Inspection robot autonomous navigation system based on deep learning

Abstract

The invention relates to the technical field of inspection robots, and particularly discloses an inspection robot autonomous navigation system based on deep learning, which comprises an environment sensing module, a dynamic map construction module, a path planning module, an obstacle avoidance control module and an execution feedback module. According to the scheme, a convolutional neural network and an attention mechanism are introduced, deep feature extraction and dynamic target processing are performed on multi-modal data of an inspection environment, and high-precision global dynamic map construction is realized; equipment importance and abnormal position detection are introduced to form a path priority, an important target is preferentially accessed, and the dynamic adaptability of an inspection path is improved according to a global dynamic map and a path risk value updated in real time, so that the inspection robot can efficiently complete a task in a complex dynamic environment; real-time obstacle avoidance and local path re-planning in a dynamic environment are realized, the future collision time is predicted by combining relative motion modeling with a kinematic model, the method has strong local path adaptability, and the stability of robot inspection control is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of inspection robots, in particular to an autonomous navigation system of an inspection robot based on deep learning. BACKGROUND

[0002] With the continuous improvement of the level of industrial intelligence, inspection robots are increasingly widely used in complex industrial environments. The inspection environment usually has problems such as irregular spatial structure, various types of obstacles, large changes in lighting conditions, and frequent appearance of dynamic objects, which leads to insufficient ability of the robot to identify and track dynamic targets; general inspection path planning usually only considers static access cost, ignores target importance and abnormal scenes, and lacks adaptation to environmental changes; and general inspection obstacle avoidance control strategies are simple and cannot finely control speed and steering in obstacle areas, and easily ignore the motion trend of local obstacles, leading to misjudgment or delayed response. SUMMARY

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present application provides an autonomous navigation system of an inspection robot based on deep learning, which is aimed at the problems of irregular spatial structure, various types of obstacles, large changes in lighting conditions, and frequent appearance of dynamic objects in the inspection environment, which leads to insufficient ability of the robot to identify and track dynamic targets. The present scheme introduces a convolutional neural network and an attention mechanism to extract deep features and process dynamic targets from multi-modal data of the inspection environment, and realizes high-precision global dynamic map construction; in view of the problem that general inspection path planning usually only considers static access cost, ignores target importance and abnormal scenes, and lacks adaptation to environmental changes, the present scheme introduces device importance and abnormal position detection to constitute path priority, preferentially accesses important targets, and improves the dynamic adaptability of the inspection path according to the global dynamic map and the real-time updated path risk value, to ensure that the inspection robot can efficiently complete the task in a complex dynamic environment; in view of the problem that general inspection obstacle avoidance control strategies are simple and cannot finely control speed and steering in obstacle areas, and easily ignore the motion trend of local obstacles, leading to misjudgment or delayed response, the present scheme realizes real-time obstacle avoidance and local path re-planning in a dynamic environment, predicts the future collision time by using relative motion modeling combined with a kinematic model, has strong local path adaptability, and ensures that the robot can avoid potential risks in advance, thereby improving the stability of the robot inspection control.

[0004] The autonomous navigation system of the inspection robot based on deep learning provided by the present application comprises an environment perception module, a dynamic map construction module, a path planning module, an obstacle avoidance control module, and an execution feedback module.

[0005] The environment perception module installs a data acquisition device inside the inspection robot, collects multi-modal data of the inspection environment in real time, generates perception input data in a unified format through data preprocessing, and sends the perception input data to a dynamic map construction module, a path planning module, and an obstacle avoidance control module;

[0006] The dynamic map construction module uses a convolutional neural network to perform semantic segmentation and feature extraction on the perception input data, constructs a global dynamic map and labels obstacle information, and sends the global dynamic map to the path planning module and the obstacle avoidance control module;

[0007] The path planning module determines device importance and abnormal positions according to the perception input data, calculates device importance scores and abnormal detection scores, and generates an inspection path based on the global dynamic map using an adaptive path optimization algorithm, and sends the inspection path to the obstacle avoidance control module;

[0008] The obstacle avoidance control module establishes a kinematic model of the inspection robot to achieve local obstacle avoidance, detects local obstacles in front in real time and adjusts the motion trajectory, and outputs the motion trajectory of the inspection robot to the execution feedback module;

[0009] The execution feedback module analyzes the motion trajectory of the inspection robot, constructs a trajectory evaluation function, and dynamically adjusts the path planning parameters according to the trajectory evaluation result to form a closed-loop optimization.

