Multi-sensor fusion based intelligent navigation detection system for wheeled inspection robot

Abstract

The application discloses a wheeled inspection robot intelligent navigation detection system based on multi-sensor fusion, relates to the technical field of industrial automation inspection, and comprises a sensor layer, a data fusion layer, a path planning layer, a navigation correction layer and an execution layer which are connected in sequence; the sensor layer collects multi-dimensional environment and robot state data; the data fusion layer adopts an improved set Kalman filtering algorithm to process multi-source data; the path planning layer generates static and dynamic paths based on a GA-PSO fusion algorithm, and the path contains path reference nodes; the navigation correction layer outputs speed and steering angle correction amounts through a DQN algorithm; and the execution layer drives an execution component through PID control; the system provided by the application improves navigation precision and dynamic obstacle adaptability, is suitable for complex inspection scenes such as transformer substations, and solves the problems of large navigation deviation and slow response of traditional inspection robots.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial automation intelligent navigation inspection, and particularly to a wheeled inspection robot intelligent navigation detection system based on multi-sensor fusion. BACKGROUND

[0002] Wheeled inspection robots are widely used in outdoor scenes such as substations and chemical parks due to their high flexibility and strong endurance, and replace manual work to complete tasks such as equipment state detection and environmental monitoring. However, the existing navigation system of the wheeled inspection robot has many technical bottlenecks and cannot meet the precise navigation requirements in complex scenes.

[0003] Firstly, the data collected by a single sensor has limitations: the laser radar is easily affected by bad weather, resulting in point cloud noise; the visual camera has low target recognition accuracy in insufficient light environment; the inertial measurement unit (IMU) will produce cumulative errors after long-term use; and relying on a single type of sensor cannot stably output the global position of the robot and the environmental obstacle information, resulting in large navigation positioning deviation and difficulty in meeting the high-precision inspection requirement of ≤0.2m.

[0004] Secondly, the performance of the multi-sensor data fusion algorithm is insufficient: the traditional Kalman filter algorithm is prone to filter divergence when processing nonlinear multi-source data, and does not consider false correlation between data, resulting in low reliability of the fused data and inability to provide accurate environment and robot state basis for path planning, further exacerbating navigation deviation.

[0005] Thirdly, the path planning and navigation correction capability is weak: the existing path planning algorithm mostly uses a single genetic algorithm or particle swarm optimization algorithm, which has low static path generation efficiency, and when facing dynamic obstacles (such as temporary maintenance tools and personnel), the real-time re-planning response time is long (usually >2s), which is prone to collision risk; at the same time, the navigation correction mostly relies on simple proportional control, which cannot dynamically adjust the speed and steering angle according to the comprehensive tracking error, resulting in the robot deviating from the trajectory when traveling along the planned path.

[0006] In addition, the control precision of the execution layer is insufficient: the traditional execution layer uses open-loop control or simple PID control, which has large speed and steering angle tracking errors and cannot quickly respond to navigation correction instructions, further reducing the stability and precision of the overall navigation system. Therefore, there is an urgent need for a wheeled inspection robot intelligent navigation detection system that integrates the advantages of multiple sensors, optimizes algorithm performance, and improves control precision to solve the above technical problems. SUMMARY

[0007] The present application relates to the technical field of industrial automation intelligent navigation inspection, and particularly to a wheeled inspection robot intelligent navigation detection system based on multi-sensor fusion.

[0008] To achieve the above objectives, the present invention provides an intelligent navigation and detection system for a wheeled inspection robot based on multi-sensor fusion, comprising a sensor layer, a data fusion layer, a path planning layer, a navigation correction layer, and an execution layer connected in sequence.

[0009] The sensor layer collects environmental 3D point cloud, images, robot posture, motion speed, and environmental temperature and humidity data and transmits them synchronously to the data fusion layer;

[0010] The data fusion layer uses an improved ensemble Kalman filter algorithm to process multi-source data, fuse multi-sensor data, and output the robot's global position, attitude, and coordinates of environmental obstacles;

[0011] The path planning layer is based on a fusion algorithm of improved genetic algorithm and particle swarm optimization. It generates a static optimal path by combining the fused obstacle information and triggers real-time replanning when dynamic obstacles appear, outputting a dynamic path.

