Based on this model graph, a risk-aware A* search algorithm is used as the backbone path for global path planning. A deep Q network is then introduced to fine-tune local paths in real time. Finally, a heuristic greedy-genetic algorithm hybrid strategy is used to iteratively optimize the task sequence, taking into account task priorities and time window constraints to output the optimal inspection scheduling sequence.

[0012] The robot performs inspection tasks according to the comprehensive risk map and the constraints of the optimal inspection scheduling sequence. During the inspection tasks, the optimal inspection scheduling sequence is continuously and dynamically corrected by an adaptive task feedback mechanism based on a multimodal perception fusion model and Bayesian state estimation.

[0013] In some embodiments, constructing the static a priori risk map includes:

[0014] Use indoor SLAM to perform offline mapping of the interior of the target distribution line and obtain a grid map of the clear area of ​​the distribution line;

[0015] On the grid map, walls, cabinet edges, and fixed equipment are marked as high-risk areas to form the static a priori risk map.

[0016] In some embodiments, the instant risk assessment process includes:

[0017] The laser radar and camera are called to collect obstacle distance information in each sensing cycle, and the obstacle distance information is mapped to the grid points of the grid map for dynamic risk perception to obtain the current instantaneous risk value of the robot. The instantaneous risk value is inserted into the corresponding position in the static prior risk map to form a time-varying risk map.

[0018] In some embodiments, the instant risk value is calculated as follows:

[0019] ;

[0020] Where d is the distance information of the obstacle, x and y are the current horizontal and vertical coordinates of the robot, respectively, and λ d is the instantaneous attenuation rate corresponding to the obstacle distance information d. When d approaches 0, p approaches 1, and as the distance increases, the risk decays rapidly in an exponential form.

[0021] In some embodiments, the Markov chain prediction process includes:

[0022] For each grid point in the time-varying risk map, a first-order Markov chain is established, and a transfer matrix is ​​used to describe the risk transfer probability and risk distribution vector of the robot between adjacent grids. The predicted risk value is calculated based on the risk transfer probability and the risk distribution vector.

[0023] In some embodiments, the scheduling objective function of the heuristic greedy-genetic algorithm hybrid strategy is:

[0024] ;

[0025] Where, is the delay penalty weight coefficient of the kth sensor cycle, is the delay of the kth sensor cycle, d(v ti ,v ti+1 ) is the path length between adjacent task points, θ is the path length weight, and m is the total number of tasks.

[0026] In some embodiments, the process of constructing the multimodal perception fusion model includes:

[0027] Let the observation of the i-th sensor at time t be z i (t), then the total observation vector is:

[0028] ;

[0029] Construct each state S k Conditional probability in observation space:

[0030] ;

[0031] Among them, P(Z|S k ) is obtained by fitting the Bayesian network with the training data, where k is the index number of the state and j is the index number of the traversal.

[0032] The beneficial effects of the present invention are: by combining exponential decay superposition with Markov prediction, a comprehensive risk map reflecting the distribution of "dynamic risks" within the distribution line is constructed to provide accurate risk information for subsequent planning, and then a heuristic greedy-genetic algorithm hybrid strategy is used to iteratively optimize the task sequence and output the optimal inspection scheduling sequence. Finally, an adaptive task feedback mechanism based on a multimodal perception fusion model and Bayesian state estimation is used to continuously and dynamically correct the optimal inspection scheduling sequence, thereby greatly improving inspection efficiency, accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of a path planning method for an all-terrain autonomous inspection robot for high-voltage transmission and distribution lines disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] To make the objectives, technical solutions, and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all of the present invention.

[0035] This embodiment proposes a path planning method for an all-terrain autonomous inspection robot for high-voltage transmission and distribution lines. Figure 1 As shown, including:

[0036] In step 1, based on the static prior map, an immediate risk assessment of static obstacles inside the target distribution line is performed, and the risk of dynamic obstacles is predicted using a Markov chain. These are defined as immediate risk and predicted risk, respectively. The immediate risk and predicted risk are superimposed by exponential decay to generate a comprehensive risk map.

