Self-adaptive safe path planning method for gas inspection robot

Abstract

The invention relates to a self-adaptive safe path planning method for a gas inspection robot, and belongs to the technical field of robot path planning. According to the method, firstly, a global path is generated by improving an A star algorithm, smoothness is optimized, then a multi-sensor data sensing environment is fused, local real-time obstacle avoidance planning is carried out by adopting a dynamic window method, and a deep reinforcement learning model is introduced to realize dynamic decision making. According to the method, the problems that in the prior art, in a dynamic complex environment, path planning is rigid, obstacle avoidance is not timely, and high-risk areas cannot be adaptively avoided are effectively solved, the efficiency and safety of path planning can be remarkably improved, and the autonomous operation capacity and the intelligent decision-making level of the robot in the underground complex environment are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of robot path planning, in particular to a gas inspection robot adaptive safety path planning method based on an improved A-star algorithm. BACKGROUND

[0002] With the rapid development of intelligent robot technology, inspection robots are increasingly widely used in complex environments such as power systems, substations, and mines. With the help of autonomous navigation technology, inspection robots can plan global paths in a pre-set environment and perform real-time obstacle avoidance, thereby significantly improving the efficiency and accuracy of inspection operations. However, in dynamic and complex environments such as underground mines, most existing inspection robots still have poor path planning flexibility and difficulty in responding to environmental changes in real time, which seriously restricts their reliability in performing tasks in dangerous places.

[0003] Traditional path planning algorithms, such as Dijkstra's algorithm and A-star algorithm, perform well in static known environments. However, in large-scale or complex scenarios, their computational complexity is high, requiring a large amount of computational resources, making it difficult to meet real-time requirements. In addition, such algorithms are usually static and cannot effectively handle dynamic obstacles. Although the rapid exploration random tree algorithm can quickly search for paths in non-convex dynamic environments, the paths generated by it are usually of poor quality, with problems such as twists and unevenness, and it is difficult to effectively balance global search and local search.

[0004] In order to overcome these limitations, researchers have begun to explore path planning solutions that combine traditional algorithms with emerging technologies. Among them, methods based on deep learning have received widespread attention due to their strong environmental perception and decision-making capabilities. For example, some solutions attempt to construct an environmental model by fusing lidar, ultrasonic, and other sensor data, and use deep learning models for online decision-making. However, how to deeply combine deep learning technology with classical path planning algorithms to achieve more efficient and safer autonomous movement remains a technical challenge that needs to be addressed. In particular, in high-risk applications such as gas inspection, robots not only need to avoid physical obstacles, but also need to respond to dangerous signals such as gas concentration exceeding limits in real time, which poses extremely high requirements for the environmental adaptability and decision-making intelligence of path planning algorithms.

[0005] The prior art such as the invention patent with publication number CN115079705A provides an optimization method based on improved A-star algorithm fusion dynamic window method, which improves the smoothness and planning efficiency of the path by optimizing the node search strategy and evaluation function. However, when dealing with completely unknown dynamic environment, the perception layer and decision layer of this method are still insufficient, especially in the aspects of multi-sensor information fusion and risk quantification. Another publication CN119596949A proposes an artificial intelligence-based inspection robot path planning scheme, which focuses on predicting road damage through visual sensors, but its applicability in complex underground environments and the reliability of real-time obstacle avoidance still need to be verified.

[0006] Therefore, it is urgent to develop a path planning method that can deeply fuse multi-source perception information and has online learning and adaptive decision-making capabilities to significantly improve the autonomy, safety and work efficiency of the gas inspection robot in dynamic, unstructured and high-risk environments.

[0007] In addition, existing fusion algorithms mostly make decisions based on fixed rules, lacking self-learning and adaptive capabilities. In extremely complex or never-before-seen scenarios, rule-based algorithms may not make optimal decisions, or even cause system failure. Especially for gas inspection scenarios, how to effectively fuse various heterogeneous sensor information such as laser radar, ultrasonic wave, gas concentration, and make intelligent decisions that can both avoid physical collisions and stay away from dangerous chemical risks is a difficult problem that existing technologies have not solved well. Deep learning, especially deep reinforcement learning technology, provides a new approach to solving decision-making problems in such complex environments, but how to efficiently integrate it with traditional path planning algorithms and deploy it on embedded mobile platforms still faces significant challenges. SUMMARY

[0008] In the prior art, there are problems such as rigid path planning of inspection robots in dynamic complex environments, inability to respond to environmental changes in real time, and high computational complexity and low efficiency of traditional path planning algorithms. Therefore, to solve the above problems, the present application provides a gas inspection robot adaptive safe path planning method based on improved A-star algorithm.

