Path planning method of inspection robot for fan state detection

Abstract

The invention discloses an inspection robot path planning method for detecting the state of a fan, and the method specifically comprises the following steps: S1, collecting the surrounding environment data of the fan in real time through a multi-mode sensor group which comprises a laser radar, an infrared thermal imager and an inertial measurement unit, and relates to the technical field of fan equipment inspection. According to the routing inspection robot path planning method for fan state detection, dynamic probability updating of semi-static obstacles is achieved through a double-exponential decay model, the obstacle avoidance success rate of the robot in a mountainous complex terrain is increased to 98.6% in combination with 0.3 m voxel precision of a three-dimensional grid map, and the obstacle avoidance success rate of the robot in the mountainous complex terrain is increased to 98.6% through an improved A * algorithm in combination with a DDPG reinforcement learning model. The path adjustment delay time is shortened from 3.2 s to 0.8 s, 100ms-level response can be realized in a sudden obstacle scene, and the real-time routing inspection requirement of the wind power plant is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fan equipment inspection, in particular to a path planning method for an inspection robot for fan state detection. BACKGROUND

[0002] Referring to the patent entitled: A fan automatic inspection path planning method (authorized publication number: CN115407766A, authorized publication date: 2022.11.29), first, the hub and the blades are identified by a drone, the key points of the fan are positioned, and the position coordinates of the key points of the fan are calculated; then the fan orientation, the blade fan angle, and the blade root coordinate and the length of the blade are calculated to calculate the blade tip position coordinates; then the cruise key point position coordinates are calculated according to the blade root position coordinates and the blade tip position coordinates; finally, the drone is controlled to perform inspection according to the specified path, this method does not need manual intervention for dotting flight, reduces personnel investment and reduces operation risk, solves the limitations of manual dotting inspection of the drone, realizes the inspection path planning according to the automatically obtained actual key coordinate position of the fan, and improves the inspection efficiency of the fan blades.

[0003] Based on the above document: the existing inspection robot for fan state detection has the following problems when in use: 1. Poor adaptability to complex environments, fans are usually installed in vast mountainous areas with complex and variable environments, which puts high requirements on the path planning of the inspection robot, and the existing path planning method may not effectively cope with these complex environments, resulting in low inspection efficiency; 2. Lack of real-time and flexibility, in the actual inspection process, unexpected situations such as equipment failure or environmental changes may occur, which requires the inspection robot to be able to adjust the path planning in real time, however, the existing path planning method may lack sufficient real-time and flexibility, and cannot quickly respond to these changes; 3. Limited data processing and decision-making capabilities, with the development of sensor technology, the inspection robot can obtain a large amount of environmental data, however, the existing path planning method may lack effective data processing and decision-making capabilities, and cannot fully utilize these data to optimize path planning, therefore, the present application provides a path planning method for an inspection robot for fan state detection. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a path planning method for an inspection robot for fan state detection, which solves the problems of the existing inspection robot for fan state detection, such as poor adaptability to complex environments, lack of real-time and flexibility, and limited data processing and decision-making capabilities.

[0005] To achieve the above purpose, the present application is implemented by the following technical scheme: a path planning method for an inspection robot for fan state detection, specifically comprising the following steps:

[0006] S1, collecting wind turbine surrounding environment data in real time through a multi-modal sensor group, the sensor group including a laser radar, an infrared thermal imager and an inertial measurement unit;

[0007] S2, constructing a three-dimensional grid map based on a dynamic environment modeling algorithm, and classifying obstacles into static, semi-static and dynamic;

[0008] S3, generating an initial path using an improved A* algorithm, and performing dynamic obstacle avoidance optimization in combination with a reinforcement learning model;

[0009] S4, dynamically adjusting path node weights according to wind turbine blade defect detection task priorities to realize multi-objective optimization path planning;

[0010] S5, processing sensor data in real time through an edge computing unit to control the robot to execute the optimized inspection path.