[0037] Step 101, constructing a static priori risk map includes:

[0038] Use indoor SLAM to perform offline mapping of the interior of the target distribution line and obtain a grid map of the clear area of ​​the distribution line;

[0039] On the grid map, walls, cabinet edges, and fixed equipment are marked as high-risk areas to form a static a priori risk map. The static a priori risk map is defined as:

[0040] .

[0041] Step 102, the process of immediate risk assessment includes:

[0042] The laser radar and camera are used to collect obstacle distance information in each sensing cycle, and the obstacle distance information is mapped to the grid points of the grid map for dynamic risk perception. The robot's current instantaneous risk value is obtained and inserted into the corresponding position in the static prior risk map to form a time-varying risk map. In this solution, the instantaneous risk value is calculated as follows:

[0043] ;

[0044] Where d is the obstacle distance information, x and y are the current horizontal and vertical coordinates of the robot, respectively, and λ d is the instantaneous attenuation rate corresponding to the obstacle distance information d. When d approaches 0, p approaches 1, and as the distance increases, the risk decays rapidly in an exponential form. d Parameters can be adjusted according to laser ranging accuracy and environmental density.

[0045] Step 103, the Markov chain prediction process includes:

[0046] For each grid point in the time-varying risk map, a first-order Markov chain is established, and the transfer matrix is ​​used to describe the risk transfer probability and risk distribution vector of the robot between adjacent grids. The predicted risk value is calculated based on the risk transfer probability and risk distribution vector.

[0047] Specifically:

[0048] (1) Transition probability estimation

[0049] The number of times the transition from grid point j to i was observed in the past T time steps is n ij ,but

[0050] .

[0051] (2) Multi-step prediction

[0052] Use matrix power operation to predict the risk distribution vector p after τ steps t :

[0053] .

[0054] (3) Single-step prediction expansion

[0055] For a single grid i: .

[0056] Step 104, exponential decay superposition update:

[0057] The immediate risk value and the predicted risk value obtained in steps 102 and 103 are added in proportion to balance the "current observation" and the "historical trend":

[0058] ;

[0059] α is the forgetting factor, which is closer to 1 and more dependent on the current observation; the cumulative prediction term Reflect short-term dynamic trends.

[0060] Step 105: Comprehensive risk map formation:

[0061] The final comprehensive risk map M(x, y, t) is the point-by-point maximum of the static prior risk map and the updated dynamic risk after step 104:

[0062] ;

[0063] This ensures that the a priori high-risk areas are always retained, while introducing new risks brought by dynamic obstacles; M(x,y,t)∈[0,1] can be directly used as the risk penalty weight in the planning cost.

[0064] The total number of grids M in a static priori risk map is usually in the order of thousands to tens of thousands. By using the transfer matrix and neighborhood risk reduction pruning, efficiency can be greatly improved, ensuring that the full map risk update is completed within every 100 ms.

[0065] Steps 101-105 generate a time-varying risk weight M(x, y, t) by exponentially superimposing the real-time obstacle distance information of the lidar / camera and combining it with a first-order Markov chain short-term prediction. This accurately reflects the "dynamic risk" distribution within the distribution line, reducing the risk exposure rate by approximately 20% compared to solutions that rely solely on static maps.

[0066] In step 2, a model graph is generated according to the target distribution line. Based on the model graph, a risk-aware A* search algorithm is used as the backbone path for global path planning. Then, a deep Q network is introduced to fine-tune the local path in real time. Finally, a heuristic greedy-genetic algorithm hybrid strategy is used to iteratively optimize the task sequence, comprehensively considering the task priority and time window constraints, and outputting the optimal inspection scheduling sequence.

[0067] Step 201: Model diagram construction:

[0068] The distribution line space is discretized into a topological graph G = (V, E), where:

[0069] V={v1,v2,…,vn}: represents each grid center or point of interest;

[0070] E={(vi,vj)}: If there is no obstacle between two points and they are passable, then an edge is established. The edge weight is defined as:

[0071] ;

[0072] Where, dij is the Euclidean distance, is the average risk based on the dynamic risk graph in the previous section, and λ is the risk weight factor.