[0012] The navigation correction layer uses a deep reinforcement learning algorithm to output speed correction and steering angle correction based on the robot's real-time state and the overall tracking error.

[0013] The execution layer receives correction instructions and drives the DC motor, servo motor, and electromagnetic brake through the PID controller to achieve precise navigation of the robot along the planned path.

[0014] Preferably, the sensor layer includes a lidar, a vision camera, an inertial measurement unit (IMU), an odometer, and a temperature and humidity sensor;

[0015] LiDAR is used for obstacle recognition and localization: by emitting a laser beam to scan the environment and receiving the reflected light signals, it calculates three-dimensional point cloud data to obtain the robot's global position. and coordinates of environmental obstacles ( , );

[0016] Visual cameras are used for target recognition: after acquiring environmental images, the target coordinates are output through the YOLOv8 target detection algorithm;

[0017] Inertial Measurement Unit (IMU) is used to acquire robot attitude: robot attitude angles are acquired via a three-axis accelerometer and a three-axis gyroscope. By combining acceleration and velocity, the robot's posture can be obtained. For roll angle, The pitch angle, Yaw angle;

[0018] The odometer is integrated into the drive wheels to measure distance and speed. ;

[0019] Temperature and humidity sensors are used for environmental monitoring to obtain ambient temperature and humidity data.

[0020] All sensors transmit data via RS485 / Ethernet interfaces with a unified timestamp and in JSON format.

[0021] The preferred improved Kalman filter algorithm for the data fusion layer is as follows:

[0022] Step 1: Initialize the set: Set the fusion state vector , Indicates the robot's global position. The coordinates are eastward. The coordinates are north-facing. For height, Indicates the robot's attitude angle. For roll angle, The pitch angle, Yaw angle This represents the velocity; the initial set state is the sensor's initial measurements superimposed with a Gaussian perturbation, expressed as: , In the formula, Let i be the state vector of the i-th member of the set at the initial time. This is the mean vector of sensor measurements at the initial moment. Let be the initial perturbation covariance matrix. The symbol for Gaussian distribution;

[0023] Step 2: Based on the differential motion physical model of the wheeled robot, predict the state of each set member, as shown in the following expression:

[0024] ;

[0025] In the formula, for Time of the first The predicted state vector of each set member, This is a differential motion physical model function for a wheeled robot. for The robot control input vector at time 1. for Time of the first The analytical state vector of each set member, The model error covariance matrix;

[0026] Step 3: Calculate the forecast ensemble mean and forecast error covariance matrix, and calculate the Kalman gain. The specific process is as follows:

[0027] S31. Calculate the mean of the forecast ensemble and the covariance matrix of the forecast error to quantify the forecast uncertainty. The expression is as follows:

[0028] ;

[0029] ;

[0030] S32, Introduction Time observation vector Observation matrix H, observation error covariance matrix R;

[0031] S33. Calculate the Kalman gain and determine the correction weights of the observed data for the prediction. The expression is as follows:

[0032] ;

[0033] In the formula, express The mean vector of the time forecast set; Indicates the total number of members in the set; express Time forecast error covariance matrix; express The Kalman gain matrix at each time step;

[0034] Step 4: Analyze and update the set members, and output the mean of the merged states, as shown in the following expression:

[0035] ;

[0036] In the formula, express Time of the first The analysis state vector of each set member;

[0037] Step 5: Calculate the mean of the analysis set The first three components are the global position. Points in the processed LiDAR point cloud data that are less than a set threshold are selected as the coordinates of environmental obstacles. ).

[0038] Preferably, the forecast error covariance matrix in step 3 is localized to avoid spurious correlations during the fusion process. The expression is as follows:

[0039] ;

[0040] In the formula, Indicates the localization factor; For the fusion state vector The The component and the first The correlation distance of each component; Indicates the localization radius;

[0041] Localized forecast error covariance matrix The expression is as follows:

[0042] ;

[0043] In the formula, For Schuler products, This represents a localized matrix.

[0044] Preferably, the output process of the dynamic path in the path planning layer is as follows:

[0045] A1. Construct an environment map using the grid method, with obstacle grids marked as 1 and free grids marked as 0;

[0046] B1. The static optimal path containing path reference nodes is generated by the Particle Swarm Optimization (PSO) algorithm.