[0009] To achieve the above purpose, the present application provides the following technical solutions:

[0010] An adaptive safe path planning method for a gas inspection robot, comprising the following steps:

[0011] S1: initialize the state parameters and environment grid map of the inspection robot, wherein the state parameters include position, attitude, maximum linear velocity, and maximum angular velocity;

[0012] S2: Real-time perception of environmental information through multi-sensor fusion, including LIDAR, ultrasonic sensor and gas concentration sensor, and updating of the environmental grid map;

[0013] S3: Global path planning based on improved A-star algorithm, forming a preliminary global path, and through a greedy optimization strategy, removing redundant nodes and smoothing the preliminary global path to obtain an optimized global path;

[0014] S4: Extracting local key nodes from the optimized global path and performing local path planning and real-time obstacle avoidance based on the Dynamic Window Approach (DWA) optimization algorithm;

[0015] S5: Using a deep reinforcement learning model to make dynamic decisions on the local path planning results and output the final safe path;

[0016] S6: Controlling the inspection robot to move along the safe path and dynamically adjusting according to real-time environmental information.

[0017] Further, in S2, real-time perception of environmental information through multi-sensor fusion, specifically including:

[0018] Collecting obstacle distance and angle information through the LIDAR;

[0019] Supplementing the collection of close-range obstacle information through the ultrasonic sensor;

[0020] Collecting environmental gas concentration data through the gas concentration sensor;

[0021] Using Kalman filtering algorithm to fuse the multi-source sensor data and update the obstacle information and dangerous area information in the environmental grid map.

[0022] Further, in S3, the evaluation function of the improved A-star algorithm is:

[0023]

[0024] Where F(n) represents the comprehensive evaluation cost of node n, G(n) represents the actual cost from the starting point to node n, H(n) represents the estimated cost from node n to the end point, S sc represents the distance from the starting point to the current node, S st represents the distance from the starting point to the end point.

[0025] Further, in S3, the greedy optimization strategy specifically includes:

[0026] Calculate the vector angle θ1 between the current node S0, the child node S1 and the grandchild node S2:

[0027]

[0028] When θ1 = 0, delete the node S1;

[0029] When θ1 ≠ 0, calculate the line equation of the nodes S0 and S2, and determine the positional relationship between the line and the obstacle. If there is no collision risk, delete the node S1.

[0030] Further, in the S4, the evaluation function of the DWA optimization algorithm is:

[0031]

[0032] Wherein, G heading (v,w) represents the angle difference between the trajectory end direction and the local target point, G dist (v,w) represents the distance between the trajectory end and the local target point, S Kdist (v,w) represents the minimum distance between the trajectory end and the known static obstacle, S Udist (v,w) represents the minimum distance between the trajectory end and the unknown static obstacle, D Udist (v,w) represents the minimum distance between the trajectory end and the unknown dynamic obstacle, velocity(v,w) represents the current speed, α1, α2, β1, β2, β s , γ are weight coefficients, and σ represents normalization processing.

[0033] Further, the weight coefficient β s is an adaptive weight coefficient, and the value rule is:

[0034]

[0035] Wherein, D s represents the collision radius, R represents the turning radius of the robot, and β3 is a constant weight value.

[0036] Further, in the S5, the deep reinforcement learning model adopts a deep deterministic policy gradient algorithm (Deep Deterministic Policy Gradient, DDPG), and the input includes environment grid map information, robot state information and sensor data, and the output is an action policy and a value function.

[0037] Further, the reward function of the deep reinforcement learning model is designed as:

[0038] R t = ω1·R path + ω2·R safety+ω3·R efficiency

[0039] Among them, R path R represents the path-tracking reward. safety R represents a reward for safe obstacle avoidance. efficiency This represents the performance bonus, with ω1, ω2, and ω3 being the weighting coefficients.

[0040] Furthermore, in S6, the dynamic adjustment specifically includes:

[0041] When the Euclidean distance between the robot and the target point is less than the safety threshold, the robot moves directly toward the target point.

[0042] When the Euclidean distance between the robot and the target point is greater than the safety threshold, obstacle avoidance movement is performed based on the local cost map.