[0073] Step 202, shortest path search: improved A* search algorithm:

[0074] Use the risk-aware A* search algorithm as the backbone for global path planning:

[0075] Heuristic function definition:

[0076] ;

[0077] Where β is the spatial heuristic coefficient, γ is the risk heuristic coefficient, and M(v) is the comprehensive risk value of the current point.

[0078] Compared with the traditional A* which only uses distance as a reference, step 202 introduces environmental risks into the path cost estimation, which can effectively avoid high-risk obstacle areas.

[0079] Step 203, local obstacle avoidance and path correction: Deep Q Network (DQN):

[0080] When an AGV encounters an unexpected obstacle (such as a moving person or a door closing suddenly) during its path execution, the global path may become unavailable. Therefore, DQN is introduced to fine-tune the local path in real time:

[0081] (1) State space S: includes the AGV’s current position, target direction, and laser point cloud scan;

[0082] (2) Action space A = {move left, move forward, move right, stop};

[0083] (3) Reward function design:

[0084] ;

[0085] (4) Update the Q value formula:

[0086] .

[0087] The AGV learns the optimal action strategy in real time to avoid obstacles and maintain the overall path direction stability.

[0088] Based on the comprehensive risk map obtained in step 1, steps 201-203 introduce risk-weighted optimized A* global search and deep Q network local fine-tuning to keep the inspection path as short as possible while avoiding high-risk areas. In actual tests, compared with the pure shortest distance algorithm, the obstacle avoidance success rate increased by 25%.

[0089] Step 204: multi-task scheduling and priority optimization:

[0090] Assume that the inspection task set is T={t1,t2,…,tm}, each task contains the following parameters: location node v tk , priority π k , feasible time window[e k ,l k ].

[0091] Finally, a heuristic greedy-genetic algorithm hybrid strategy is used to iteratively optimize the task sequence, comprehensively considering the priority and time window constraints, and output the optimal inspection scheduling sequence.

[0092] Among them, the scheduling objective function of the heuristic greedy-genetic algorithm hybrid strategy is:

[0093] ;

[0094] Where, is the delay penalty weight coefficient of the kth sensor cycle, is the delay of the kth sensor cycle, where ( represents the actual arrival time of task k, represents the expected completion time threshold of task k), d(v ti ,v ti+1 ) is the path length between adjacent task points (the i-th task point and the i+1-th task point, where i represents the position in the task execution sequence), θ is the path length weight, and m is the total number of tasks.

[0095] In step 204, a heuristic greedy-genetic algorithm hybrid strategy is used to construct the scheduling objective function. The inspection point sequence is iteratively optimized while taking into account the task priority π_k and the time window [e_k, l_k]. Experiments show that the total trip load is shortened by an average of 15% and the task response delay is reduced by 18%.

[0096] In step 3, the robot performs inspection tasks according to the comprehensive risk map and the constraints of the optimal inspection scheduling sequence. During the inspection task, the adaptive task feedback mechanism based on the multimodal perception fusion model and Bayesian state estimation continuously and dynamically corrects the optimal inspection scheduling sequence.

[0097] Step 301, the process of constructing a multimodal perception fusion model includes:

[0098] AGV is equipped with multiple sensor modules (IMU, lidar, visual camera, ambient temperature and humidity sensor, etc.), so let the observation of the i-th sensor at time t be z i (t), then the total observation vector is:

[0099] ;

[0100] Construct each state S k (e.g., normal, deviation, stuck, completed, etc. task status) The conditional probability in the observation space is:

[0101] ;

[0102] in, Obtained by fitting a Bayesian network to the training data (or using a Gaussian mixture model (GMM)), Indicates that at time t in state S k The conditional probability of observing Z(t) is obtained by training samples through the Bayesian network The fitting is obtained, where is the network parameter vector, k is the index number of the state, and j is the index number of the traversal.