[0010] Further, the dynamic map construction module includes a model construction unit, a feature extraction unit, and a dynamic map generation unit.

[0011] The model construction unit initializes a convolutional neural network, loads pre-trained weights, inputs perception input data, and constructs an encoder-decoder architecture to extract inspection environment features: the encoder uses a ResNet structure to capture high-level semantic information in the perception input data; the decoder recovers the spatial resolution through upsampling and reconstructs the environment details;

[0012] The feature extraction unit introduces an attention mechanism in the convolutional neural network, generates feature representations of dynamic targets based on the inspection environment features, including a feature extraction layer, a weight calculation layer, and a weighted fusion layer:

[0013] The feature extraction layer uses a 1x1 convolution kernel to reduce the dimensionality of the perception input data to obtain a set of low-dimensional feature vectors;

[0014] The weight calculation layer calculates a similarity matrix between feature vectors in the set of low-dimensional feature vectors using cosine similarity, and obtains attention weights after normalization, using the following formula: ; ;

[0015] In the formula, Represents the feature similarity matrix. , Indicates the index of the feature vector. and They represent the first in the low-dimensional feature vector set. The and the first 1 eigenvector Represents cosine similarity. Indicates attention weights, Represents the th element in the similarity matrix The importance of each feature vector The total number of eigenvectors in the low-dimensional eigenvector set;

[0016] The weighted fusion layer performs weighted fusion of feature vectors according to attention weights to generate a feature representation of the dynamic target, using the following formula: ;

[0017] In the formula, Characteristic representation of dynamic targets;

[0018] The dynamic map generation unit fuses the feature representation of dynamic targets with the perceived input data, labels the categories, locations and movement trends of static obstacles and dynamic targets, and outputs a global dynamic map.

[0019] Furthermore, the path planning module includes an equipment importance scoring module, an anomaly detection scoring module, a path priority calculation module, a path risk calculation module, a path cost calculation module, and an inspection path generation module;

[0020] The equipment importance scoring module determines the importance of equipment based on the perceived input data, counts the number of equipment failures over a period of time to obtain the historical failure frequency of the equipment, quantifies the criticality of the equipment based on its role in the production process, and defines the equipment importance scoring function.

[0021] The anomaly detection and scoring module determines the location of the anomaly based on the sensing input data, classifies its hazard level according to the degree of impact of the anomaly location on the operation of the production line, assesses the severity of the anomaly type based on the potential losses, and defines the anomaly detection and scoring function.

[0022] The path priority calculation module plans static inspection paths based on equipment importance scores and anomaly detection scores, and calculates the priority scores of the static inspection paths.

[0023] The path risk calculation module analyzes the distribution of obstacles in the global dynamic map, calculates the risk value of each grid cell in the global dynamic map, and generates a risk-weighted path map.

[0024] The path cost calculation module calculates the comprehensive generation value of each grid cell based on the priority score of the static inspection path and the risk weighted path graph using the improved A* algorithm;

[0025] The inspection path generation module adjusts the static inspection path of the robot according to the comprehensive generation value of the grid cell, preferentially expands the grid cell with small comprehensive generation value, and reaches the target grid cell. After reaching the target grid cell, backtracking is performed in reverse to generate a complete inspection path.

[0026] Further, the obstacle avoidance control module establishes a kinematic model for the inspection robot to realize local obstacle avoidance, including the following steps:

[0027] Step S1: Local environment perception, the local obstacle when the robot moves along the inspection path is detected by using the perception input data, and a three-dimensional obstacle model of the front area of the inspection path is constructed: the distance and profile of the obstacle are measured by laser radar, the blind area near distance obstacle is detected by ultrasonic sensor, and the obstacle category, texture and color information are recognized by visual camera;

[0028] Step S2: Dynamic obstacle prediction, the motion state of the robot is obtained from the perception input data, the possible collision time is predicted based on the three-dimensional obstacle model, and the safety threshold is set;