[0043] Furthermore, the deep reinforcement learning model employs an actor-critic network architecture; wherein the actor network (Actor-Network) uses a policy function μ(s|θ) μ This maps the observation information in the state space to the optimal action proposal in the continuous action space, where s represents the environmental state observation and θ... μ The weight parameters of the actor network are represented by the value function Q(a|θ). Q Evaluate the long-term expected reward of performing a given action in a specific state, where a represents the action vector in the action space, and θ Q The weight parameters represent the weight parameters of the critic network; the actor network structure is a fully connected layer, while the critic network structure adds a feature splicing layer before the fully connected layer to fuse state and action information; both networks are trained using a target network mechanism and an experience replay pool to improve the stability of the learning process.

[0044] Furthermore, the specific process of fusing multi-source sensor data using the Kalman filter algorithm includes: first, establishing a system state equation with the robot pose as the core and an observation equation with sensor observations as the core; then, using obstacle distance and angle information provided by LiDAR (Light Detection and Ranging), near-range blind spot information provided by ultrasonic sensors, and heading angle information provided by Inertial Measurement Unit (IMU) as observation inputs; next, performing a prediction step to predict the current state based on the state estimate and motion model from the previous moment; finally, performing an update step to optimally correct the predicted state using the sensor observations from the current moment, outputting the fused environmental state estimate and the corresponding uncertainty covariance, and updating the environmental grid map accordingly.

[0045] An adaptive safety path planning system for a gas inspection robot, used to implement the method described above, the system being integrated onto the inspection robot body, comprising:

[0046] The central processing unit is used to load and run the path planning algorithm;

[0047] A multi-sensor kit, electrically connected to the central processing unit, is used to collect environmental information in real time; the multi-sensor kit includes at least a LiDAR, an ultrasonic sensor, and a gas concentration sensor.

[0048] The motion control unit is electrically connected to the central processing unit and is used to receive speed control commands issued by the central processing unit and drive the chassis motor to complete the movement.

[0049] The navigation and positioning unit is used to provide the robot's real-time pose information;

[0050] The memory stores instructions executable by the central processing unit, which, when executed by the central processing unit, control the system to implement the following software functional modules:

[0051] The environment modeling module is used to initialize and update the environmental grid map in real time based on the data from the multi-sensor suite;

[0052] The global path planning module is used to run the improved A* algorithm to generate the globally optimal path;

[0053] The local obstacle avoidance planning module is used to run the Dynamic Window Method (DWA) optimization algorithm for real-time local path planning and obstacle avoidance.

[0054] The intelligent decision-making module is used to run deep reinforcement learning models and output dynamic decision-making strategies.

[0055] The control execution module is used to generate control commands and send them to the motion control unit.

[0056] The beneficial effects of this invention are as follows:

[0057] (1) By improving the node expansion strategy and evaluation function of the traditional A* algorithm and combining it with the greedy optimization method to remove redundant nodes, the generated global path is not only shorter and has fewer turns, but also has good continuity and trackability, making it more suitable for the actual movement of the robot.

[0058] (2) By using multi-sensor fusion technology to perceive environmental changes in real time, and by using the optimized dynamic window method local planning layer, the robot can sensitively and reliably avoid static and dynamic obstacles, effectively coping with the complexity and uncertainty of the underground environment.

[0059] (3) By introducing a gas concentration monitoring and hazardous area identification mechanism, the path planning process can not only take physical obstacles into account, but also autonomously avoid high-risk areas, which significantly improves the safety of the robot in gas environments.

[0060] (4) With the help of deep reinforcement learning models, robots can autonomously adjust their behavior strategies based on real-time environmental information, and have the ability to learn online and optimize dynamically, thereby making more reasonable and efficient travel decisions in unstructured environments.

[0061] (5) By extracting key nodes to guide local planning, the advantages of the global path are preserved, while the flexibility of local real-time adjustment is given, so that the robot can respond flexibly to local emergencies while taking into account overall efficiency.

[0062] (6) By improving the algorithm structure and data processing flow, the computational load is reduced while ensuring the planning quality, which is conducive to real-time operation on the embedded platform and broadens the application prospects of this technology in practical engineering.

[0063] (7) By adopting a deep reinforcement learning model with an actor-critic architecture and clarifying its network structure and training mechanism, the robot is equipped with the ability to make optimal decisions in a continuous action space. Its learning process is more stable, its convergence is better, and the final decision strategy generated is better.