[0103] Step 302, state estimation and self-correction mechanism (improved Kalman filter):

[0104] State transfer uses linear prediction: ;

[0105] Observation Update: ;

[0106] in, is the Kalman gain, x^t is the state estimate (position / bias / error), zt is the current observation of the sensor, and R is the measurement noise covariance.

[0107] In this way, the AGV status can be continuously and dynamically predicted and errors corrected, enabling self-adjustment in the event of task interruption or missed inspection.

[0108] Step 303, task completion and information entropy determination mechanism:

[0109] Information entropy H is introduced to judge the "credibility" of the current task completion. The lower the entropy, the clearer the status. The formula is as follows:

[0110] ;

[0111] If the confidence level of a status (such as "task completed") And the entropy value is less than the set threshold , the task is considered completed:

[0112] ;

[0113] Otherwise, feedback is abnormal or continue monitoring.

[0114] Step 304: abnormal feedback and scheduling linkage:

[0115] When the system determines that the status is abnormal (such as path deviation, task failure, perception failure, etc.), it immediately performs the following operations:

[0116] Report the state flag Sk and current observation Z(t) to the main control and scheduling system;

[0117] The scheduling system replans the task path based on the task priority table and the currently available paths;

[0118] If there are continuous abnormalities within a short period of time, the system will mark the area as a "high-risk area" and subsequent tasks will be avoided or scheduled for manual re-inspection.

[0119] The above steps 301-304 are based on the multimodal observation vector Z(t) and Bayesian state estimation, combined with the information entropy H(Z) judgment mechanism, which can trigger re-planning in real time when anomalies such as deviation from the path and sensor failure occur. The measured task completion rate is 10% higher than that of the no-feedback solution, and the system robustness is significantly enhanced.

[0120] Step 305: Exception log recording and visualization report generation:

[0121] In order to achieve full-cycle recording and tracing of the system operation status, this solution system has designed an abnormal event log recording mechanism and a periodic inspection report automatic generation module.

[0122] The structure of each abnormal event record is as follows:

[0123] ;

[0124] Among them, t i The timestamp of the exception; Type i Normal type (such as path deviation, sensor failure, stuck, etc.); Z i is the observation vector when anomaly occurs; (x i ,y i ) is the grid coordinate of the abnormal location; Risk i is the comprehensive risk value corresponding to the abnormality.

[0125] Step 306: High-risk area management and dynamic threshold adjustment:

[0126] When anomalies occur continuously at a certain location, the system marks it as a "high-risk area". However, in order to avoid misjudgment and information redundancy, the following mechanism is introduced for automatic clearing and dynamic risk adjustment.

[0127] (1) High-risk mark revocation mechanism

[0128] If a high-risk grid (x, y) does not have any anomalies within Mclear cycles, the risk mark of the point will be automatically revoked:

[0129] ;

[0130] Among them, M0(x,y) is the initial static risk.

[0131] (2) Dynamic threshold adjustment for risk assessment

[0132] In scenarios where the environment changes dramatically or false alarms occur frequently, the risk confidence threshold δ and the information entropy threshold ε can be dynamically adjusted to:

[0133] ;

[0134] Among them, σ t is the abnormal fluctuation intensity in the recent period; η1, η2 are adjustment factors; δ0, ε0 are the initial confidence and entropy thresholds.

[0135] Step 307: Manual review task scheduling and queue-jumping mechanism:

[0136] (1) Rules for generating manual review tasks

[0137] If the grid (x, y) appears more than N times in the time window [t, t+T] abn The following exception triggers a manual re-inspection task:

[0138] ;

[0139] (2) Feedback mechanism for re-examination results

[0140] Operation and maintenance personnel can enter the re-inspection results through the platform:

[0141] Normal: The risk mark is cleared and the scheduling system returns to the original plan;

[0142] Abnormal confirmation: Update the risk value of the point to the highest level M(x,y)=1, and restrict the robot from entering.

[0143] (3) Queue-jumping mechanism

[0144] The manual review task has the highest priority, denoted as πmanual=1. The system inserts it into the head of the task queue and executes it as follows:

[0145] Original queue task set T={t1,t2,...,tn};

[0146] After insertion: T′={manual_review,t1,...,t n}.