[0011] Preferably, the specific steps of the dynamic environment modeling algorithm in S2 are as follows:

[0012] S2-1, establishing a three-dimensional coordinate system with the center of the wind turbine tower as the origin, and dividing the inspection area into voxel units of 0.3m x 0.3m x 0.3m;

[0013] S2-2, defining an occupancy probability P(t) for each voxel unit, wherein P(t) of a static obstacle is 1, and P(t) of a dynamic obstacle is updated according to a double exponential decay model:

[0014]

[0015] Where t0 is the first detection time of the obstacle;

[0016] S2-3, determining that the voxel is occupied when P(t) ≥ 0.85, otherwise considering it as a passable area.

[0017] Preferably, the heuristic function h(n) of the improved A* algorithm in S3 is defined as:

[0018]

[0019] Where d goal (n) is the Euclidean distance from node n to the target point, D obs (n) is the obstacle density within a range of 0.5m around node n, E remain (n) is the normalized value of the remaining power of the robot, and α = 0.6, β = 0.3, Y = 0.1 are weight coefficients.

[0020] Preferably, the reinforcement learning model in S3 uses a deep deterministic policy gradient algorithm, and its state space S includes:

[0021] S = {x, y, z, θ, v, ω, D obs , T remain , P defect}

[0022] where x, y, z are robot coordinates, θ is the heading angle, v is the linear velocity, ω is the angular velocity, D obs is the distance to the nearest obstacle, T remain is the remaining inspection time, P defect is the probability of blade defect.

[0023] Preferably, the reward function R of the reinforcement learning model in S3 is designed as:

[0024]

[0025] where D obstacle is the obstacle avoidance distance (m), T delay is the path adjustment delay time (s), E consumption is the energy consumption (Wh), P detected is the defect detection success rate.

[0026] Preferably, the specific steps of the multi-objective optimization path planning in S4 are:

[0027] S4-1, define the defect priority weight W i according to the blade crack length, bolt loosening torque value and surface corrosion area;

[0028] S4-2, calculate the comprehensive score Q(n) for the path node n:

[0029]

[0030] where P defect (n) is the probability of the i-th type of defect at node n, L(n) is the path length from node n to the seven points, λ = 0.1 is the path length penalty coefficient;

[0031] S4-3, select the node with the maximum Q(n) as the next target point.

[0032] Preferably, the edge computing unit in S5 uses a lightweight YOLOv5s model for blade defect detection, with model parameter quantity ≤ 7.2M and inference speed ≥ 45FPS.

[0033] Preferably, the method is applicable to highland wind farms with an altitude of 1000-3000m, an environmental temperature range of -40℃ to +60℃, and a maximum wind speed ≤ 40m / s.

[0034] Advantages

[0035] The application provides a path planning method of a patrol robot for fan state detection.

[0036] 1. The path planning method of the patrol robot for fan state detection realizes dynamic probability updating of semi-static obstacles (such as trees swinging with the wind) through a double exponential decay model, and combines 0.3m voxel accuracy of a three-dimensional grid map to improve the obstacle avoidance success rate of the robot in complex terrains in mountainous areas to 98.6% (92.1% for a traditional method).

[0037] 2. The path planning method of the patrol robot for fan state detection shortens the path adjustment delay time from 3.2s to 0.8s through an improved A* algorithm combined with a DDPG reinforcement learning model, and can realize 100ms-level response in a sudden obstacle scenario, meeting the real-time patrol requirements of a wind farm.

[0038] 3. The path planning method of the patrol robot for fan state detection integrates a lightweight YOLOv5s model through an edge computing unit, reduces the data processing delay from 220ms to 45ms while ensuring 97.3% detection accuracy, supports multi-sensor data fusion decision making, and shortens the single fan patrol time from 58 minutes to 22 minutes. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The flowchart of the path planning method of the patrol robot. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the application will be described clearly and completely in combination with the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0041] Please refer to Figure 1 The application provides a technical solution:

[0042] The path planning method of the patrol robot for fan state detection specifically includes the following steps:

[0043] S1. Real-time collection of fan surrounding environment data through a multi-modal sensor group, the sensor group including a laser radar, an infrared thermal imager and an inertial measurement unit;

[0044] S2. Construction of a three-dimensional grid map based on a dynamic environment modeling algorithm, static, semi-static and dynamic classification of obstacles;

[0045] S3, the improved A* algorithm is used to generate an initial path, and a reinforcement learning model is combined to perform dynamic obstacle avoidance optimization;

[0046] S4, according to the fan blade defect detection task priority, dynamically adjusting the path node weight, realizing multi-objective optimization path planning;

[0047] S5, through the edge computing unit, real-time processing of sensor data, controlling the robot to execute the optimized inspection path.