[0064] (8) By clearly defining the multi-sensor fusion process based on Kalman filtering, an accurate system state estimation and update method was established, which significantly reduced the risk of false alarms and false alarms of a single sensor, provided a more reliable and accurate environmental perception data foundation for path planning, and improved the safety and reliability of the entire system from the source.

[0065] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0066] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0067] Figure 1 A diagram illustrating the greedy optimization strategy;

[0068] Figure 2 Here is the overall flowchart of the algorithm;

[0069] Figure 3 A comparison chart of global path planning algorithms;

[0070] Figure 4 A comparison chart of path planning for the fusion algorithms;

[0071] Figure 5 This is a sequence diagram of the physical verification process for the fusion algorithm. Detailed Implementation

[0072] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0073] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0074] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0075] Figure 1 This is a schematic diagram of the greedy optimization strategy of this invention. The diagram illustrates the geometric relationship between path nodes S0, S1, and S2, and the process of optimizing the path by deleting redundant nodes. This strategy effectively improves the smoothness of the path and execution efficiency.

[0076] Figure 2 This is a flowchart of the overall algorithm of this invention. The diagram illustrates the complete process from environment modeling, global planning, local optimization to motion execution, and shows the logical relationships and data flow between the modules.

[0077] Figure 3This is a comparison chart of global path planning algorithms. By comparing it with other algorithms, the comprehensive advantages of the algorithm of this invention in terms of path length, number of turns, and planning time are clearly demonstrated.

[0078] Figure 4 This is a comparison chart of path planning using fusion algorithms. The chart compares the performance of different fusion algorithms from multiple dimensions, highlighting the improvements in path quality and obstacle avoidance capabilities achieved by the algorithm of this invention.

[0079] Figure 5 This is a sequence diagram of the physical verification process of the fusion algorithm. The diagram details the interaction process between the robot and the control system in a real-world application scenario, including key aspects such as command transmission, status reporting, and anomaly handling.

[0080] Example 1: Basic Path Planning Based on the Fusion of Improved A-Star and Dynamic Window Method

[0081] First, the system is initialized. The initial position coordinates of the inspection robot are set to the origin, the attitude angle is zero, the maximum linear velocity is 0.5 meters per second, and the maximum angular velocity is 1 radian per second. At the same time, a 20-meter by 20-meter grid environment map is constructed, with a grid resolution of 0.1 meters. Dark grids in the map represent known static obstacles.

[0082] Next, the improved A* algorithm is executed for global path planning. For example... Figure 2 As shown, the overall algorithm flow begins with environment modeling and initialization, followed by the global path planning phase. After inputting the starting point and target point into the algorithm, an improved evaluation function is used for node expansion. This function introduces adaptive weight coefficients, increasing the heuristic function weights to speed up the search when the target is far away, and decreasing the weights to improve path quality when the target is close.

[0083] After completing the initial path planning, a greedy optimization strategy is used to smooth the path. For example... Figure 1 As shown, this strategy optimizes the path by analyzing the geometric relationships between three consecutive path nodes. It calculates the vector from the current node S0 to its child node S1, and the vector from child node S1 to its grandchild node S2, and calculates the angle θ1 between the two vectors. When the angle is zero, it indicates that the three points are collinear, and the intermediate node S1 is deleted. When the angle is not zero, it checks whether the line segment directly connecting S0 and S2 intersects with an obstacle. If there is no risk of collision, node S1 is deleted, thus simplifying and smoothing the path.

[0084] After obtaining the optimized global path, key nodes are extracted as local target points to guide the dynamic window method for local planning. The dynamic window method generates multiple candidate trajectories by sampling in the velocity space based on the robot's kinematic model and environmental perception information, and selects the optimal trajectory through a comprehensive evaluation function. This evaluation function comprehensively considers factors such as the distance between the trajectory endpoint and the local target, directional alignment, and distance to obstacles.

[0085] Finally, in the execution phase, the robot moves along the planned path and monitors environmental changes in real time. When an unknown obstacle is detected, local replanning is immediately triggered to generate an obstacle avoidance trajectory. After obstacle avoidance is completed, the robot automatically returns to the global path and continues its journey.