[0147] Step 308: Exception type extension and system interface opening:

[0148] (1) Custom exception type configuration

[0149] The system supports extending exception types through configuration files, such as:

[0150] Abnormal types include deviation from the path, lidar failure, image recognition failure, communication interruption, and equipment jamming, which are automatically classified and matched by the perception module after loading.

[0151] (2) Interface protocol definition

[0152] Custom exception types: In addition to standard exceptions such as deviation and sensor failure, new exception types such as jamming, collision, and communication interruption can be added through the configuration file and automatically loaded during system initialization.

[0153] Open interface: Provide RESTful API or message bus interface to push abnormal data to third-party operation and maintenance systems, realizing multi-system linkage and secondary development.

[0154] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made based on the essence of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. A path planning method for an all-terrain autonomous inspection robot for high-voltage transmission and distribution lines, characterized in that: include: Based on the static prior risk map, an immediate risk assessment is performed on static obstacles within the target distribution line, and the risk of dynamic obstacles is predicted using a Markov chain. These are defined as immediate risk and predicted risk, respectively. The immediate risk and predicted risk are superimposed through exponential decay to generate a comprehensive risk map. A model graph is generated based on the target distribution line. Based on this model graph, a risk-aware A* search algorithm is used as the backbone path for global path planning. A deep Q network is then introduced to fine-tune local paths in real time. Finally, a heuristic greedy-genetic algorithm hybrid strategy is used to iteratively optimize the task sequence, taking into account task priorities and time window constraints to output the optimal inspection scheduling sequence. The robot performs inspection tasks according to the comprehensive risk map and the constraints of the optimal inspection scheduling sequence. During the inspection tasks, the optimal inspection scheduling sequence is continuously and dynamically corrected by an adaptive task feedback mechanism based on a multimodal perception fusion model and Bayesian state estimation; The static priori risk map is constructed by using indoor SLAM to perform offline mapping of the interior of the target distribution line to obtain a grid map of the distribution line clearance area; on the grid map, walls, cabinet edges, and fixed equipment are marked as high-risk areas to form the static priori risk map; The instant risk assessment process includes: calling a laser radar and a camera to collect obstacle distance information in each sensing cycle, mapping the obstacle distance information to grid points of the grid map for dynamic risk perception, obtaining the robot's current instant risk value, and inserting the instant risk value into the corresponding position in the static prior risk map to form a time-varying risk map; The instant risk value is calculated as follows: ； Where d is the distance information of the obstacle, x and y are the current horizontal and vertical coordinates of the robot, respectively, and λ d is the instantaneous attenuation rate corresponding to the obstacle distance information d. When d approaches 0, p approaches 1, and as the distance increases, the risk decays rapidly in an exponential form.

2. The path planning method for an all-terrain autonomous inspection robot for high-voltage power transmission and distribution lines according to claim 1, characterized in that: The Markov chain prediction process includes: For each grid point in the time-varying risk map, a first-order Markov chain is established, and a transfer matrix is ​​used to describe the risk transfer probability and risk distribution vector of the robot between adjacent grids. The predicted risk value is calculated based on the risk transfer probability and the risk distribution vector.

3. The path planning method for an all-terrain autonomous inspection robot for high-voltage power transmission and distribution lines according to claim 1, characterized in that: The scheduling objective function of the heuristic greedy-genetic algorithm hybrid strategy is: ； Where, is the delay penalty weight coefficient of the kth sensor cycle, is the delay of the kth sensor cycle, d(v ti ,v ti+1 ) is the path length between adjacent task points, θ is the path length weight, and m is the total number of tasks.

4. The path planning method for an all-terrain autonomous inspection robot for high-voltage power transmission and distribution lines according to claim 1, characterized in that: The construction process of the multimodal perception fusion model includes: Let the observation of the i-th sensor at time t be z i (t), then the total observation vector is: ； Construct each state S k Conditional probability in observation space: ； Among them, P(Z|S k ) is obtained by fitting the Bayesian network with the training data, where k is the index number of the state and j is the index number of the traversal.

Images