[0029] Step S3: Local obstacle avoidance path generation, if the predicted collision time is less than the safety threshold, a local obstacle avoidance path is generated between the current position and the next position;

[0030] Step S4: Motion state estimation, the motion state of the robot is decomposed by using Kalman filter, the velocity command is calculated on the local obstacle avoidance path by using the improved dynamic window method, and the formula is as follows: ; ; ;

[0031] In the formula, represents the time index, 、 、 respectively represent the horizontal, vertical coordinates and angle of the robot at the first time in the global dynamic map, 、 respectively represent the linear velocity and angular velocity of the robot at the first time, is the time step;

[0032] Step S5: instruction execution, driving the inspection robot to move according to the calculated speed instruction, outputting the motion trajectory of the inspection robot, repeating the calculation process of the speed instruction in step S4 until the inspection robot reaches the tolerance range of the target inspection position, and stopping moving.

[0033] Further, in step S2, the dynamic obstacle prediction comprises the following steps:

[0034] Step S21: extracting the distance and position of the local obstacle based on the three-dimensional obstacle model;

[0035] Step S22: converting the distance and position of the local obstacle into a motion vector in the three-dimensional space by using the optical flow method to predict the future position of the obstacle, and the formula used is as follows: ,

[0036] In the formula, represents the future position of the obstacle, represents the current position of the local obstacle, represents the motion speed of the local obstacle, is the prediction time interval;

[0037] Step S23: predicting the collision time, comprising the following steps:

[0038] Step S231: relative state definition, obtaining the position and speed of the inspection robot in the three-dimensional space at the current time, combining the obstacle information labeled by the global dynamic map and the predicted future position of the obstacle, and calculating the relative position vector and the relative speed vector of the inspection robot and the local obstacle, and the formula used is as follows: ; ;

[0039] In the formula, and represent the current position and speed of the obstacle, and represent the position and speed of the inspection robot in the three-dimensional space, and represent the relative position vector and the relative speed vector between the inspection robot and the local obstacle;

[0040] Step S232: shortest approach time calculation, minimizing the square of the distance between the robot and the local obstacle, calculating the shortest approach time under relative motion, i.e. the time before potential collision occurs, and the formula used is as follows: ;

[0041] In the formula, represents the shortest approaching time;

[0042] Step S233: collision time output, predicting the minimum distance between the inspection robot and the local obstacle at the shortest approaching time, setting a safety threshold, if the minimum distance is less than the safety threshold, it is judged that there is a collision risk, and the shortest approaching time at this time is the predicted collision time, and the formula is as follows: ;

[0043] In the formula, represents the minimum distance.

[0044] The beneficial effects obtained by the above scheme are as follows:

[0045] (1) In view of the problems that the inspection environment usually has irregular spatial structure, various types of obstacles, large changes in lighting conditions, and frequent appearance of dynamic objects, which leads to insufficient identification and tracking ability of the robot to dynamic targets, the scheme introduces a convolutional neural network and an attention mechanism to extract deep features and process dynamic targets of multi-modal data of the inspection environment, and realizes high-precision global dynamic map construction.

[0046] (2) In view of the problems that general inspection path planning usually only considers static access cost, ignores target importance and abnormal scenes, and lacks adaptation to environmental changes, the scheme introduces device importance and abnormal position detection to constitute path priority, preferentially accesses important targets, and improves the dynamic adaptability of the inspection path according to the global dynamic map and the real-time updated path risk value, to ensure that the inspection robot can efficiently complete the task in a complex dynamic environment.

[0047] (3) In view of the problems that general inspection obstacle avoidance control strategy is simple and cannot finely control speed and steering in obstacle areas, and is easy to ignore the motion trend of local obstacles, leading to misjudgment or slow response, the scheme realizes real-time obstacle avoidance and local path re-planning in a dynamic environment, predicts the future collision time by using relative motion modeling combined with a kinematic model, has strong local path adaptability, and ensures that the robot can avoid potential risks in advance, and improves the stability of the robot inspection control. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is a schematic diagram of the autonomous navigation system of the inspection robot based on deep learning proposed by the present application;

[0049] Figure 2 is a process schematic diagram for establishing a kinematic model of the inspection robot to realize local obstacle avoidance.