[0086] Example 2: Enhanced Path Planning Incorporating Intelligent Decision-Making

[0087] After completing system initialization and global path planning, this embodiment introduces a deep reinforcement learning model for intelligent decision-making. For example... Figure 4 As shown, compared with traditional algorithms, the fusion algorithm of this invention has significant improvements in both path quality and obstacle avoidance performance.

[0088] The multi-sensor information fusion module begins operation: lidar provides high-precision obstacle information, ultrasonic sensors handle close-range detection, and gas concentration sensors monitor the environmental safety status in real time. All this information is fused into a unified environmental model.

[0089] The deep reinforcement learning model employs a dual-deep Q-network structure, whose state space includes multiple dimensions such as robot pose, sensor readings, global path information, and local environmental features. The action space is defined as the robot's motion control commands. The reward function is carefully designed, comprising three parts: path tracking reward, obstacle avoidance reward, and efficiency reward.

[0090] like Figure 3 As shown in the comparison test with other global path planning algorithms, the algorithm of this invention has significant advantages in terms of path length, smoothness, and planning efficiency. Especially in complex environments, the algorithm of this invention can maintain better performance stability.

[0091] In actual operation, such as Figure 5 As shown, the robot interacts with the control platform through serialized instructions. After the control platform sends task instructions, the robot autonomously completes the entire process, including environmental perception, decision-making and planning, and motion execution, and reports status information in real time. When encountering emergencies, the robot can autonomously activate emergency response mechanisms to ensure operational safety.

[0092] Example 3: Complete System Implementation Based on Hardware-Software Collaboration

[0093] I. System Hardware Composition and Connection Relationships

[0094] The system is integrated into the inspection robot body, and its hardware architecture mainly includes:

[0095] Central Processing Unit: A heterogeneous computing platform consisting of an ARM-based processor and an NVIDIA Jetson series GPU. The CPU is responsible for tasks such as logic control and sensor driving, while the GPU accelerates inference computations for deep reinforcement learning models.

[0096] Multi-sensor kit: Electrically connected to the central processing unit via serial port or CAN bus. Includes: ① LiDAR, using SLAMTEC's RPLIDAR series, with a scanning range of no less than 12 meters, used to construct a 2D environmental map and provide obstacle distance information; ② Ultrasonic sensors, located around the robot body, with a detection range of 0.1 to 5 meters, used to detect short-range obstacles within the LiDAR blind zone; ③ Methane concentration sensor, using a catalytic combustion sensor, to monitor methane concentration in real time; ④ Navigation and positioning unit, including IMU and wheeled odometry, used to provide real-time pose estimation of the robot.

[0097] Motion control unit: It uses an STM32 series microcontroller as the lower-level controller, communicates with the central processing unit through a serial port, receives the speed control commands (v, w) issued by the central processing unit, and drives the DC servo motor to complete the movement.

[0098] Memory: including eMMC flash memory and DDR4 memory, used to store the operating system, environment map, and software and algorithm models executable by the central processing unit.

[0099] II. Implementation of Software Functional Modules

[0100] When instructions stored in memory are executed by the central processing unit, the system is controlled to implement the following modules:

[0101] Environmental Modeling Module: This module uses the Kalman filter algorithm for multi-sensor fusion. The specific process is as follows: First, based on the robot's pose X = [x, y, θ]... T And velocity U = [v, w] T Establish the system state equations for the core:

[0102] X t =F·X t-1 +B·U t +w

[0103] Where F is the state transition matrix, B is the control matrix, and U... t The variable is the control variable, and w is the process noise.

[0104] Then, establish the observation equation:

[0105] Z t =H·X t +v

[0106] Where H is the observation matrix, v is the observation noise, and Z is the observation input. t Obstacle distance from lidar, ultrasonic blind spot information, and heading angle from IMU.

[0107] Next, the prediction step is performed: based on the state estimate from the previous time step. Covariance P t-1 Predict the current state Covariance

[0108]

[0109] Where Q is the process noise covariance matrix.

[0110] Finally, perform the update step: calculate the Kalman gain K. t Then, using the current sensor observation value Z t After making corrections, the optimal state estimate is obtained. And update the estimated covariance P t :

[0111]

[0112] Where R is the observation noise covariance matrix and I is the identity matrix. The fused environmental state estimates are used to construct and update the environmental raster map.

[0113] Intelligent Decision-Making Module: This module utilizes a deep reinforcement learning model. The model employs an actor-critic architecture.