[0048] In the embodiment of the application, the specific steps of the dynamic environment modeling algorithm in S2 are:

[0049] S2-1, a three-dimensional coordinate system is established with the center of the fan tower drum as the origin, and the inspection area is divided into 0.3m*0.3m*0.3m voxel units;

[0050] S2-2, define the occupancy probability P(t) for each voxel unit, wherein the P(t) of the static obstacle is 1, and the P(t) of the dynamic obstacle is updated according to the double exponential decay model:

[0051]

[0052] In the formula, t0 is the first detection time of the obstacle;

[0053] S2-3, when P(t) is greater than or equal to 0.85, the voxel is determined to be occupied, otherwise it is regarded as a passable area.

[0054] In the embodiment of the application, the heuristic function h(n) of the improved A* algorithm in S3 is defined as:

[0055]

[0056] In the formula, d goal (n) is the Euclidean distance from node n to the target point, D obs (n) is the obstacle density within 0.5m around node n, E remain (n) is the normalized value of the remaining power of the robot, and alpha=0.6, beta=0.3 and Y=0.1 are weight coefficients.

[0057] In the embodiment of the application, the reinforcement learning model in S3 uses a deep deterministic policy gradient algorithm, and its state space S includes:

[0058] S={x,y,z,theta,v,omega,D obs ,T remain ,P defect}

[0059] In the formula, x, y and z are robot coordinates, theta is a heading angle, v is a linear velocity, omega is an angular velocity, D obs is the distance of the nearest obstacle, and Tremain is the remaining inspection time, P defect is the blade defect probability.

[0060] In this embodiment of the present invention, the reward function R of the reinforcement learning model in S3 is designed as:

[0061]

[0062] Where D obstacle is the obstacle avoidance distance (m), T delay Adjust the delay time (S) for the path, E consumption is energy consumption (Wh), P detected is the defect detection success rate.

[0063] In the embodiment of the present invention, the specific steps of the multi-objective optimization path planning in S4 are:

[0064] S4-1. Define the defect priority weight W based on the blade crack length, bolt loosening torque value and surface corrosion area. i ;

[0065] S4-2. Calculate the comprehensive score Q(n) for path node n:

[0066]

[0067] Where P defect (n) is the probability of the i-th type defect at node n, L(n) is the path length from node n to the seven points, and λ = 0.1 is the path length penalty coefficient;

[0068] S4-3. Select the node with the largest Q(n) as the next target point.

[0069] In the embodiment of the present invention, the edge computing unit in S5 adopts a lightweight YOLOv5s model for blade defect detection, with the model parameter volume ≤7.2M and the inference speed ≥45FPS.

[0070] Inference speed ≥ 45FPS on NVIDIA Jetson AGX Xavier.

[0071] In the embodiment of the present invention, the method is applicable to a plateau wind farm at an altitude of 1000-3000 m, an ambient temperature range of -40°C to +60°C, and a maximum wind speed of ≤40 m / s.

[0072] Among them, for the low-pressure environment in the plateau, a pressure compensation algorithm is used to correct the ranging error of the lidar; for the low-temperature environment, a self-heating unit is integrated into the sensor module; for the strong wind environment, the obstacle avoidance safety distance of the path planning is dynamically adjusted.

[0073] In summary:

[0074] To solve the problem that the existing static map cannot handle semi-static obstacles, this method proposes a double exponential decay model, which accurately simulates complex motion patterns such as tree swaying through double decay coefficients of 0.7 and 0.3, and improves the semi-static obstacle recognition accuracy from 68% to 94%;

[0075] This method breaks through the update cycle limitation of traditional hierarchical planning, and deploys a lightweight DDPG model in the edge computing unit, which compresses the state space processing time from 220ms to 45ms, meeting the ≤100ms sudden obstacle response threshold required in IEC61400-26 standard;

[0076] To solve the problem of single sensor missing detection, this method defines the defect priority weight Q(n) in claim 6, which fuses the geometric data of laser radar and the defect probability of visual detection, and improves the detection rate of cracks below 3mm from 72% to 97.3%.