[0050] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification. They are used together with embodiments of the present application to explain the present application, and do not constitute a limitation of the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0052] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0053] Embodiment one, refer to Figure 1 The autonomous navigation system of the inspection robot based on deep learning provided by the present application comprises an environment perception module, a dynamic map construction module, a path planning module, an obstacle avoidance control module and an execution feedback module.

[0054] The environment perception module internally installs a data acquisition device in the inspection robot, acquires multi-modal data of the inspection environment in real time, including visual images, point cloud data, ultrasonic ranging information and robot pose data, generates unified format perception input data through data preprocessing, and sends the perception input data to the dynamic map construction module, the path planning module and the obstacle avoidance control module.

[0055] The dynamic map construction module adopts a convolutional neural network to perform semantic segmentation and feature extraction on the perception input data, constructs a global dynamic map and labels obstacle information, and sends the global dynamic map to the path planning module and the obstacle avoidance control module.

[0056] The path planning module determines the importance of the equipment and the abnormal position according to the perception input data, calculates the equipment importance score and the abnormal detection score, and generates an inspection path based on the global dynamic map using an adaptive path optimization algorithm, and sends the inspection path to the obstacle avoidance control module.

[0057] The obstacle avoidance control module establishes a kinematic model of the inspection robot to realize local obstacle avoidance, detects the local obstacle in front in real time and adjusts the motion trajectory, and outputs the motion trajectory of the inspection robot to the execution feedback module.

[0058] The execution feedback module analyzes the motion trajectory of the inspection robot, constructs a trajectory evaluation function, and dynamically adjusts the path planning parameters according to the trajectory evaluation result to form a closed loop optimization.

[0059] Embodiment two, refer to Figure 1 , which is based on the above embodiment, in the environment perception module, the data acquisition device includes a laser radar, a visual camera, an ultrasonic sensor and an inertial measurement unit: the laser radar obtains high-precision distance point cloud data, and converts the point cloud data into a three-dimensional grid representation; the visual camera collects visual images, denoises and normalizes the visual images, and retains texture and color information; the ultrasonic sensor detects near-distance obstacles, and filters out noise interference; the inertial measurement unit records the attitude change of the robot. The visual image, the point cloud data, the ultrasonic ranging information and the robot attitude data are converted into unified format perception input data and sent to the dynamic map construction module.

[0060] Embodiment three, refer to Figure 1 , which is based on the above embodiment, the dynamic map construction module includes a model construction unit, a feature extraction unit, a dynamic map generation unit;

[0061] The model construction unit initializes the convolutional neural network and loads the pre-trained weight, inputs the perception input data, and constructs an encoder-decoder architecture to extract the inspection environment features: the encoder uses the ResNet structure to capture the high-level semantic information in the perception input data; the decoder restores the spatial resolution through up-sampling, and reconstructs the environment details;

[0062] The feature extraction unit introduces an attention mechanism in the convolutional neural network, generates a feature representation of the dynamic target based on the inspection environment features, including a feature extraction layer, a weight calculation layer and a weighted fusion layer:

[0063] The feature extraction layer uses a 1x1 convolution kernel to reduce the dimension of the perception input data to obtain a set of low-dimensional feature vectors;

[0064] The weight calculation layer calculates the similarity matrix between the feature vectors in the set of low-dimensional feature vectors using cosine similarity, and obtains the attention weight after normalization, and the formula is as follows: ; ;

[0065] In the formula, represents the feature similarity matrix, , represents the index of the feature vector, and represent the first and the first a feature vector, denotes a cosine similarity, denotes an attention weight, denotes the importance of the i-th feature vector in the similarity matrix, denotes the total number of feature vectors in the low-dimensional feature vector set;

[0066] The weighted fusion layer weights and fuses the feature vectors according to the attention weights to generate a feature representation of the dynamic target, and the formula used is as follows: ;

[0067] In the formula, denotes the feature representation of the dynamic target;

[0068] The dynamic map generation unit fuses the feature representation of the dynamic target with the perception input data, labels the categories, positions and motion trends of the static obstacles and dynamic targets, and outputs a global dynamic map.