[0114] Actor Network: Its policy function is μ(s|θ) μ The input is observation information s in the state space (including sensor data, target point position, etc.), and the output is the optimal action suggestion a in the continuous action space (i.e., the robot's linear velocity and angular velocity (v, w)). The actor network consists of three fully connected layers, and the activation function is ReLU.

[0115] Commentator Network: Its value function is Q(a|θ) Q This network is used to evaluate the long-term expected reward of performing action a in state s. Its network structure concatenates the features of state s and action a after the input layer, and then inputs them into subsequent fully connected layers.

[0116] The weight parameters θ of the two networks μ and θ QAll updates are performed using stochastic gradient descent. To improve stability, a target network mechanism is introduced, which involves creating actor and critic target networks whose parameters are slowly synchronized from the main network via soft updates.

[0117] θ μ' ←τ·θ μ +(1-τ)·θ μ'

[0118] θ Q' ←τ·θ Q +(1-τ)·θ Q'

[0119] Where, θ μ' and θ Q' Let τ be the target network parameter, and τ be the update coefficient (usually τ << 1).

[0120] Simultaneously, the experience replay pool is used to store the transferred samples (s t ,a t ,r t ,s t+1 During training, random sampling is performed from the pool to break the correlation between data.

[0121] III. System Workflow

[0122] After the system is powered on, each module operates according to its process. At its core, the environment modeling module provides accurate environmental information for subsequent planning through Kalman filtering; the intelligent decision-making module makes advanced decisions through a deep reinforcement learning model; and finally, the control execution module sends instructions to the motion control unit to drive the physical motor, forming a complete perception-decision-control closed loop.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive safe path planning method for a gas inspection robot, characterized in that: The method comprises the following steps: S1: initializing state parameters of the inspection robot and an environment grid map, the state parameters comprising position, pose, maximum linear velocity, and maximum angular velocity; S2: perceiving environment information in real time through multi-sensor fusion, and updating the environment grid map, the multi-sensor comprising a laser radar (LIDAR), an ultrasonic sensor, and a gas concentration sensor; S3: performing global path planning based on an improved A-star algorithm, forming a preliminary global path, and performing redundant node elimination and smoothing processing on the preliminary global path through a greedy optimization strategy to obtain an optimized global path; S4: extracting local key nodes in the optimized global path, and performing local path planning and real-time obstacle avoidance based on a dynamic window method (DWA) optimization algorithm; S5: adopting a deep reinforcement learning model to make dynamic decisions on the local path planning result, and outputting a final safe path; S6: controlling the inspection robot to move along the safe path, and dynamically adjusting according to real-time environment information.

2. The adaptive safe path planning method of the gas inspection robot according to claim 1, characterized in that: In the S2, the environment information is perceived in real time through multi-sensor fusion, specifically comprising: collecting obstacle distance and angle information through the LIDAR; supplementarily collecting close-range obstacle information through the ultrasonic sensor; collecting environment gas concentration data through the gas concentration sensor; adopting a Kalman filtering algorithm to fuse the multi-source sensor data, and updating obstacle information and dangerous area information in the environment grid map. 3.The adaptive safe path planning method of the gas inspection robot according to claim 1, characterized in that: In the S3, the evaluation function of the improved A-star algorithm is: Wherein, F(n) represents the comprehensive evaluation cost of node n, G(n) represents the actual cost from the starting point to node n, H(n) represents the estimated cost from node n to the end point, S sc represents the distance from the starting point to the current node, S st represents the distance from the starting point to the end point.

4. The adaptive safe path planning method of the gas inspection robot according to claim 3, characterized in that: In the S3, the greedy optimization strategy specifically comprises: calculating a vector angle θ1 between a current node S0, a child node S1, and a grandchild node S2: when θ1 = 0, deleting the node S1; when θ1 ≠ 0, calculating a line equation of the nodes S0 and S2, and judging a positional relationship between the line and an obstacle, if there is no collision risk, deleting the node S1.

5. The adaptive safe path planning method of the gas inspection robot according to claim 1, characterized in that: In the S4, the evaluation function of the DWA optimization algorithm is: where G heading (v,w) represents the angle difference between the trajectory end direction and the local target point, G dist (v,w) represents the distance between the trajectory end and the local target point, S Kdist (v,w) represents the minimum distance between the trajectory end and the known static obstacle, S Udist (v,w) represents the minimum distance between the trajectory end and the unknown static obstacle, D Udist (v,w) represents the minimum distance between the trajectory end and the unknown dynamic obstacle, velocity (v,w) represents the current speed, α1, α2, β1, β2, β s , γ are weight coefficients, and σ represents normalization processing. 6.The adaptive safe path planning method of the gas inspection robot according to claim 5, characterized in that: The weight coefficient β s is an adaptive weight coefficient, and the value rule is: where D s represents the conflict radius, R represents the turning radius of the robot, and β3is a constant weight value.