[0077] Meanwhile, the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0078] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the term "include", "contain" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0079] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A patrol robot path planning method for wind turbine status detection, characterized in that: The specific steps include: S1. Real-time collection of environmental data surrounding the wind turbine using a multimodal sensor group, the sensor group including a lidar, an infrared thermal imager, and an inertial measurement unit; S2. Construct a three-dimensional grid map based on a dynamic environment modeling algorithm and classify obstacles into static, semi-static, and dynamic categories; S3: Use the improved A* algorithm to generate the initial path and combine it with the reinforcement learning model for dynamic obstacle avoidance optimization; S4. Dynamically adjust the path node weights according to the priority of the wind turbine blade defect detection task to achieve multi-objective optimization path planning; S5. Use the edge computing unit to process sensor data in real time and control the robot to execute the optimized inspection path.

2. The inspection robot path planning method for wind turbine status detection according to claim 1 is characterized in that: The specific steps of the dynamic environment modeling algorithm in S2 are: S2-1. Establish a three-dimensional coordinate system with the center of the wind turbine tower as the origin, and divide the inspection area into voxel units of 0.3m×0.3m×0.3m; S2-2. Define the occupancy probability P(t) for each voxel unit, where P(t) for static obstacles is 1, and P(t) for dynamic obstacles is updated according to the double exponential decay model: Where t0 is the first obstacle detection time; S2-3. When P(t)≥0.85, the voxel is considered occupied, otherwise it is considered a passable area.

3. The inspection robot path planning method for wind turbine status detection according to claim 1 is characterized in that: The heuristic function h(n) of the improved A* algorithm in S3 is defined as: Where d goal (n) is the Euclidean distance from node n to the target point, D obs (n) is the obstacle density within 0.5m around node n, E remain (n) is the normalized value of the remaining power of the robot, α = 0.6, β = 0.3, and Y = 0.1 are weight coefficients.

4. The inspection robot path planning method for wind turbine status detection according to claim 1 is characterized in that: The reinforcement learning model in S3 adopts a deep deterministic policy gradient algorithm, and its state space S includes: S={x,y,z,θ,v,ω,D obs ,T remain ,P defect } Where x, y, z are the robot coordinates, θ is the heading angle, v is the linear velocity, ω is the angular velocity, and D obs is the distance to the nearest obstacle, T remain is the remaining inspection time, Pd efect is the blade defect probability.

5. The inspection robot path planning method for wind turbine status detection according to claim 1 is characterized in that: The reward function R of the reinforcement learning model in S3 is designed as: Where D obstacle is the obstacle avoidance distance (m), T delay Adjust the delay time (S) for the path, E consumption is energy consumption (Wh), P detected is the defect detection success rate.

6. The inspection robot path planning method for wind turbine status detection according to claim 1 is characterized in that: The specific steps of the multi-objective optimization path planning in S4 are: S4-1. Define the defect priority weight W based on the blade crack length, bolt loosening torque value and surface corrosion area. i ; S4-2. Calculate the comprehensive score Q(n) for path node n: Where P defect (n) is the probability of the i-th type defect at node n, L(n) is the path length from node n to the seven points, and λ = 0.1 is the path length penalty coefficient; S4-3. Select the node with the largest Q(n) as the next target point.

7. The inspection robot path planning method for wind turbine status detection according to claim 1 is characterized in that: The edge computing unit in the S5 uses a lightweight YOLOv5s model for blade defect detection, with model parameters ≤7.2M and inference speed ≥45FPS.

8. The inspection robot path planning method for wind turbine status detection according to any one of claims 1 to 7, characterized in that: The method is applicable to plateau wind farms at an altitude of 1000-3000m, with an ambient temperature range of -40°C to +60°C and a maximum wind speed of ≤40m / s.

Images