[0069] By performing the operation, for the problems of irregular spatial structure, various obstacle types, large changes in lighting conditions and frequent appearance of dynamic objects in the inspection environment, the robot has insufficient recognition and tracking ability for dynamic targets. The present scheme introduces a convolutional neural network and an attention mechanism to extract deep features and process dynamic targets from multi-modal data of the inspection environment, and realizes high-precision global dynamic map construction.

[0070] Embodiment four, see Figure 1 This embodiment is based on the above-mentioned embodiments, and the path planning module includes a device importance scoring module, an anomaly detection scoring module, a path priority calculation module, a path risk calculation module, a path cost calculation module and an inspection path generation module;

[0071] The device importance scoring module determines the device importance according to the perception input data, counts the number of faults of the device in the past period of time to obtain the historical fault frequency of the device, quantifies the criticality of the device according to the role of the device in the production process, defines a device importance scoring function, and the formula used is as follows: ;

[0072] In the formula, denotes the device index, denotes the device importance score, denotes the criticality of the device, denotes the historical fault frequency of the device, and denote the weight coefficients for balancing the influence of criticality and fault frequency;

[0073] ​The abnormality detection score module determines an abnormal position according to the perception input data, classifies a danger level of the abnormal position according to an influence degree of the production line operation, evaluates a severity of an abnormal type according to a possible loss, defines an abnormality detection score function, and the formula is as follows: ;

[0074] In the formula, represents the abnormal position, represents the abnormality detection score, represents the danger level of the abnormal position, represents the severity of the abnormal type, and represent weight coefficients;

[0075] The path priority calculation module plans a static inspection path according to the equipment importance score and the abnormality detection score, calculates a priority score of the static inspection path, and the formula is as follows: ;

[0076] In the formula, represents the priority score of the static inspection path, and represent path priority weights;

[0077] The path risk calculation module analyzes the obstacle distribution in the global dynamic map, calculates a risk value of each grid unit in the global dynamic map, generates a risk-weighted path map, and the formula is as follows: ;

[0078] In the formula, represents coordinates of the grid unit, represents the risk value of the grid unit, represents a probability of an obstacle appearing in the grid unit, represents a deviation value of the grid unit and the inspection path, represents a weight coefficient of the obstacle probability, represents a weight coefficient of the path deviation;

[0079] The path cost calculation module calculates a comprehensive cost value of each grid unit based on the priority score of the static inspection path and the risk-weighted path map, using an improved A* algorithm, and the formula is as follows: ;

[0080] In the formula, represents the grid unit, represents the comprehensive cost value of the grid unit, This represents the actual cost from the inspection starting point to the current grid cell. This represents the heuristically estimated cost from the current grid cell to the inspection target cell. and These represent the risk weighting coefficient and the priority weighting coefficient, respectively.

[0081] The inspection path generation module adjusts the robot's static inspection path based on the comprehensive cost value of the grid cells, prioritizing the expansion of grid cells with lower comprehensive cost value until the target grid cell is reached. After reaching the target grid cell, the robot backtracks to generate a complete inspection path.

[0082] By performing the aforementioned operations, this solution addresses the common problem that general inspection path planning typically only considers static access costs, ignores target importance and abnormal scenarios, and lacks adaptability to environmental changes. Instead, it introduces equipment importance and abnormal location detection to form path priorities, prioritizing access to important targets. Furthermore, it improves the dynamic adaptability of inspection paths based on a global dynamic map and real-time updated path risk values, ensuring that the inspection robot can efficiently complete tasks in complex dynamic environments.

[0083] Example 5, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. The obstacle avoidance control module establishes a kinematic model of the inspection robot to achieve local obstacle avoidance, including the following steps:

[0084] Step S1: Local environment perception. Utilize perception input data to detect local obstacles as the robot moves along the inspection path and construct a 3D obstacle model of the area in front of the inspection path: use lidar to measure the distance and contour of obstacles, ultrasonic sensors to detect nearby obstacles in blind spots, and visual cameras to identify obstacle categories, textures, and color information.

[0085] Step S2: Dynamic obstacle prediction. Obtain the robot's motion state from the sensor input data, predict possible collision times based on the 3D obstacle model, and set a safety threshold.