7. The adaptive safe path planning method of the gas inspection robot according to claim 1, characterized in that: In the S5, the deep reinforcement learning model adopts a deep deterministic policy gradient (DDPG) algorithm, the input of which comprises environment grid map information, robot state information, and sensor data, and the output of which is an action policy and a value function. 8.The adaptive safe path planning method of the gas inspection robot according to claim 7, characterized in that: The reward function of the deep reinforcement learning model is designed as: R t = ω1 · R path + ω2 · R safety + ω3 · R efficiency wherein R path represents the path tracking reward, R safety represents the safety obstacle avoidance reward, R efficiency represents the running efficiency reward, and ω1, ω2, and ω3 are weight coefficients. 9.The adaptive safe path planning method of the gas inspection robot according to claim 1, wherein: In the S6, the dynamic adjustment specifically comprises: when the Euclidean distance between the robot and a target point is less than a safety threshold, directly moving to the target point; when the Euclidean distance between the robot and the target point is greater than the safety threshold, performing obstacle avoidance movement according to a local cost map.

10. The adaptive safety path planning method of the gas inspection robot according to claim 1 or 7, characterized in that: The deep reinforcement learning model adopts an actor-critic network architecture; wherein an actor network Actor-Network maps observation information in a state space to an optimal action suggestion in a continuous action space through a policy function μ(s|θ μ ), wherein s represents an environment state observation value, and θ μ represents a weight parameter of the actor network; a critic network Critic-Network evaluates a long-term expected return of performing a given action in a specific state through a value function Q(a|θ Q ), wherein a represents an action vector in an action space, and θ Q represents a weight parameter of the critic network; the actor network structure is a fully connected layer, and the critic network structure is increased with a feature splicing layer before the fully connected layer to fuse state and action information; both networks are trained using a target network mechanism and an experience replay pool to improve the stability of the learning process.

11. The adaptive safe path planning method of the gas inspection robot according to claim 2, characterized in that: The specific process of fusing multi-source sensor data by using the Kalman filtering algorithm comprises the following steps: firstly, a system state equation with a robot pose as a core and an observation equation with a sensor observation value as a core are established; secondly, obstacle distance and angle information provided by a laser radar (LIDAR), near-distance blind-filling information provided by an ultrasonic sensor and a heading angle provided by an inertial measurement unit (IMU) are taken as observation inputs; thirdly, a prediction step is performed to predict a current state according to a state estimation at a previous moment and a motion model; finally, an update step is performed to optimally correct the predicted state by using sensor observation values at a current moment, to output fused environment state estimation values and corresponding uncertainty covariance, and to update an environment grid map according to the fused environment state estimation values and the corresponding uncertainty covariance.

12. An adaptive safe path planning system for a gas inspection robot, configured to implement the method of any one of claims 1 to 11. The system is integrated on a patrol robot body and comprises: a central processing unit configured to load and run the path planning algorithm; a multi-sensor suite electrically connected to the central processing unit and configured to collect environment information in real time; the multi-sensor suite at least comprises a laser radar (LIDAR), an ultrasonic sensor and a gas concentration sensor; a motion control unit electrically connected to the central processing unit and configured to receive a speed control instruction issued by the central processing unit and to drive a chassis motor to move; a navigation positioning unit configured to provide real-time pose information of the robot; a memory configured to store instructions executable by the central processing unit, and the instructions, when executed by the central processing unit, control the system to implement the following software function modules: an environment modeling module configured to initialize and update an environment grid map in real time according to data of the multi-sensor suite; a global path planning module configured to run an improved A-star algorithm to generate a globally optimal path; a local obstacle avoidance planning module configured to run a dynamic window method (DWA) optimization algorithm to perform real-time local path planning and obstacle avoidance; an intelligent decision-making module configured to run a deep reinforcement learning model to output a dynamic decision-making strategy; a control execution module configured to generate a control instruction and issue the control instruction to the motion control unit.

Images