[0086] Step S3: Local obstacle avoidance path generation. If the predicted collision time is less than the safety threshold, call RRT* (Fast Extended Random Tree Algorithm) to generate a local obstacle avoidance path between the current position and the next position.

[0087] Step S4: Motion state estimation. The Kalman filter is used to decompose the robot's motion state. An improved dynamic window method is used to calculate the velocity command on the local obstacle avoidance path. The formula used is as follows: ; ; ;

[0088] In the formula, Indicates the time index. , , They represent the robot's number 1 and 2. The horizontal and vertical coordinates and angles are always displayed on the global dynamic map. , These represent the robot's position in the [missing information]. The linear velocity and angular velocity at time t. For time step;

[0089] Step S5: Command execution. Drive the inspection robot to move according to the calculated speed command, and output the movement trajectory of the inspection robot. Repeat the speed command calculation process in step S4 until the inspection robot reaches the tolerance range of the target inspection position and stops moving.

[0090] Example 6, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. Step S2, dynamic obstacle prediction, includes the following steps:

[0091] Step S21: Extract the distance and position of local obstacles based on the 3D obstacle model;

[0092] Step S22: The distance and position of local obstacles are converted into motion vectors in three-dimensional space using optical flow to predict the future position of the obstacles. The formula used is as follows: ,

[0093] In the formula, Indicates the future position of the obstacle. Indicates the current position of the local obstacle. Indicates the speed of movement of a local obstacle. For the prediction time interval;

[0094] Step S23: Predict the collision time, including the following steps:

[0095] Step S231: Define the relative state. Obtain the current position and velocity of the inspection robot in three-dimensional space. Combine the obstacle information marked on the global dynamic map and the predicted future positions of the obstacles, calculate the relative position vector and relative velocity vector between the inspection robot and the local obstacles. The formulas used are as follows: ; ;

[0096] In the formula, and These represent the current position and velocity of the obstacle, respectively. and These represent the position and velocity of the inspection robot in three-dimensional space, respectively. and These represent the relative position vector and relative velocity vector between the inspection robot and the local obstacle, respectively.

[0097] Step S232: Shortest approach time calculation. Minimize the square of the distance between the robot and the local obstacle to calculate the shortest approach time under relative motion, which is the time before a potential collision occurs. The formula used is as follows: ;

[0098] In the formula, Indicates the shortest approach time;

[0099] Step S233: Collision time output. Predict the minimum distance between the inspection robot and the local obstacle at the shortest approach time. Set a safety threshold. If the minimum distance is less than the safety threshold, it is determined that there is a collision risk. The shortest approach time at this time is the predicted collision time. The formula used is as follows: ;

[0100] In the formula, This represents the minimum distance.

[0101] Example 7, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. Step S3, local obstacle avoidance path generation, includes the following steps:

[0102] Step S31: Set the planning area. In the global dynamic map, with the current position of the inspection robot as the center, set the local planning radius and extract the local environmental mesh as the sampling space of RRT*.

[0103] Step S32: Initialize the tree structure. Initialize the tree structure with the current position as the root node and set the next grid cell on the inspection path as the termination node.

[0104] Step S33: Perform random sampling. Randomly sample a point in the sampling space, search for the node closest to the random point in the search tree, connect the nearest node and generate a new node;

[0105] Step S34: Collision detection. Based on the 3D obstacle model, determine whether the line connecting the nearest node and the new node intersects with the obstacle. If there is no collision, add the new node to the tree.

[0106] Step S35: Path optimization, find the set of candidate parent nodes for the new node, select the parent node with the lowest cost, and backtrack to optimize the entire path;

[0107] Step S36: Upon reaching the target point, if the distance between a new node and the target point during the sampling process is less than the set tolerance, the connection is considered successful; backtrack the complete path from the tree and use that path as the local obstacle avoidance path at the current moment.

[0108] By performing the aforementioned operations, this solution addresses the problems of conventional obstacle avoidance control strategies being simple, unable to precisely control speed and steering within obstacle areas, and prone to overlooking the movement trends of local obstacles, leading to misjudgments or delayed responses. This solution achieves real-time obstacle avoidance and local path replanning in dynamic environments. By utilizing relative motion modeling combined with kinematic models to predict future collision times, it possesses strong local path adaptability, ensuring that the robot can avoid potential risks in advance and improving the stability of robot inspection control.

[0109] Example 8, see Figure 1 Figure 2 This embodiment is based on the above embodiment. The execution feedback module constructs a trajectory evaluation function based on three indicators: safety, smoothness, and efficiency of the motion trajectory. Specifically, it includes the following steps:

[0110] Step A1: Trajectory data reception. Based on the motion trajectory of the inspection robot output by the obstacle avoidance control module, obtain the robot's real-time timestamp, position coordinates, attitude angle, linear velocity, and angular velocity during inspection.

[0111] Step A2: Trajectory performance evaluation, constructing quantitative evaluation indicators from three dimensions: safety, smoothness, and efficiency. The specific steps are as follows:

[0112] Step A21: Safety evaluation. Calculate the minimum distance predicted and output by the obstacle avoidance control module at all times in the trajectory, and calculate the proportion of distances below the safety threshold. The lower the proportion, the higher the safety score.

[0113] Step A22: Smoothness evaluation. Calculate the rate of change of the robot's actual angular velocity and linear velocity. The smaller the average rate of change, the higher the smoothness score.

[0114] Step A23: Efficiency evaluation. Calculate the actual time taken for the robot to complete the inspection path and compare it with the estimated time when the path planning module generated the path. The closer the actual time is to the estimated time, the higher the efficiency score.

[0115] Step A3: Comprehensive evaluation. The scores for safety, smoothness, and efficiency are weighted and summed to obtain the overall trajectory evaluation function. When the score remains low, the risk weight coefficient and safety threshold in the path planning module are automatically adjusted. The formula used is as follows: ;

[0116] In the formula, This represents the trajectory evaluation function. The safety score indicates the trajectory. The score represents the smoothness of the trajectory. Efficiency score representing the trajectory. , , These are the weighting coefficients for each indicator.

[0117] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0119] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A deep learning-based autonomous navigation system for inspection robots, characterized in that: It includes an environmental perception module, a dynamic map construction module, a path planning module, an obstacle avoidance control module, and an execution feedback module; The environmental perception module is equipped with data acquisition devices inside the inspection robot to collect multimodal data of the inspection environment in real time. Through data preprocessing, it generates perception input data in a unified format and sends the perception input data to the dynamic map construction module, path planning module, and obstacle avoidance control module. The dynamic map construction module uses a convolutional neural network to perform semantic segmentation and feature extraction on the perceptual input data, constructs a global dynamic map and labels obstacle information, and sends the global dynamic map to the path planning module and the obstacle avoidance control module. The path planning module determines the importance of the equipment and the location of the anomaly based on the perception input data, calculates the equipment importance score and the anomaly detection score, and generates an inspection path based on the global dynamic map using an adaptive path optimization algorithm, and sends the inspection path to the obstacle avoidance control module. The obstacle avoidance control module establishes a kinematic model for the inspection robot to achieve local obstacle avoidance, detects local obstacles in front in real time and adjusts the movement trajectory, and outputs the movement trajectory of the inspection robot to the execution feedback module. The execution feedback module analyzes the motion trajectory of the inspection robot, constructs a trajectory evaluation function, and dynamically adjusts the path planning parameters based on the trajectory evaluation results to form a closed-loop optimization.

2. The deep learning-based autonomous navigation system for inspection robots according to claim 1, characterized in that: The dynamic map construction module includes a model construction unit, a feature extraction unit, and a dynamic map generation unit; The model building unit initializes the convolutional neural network and loads pre-trained weights, inputs perceptual input data, and constructs an encoder-decoder architecture to extract inspection environment features: encoder Use the ResNet structure to capture high-level semantic information in perceptual input data; The decoder restores spatial resolution through upsampling, thus reconstructing environmental details; The feature extraction unit introduces an attention mechanism into the convolutional neural network to generate a feature representation of dynamic targets based on the characteristics of the inspection environment. The dynamic map generation unit fuses the feature representation of dynamic targets with the perceived input data, labels the categories, locations and movement trends of static obstacles and dynamic targets, and outputs a global dynamic map.

3. The deep learning-based autonomous navigation system for inspection robots according to claim 2, characterized in that: The feature extraction unit includes a feature extraction layer, a weight calculation layer, and a weighted fusion layer. The feature extraction layer uses a 1x1 convolutional kernel to reduce the dimensionality of the perceptual input data, resulting in a low-dimensional feature vector set. The weight calculation layer uses cosine similarity to calculate the similarity matrix between feature vectors in the low-dimensional feature vector set, and then normalizes it to obtain the attention weights. The weighted fusion layer performs weighted fusion of feature vectors according to attention weights to generate a feature representation of the dynamic target.

4. The deep learning-based autonomous navigation system for inspection robots according to claim 1, characterized in that: The path planning module includes an equipment importance scoring module, an anomaly detection scoring module, a path priority calculation module, a path risk calculation module, a path cost calculation module, and an inspection path generation module. The equipment importance scoring module determines the importance of equipment based on the perceived input data, counts the historical failure frequency of the equipment, quantifies the criticality of the equipment, and defines the equipment importance scoring function. The anomaly detection and scoring module determines the location of the anomaly based on the perceived input data, classifies its danger level, assesses the severity of the anomaly type, and defines the anomaly detection and scoring function. The path priority calculation module plans static inspection paths based on equipment importance scores and anomaly detection scores, and calculates the priority scores of the static inspection paths. The path risk calculation module analyzes the distribution of obstacles in the global dynamic map, calculates the risk value of each grid cell in the global dynamic map, and generates a risk-weighted path map. The path cost calculation module calculates the comprehensive cost value of each grid cell based on the priority score and risk-weighted path graph of the static inspection path using an improved A* algorithm. The inspection path generation module adjusts the robot's static inspection path based on the comprehensive cost value of the grid cells, prioritizing the expansion of grid cells with lower comprehensive cost value until the target grid cell is reached. After reaching the target grid cell, the robot backtracks to generate a complete inspection path.

5. The deep learning-based autonomous navigation system for inspection robots according to claim 1, characterized in that: The obstacle avoidance control module establishes a kinematic model for the inspection robot to achieve local obstacle avoidance, including the following steps: Step S1: Local environment perception. Detect local obstacles as the robot moves along the inspection path using the perception input data, and construct a 3D obstacle model of the area in front of the inspection path. Step S2: Dynamic obstacle prediction. Obtain the robot's motion state from the sensor input data, predict possible collision times based on the 3D obstacle model, and set a safety threshold. Step S3: Local obstacle avoidance path generation. If the predicted collision time is less than the safety threshold, a local obstacle avoidance path is generated between the current position and the next position. Step S4: Motion state estimation. The motion state of the robot is decomposed using a Kalman filter, and the velocity command is calculated using an improved dynamic window method on the local obstacle avoidance path. Step S5: Command execution. Drive the inspection robot to move according to the calculated speed command, and output the movement trajectory of the inspection robot. Repeat the speed command calculation process in step S4 until the inspection robot reaches the tolerance range of the target inspection position and stops moving.

6. The deep learning-based autonomous navigation system for inspection robots according to claim 5, characterized in that: Step S2, dynamic obstacle prediction, includes the following steps: Step S21: Extract the distance and position of local obstacles based on the 3D obstacle model; Step S22: Use optical flow to convert the distance and position of local obstacles into motion vectors in three-dimensional space to predict the future position of the obstacles; Step S23: Predict the collision time, including the following steps: Step S231: Relative state definition, obtain the position and velocity of the inspection robot in the three-dimensional space at the current moment, combine the obstacle information marked on the global dynamic map and the predicted future position of the obstacles, calculate the relative position vector and relative velocity vector between the inspection robot and the local obstacles; Step S232: Shortest approach time calculation. Minimize the square of the distance between the robot and the local obstacle to calculate the shortest approach time under relative motion, which is the time before a potential collision occurs. Step S233: Collision time output. Predict the minimum distance between the inspection robot and the local obstacle at the shortest approach time. Set a safety threshold. If the minimum distance is less than the safety threshold, it is determined that there is a collision risk. The shortest approach time at this time is the predicted collision time.

Images