[0047] C1. The static optimal path with path reference nodes generated in B1 is optimized by the genetic GA algorithm to generate a dynamic path with updated path reference nodes.

[0048] Preferably, the process of generating the static optimal path with path reference nodes using the PSO algorithm in B1 is as follows:

[0049] B11. Initialize the particle swarm: Set the number of particles and the maximum number of iterations. Each particle corresponds to a candidate path, and the path consists of a series of continuous grid nodes.

[0050] B12. Iteratively update particle velocity and position, optimize the distribution of path reference nodes, and realize path search. The expression is as follows:

[0051] ;

[0052] ;

[0053] In the formula, Indicates the first During the nth iteration The velocity vector of each particle reflects the speed and direction of the particle's movement in the solution space; Indicates inertia weight; Indicates the first During the nth iteration The velocity vector of each particle; This represents the particle's own cognitive coefficient, which controls the weight of the particle's tendency to move closer to its historical best position; , Both represent random numbers within the interval [0,1], increasing the randomness of the algorithm's search and avoiding getting trapped in local optima; Indicates the first The historical optimal position of each particle; Indicates the first During the nth iteration The current position of each particle; It represents the global social cognition coefficient, which controls the weight of particles to move closer to the best position in the history of the entire particle swarm; Indicates the globally optimal position; Indicates the first During the nth iteration The position of each particle;

[0054] B13. Based on the fitness function Calculate the fitness value of each particle and select the best historical position for the current particle. with the global optimal position ;

[0055] B14. Output static optimal path: After the maximum number of iterations is set, the path is output at the globally optimal position. The corresponding path is the static optimal path. All nodes on this path are static path reference nodes; this path satisfies the conditions of being collision-free, having the shortest path length, and having optimal smoothness.

[0056] Preferably, in C1, the GA algorithm optimizes the static optimal path with path reference nodes generated in B1 to generate a dynamic path with updated path reference nodes, as follows:

[0057] C11. Construct the initial population; use the static optimal path output by the PSO algorithm. Based on this, an initial population is generated through random perturbations (such as replacing a small number of grid nodes and fine-tuning path segments);

[0058] C12. Optimize the preset number of iterations through selection, crossover, and mutation operations;

[0059] C13. Output dynamic path; after the preset number of iterations, select the path with the highest fitness value as the dynamic path. The nodes on this path are the updated path reference nodes. When new dynamic obstacles appear in the environment, the above-mentioned GA algorithm replanning process is triggered to update the path reference nodes in real time.

[0060] Preferably, the fitness function of the GA-PSO fusion algorithm for:

[0061] ;

[0062] in, , , ; , , These are the path length weight, path smoothness weight, and obstacle avoidance success rate weight, respectively. To normalize the path length, To normalize path smoothness, The success rate of overshoot avoidance is 1 if the path and obstacle grid do not overlap, and 0 if they overlap.

[0063] Preferably, the calculation process for the overall tracking error is as follows:

[0064] A2. Calculate the horizontal position error The deviation of the robot's trajectory in the plane is reflected by the Euclidean distance between the robot's current horizontal position and the path reference node, as expressed in the following expression:

[0065] ;

[0066] in, , They are respectively The east and north coordinates after time merging , These are the east and north coordinates of the path reference node, respectively;

[0067] B2. Calculate heading error The deviation in the driving direction is reflected by the angular deviation between the robot's current heading and the reference heading of the path, as expressed in the following expression:

[0068] ;

[0069] in, for Yaw angle after time-integration The reference heading for the path reference node;

[0070] C2. Calculate the overall tracking error The weighted fusion of horizontal position error and heading error is used as input to the DQN algorithm, and the expression is:

[0071] ;

[0072] in, ; For horizontal position error weights, This represents the heading error weight.

[0073] Preferably, the execution layer includes a DC motor, a servo motor, and an electromagnetic brake, and a PID controller is used to achieve speed and steering angle tracking, as shown in the following expression:

[0074] ;

[0075] in, This is the proportionality coefficient. The integral coefficient is... These are the differential coefficients. For the desired speed, For actual speed, To track speed.

[0076] Therefore, the intelligent navigation and detection system for wheeled inspection robots based on multi-sensor fusion described above has the following beneficial effects:

[0077] (1) Improve the reliability of multi-sensor data fusion: By improving the ensemble Kalman filter algorithm and introducing localization processing to eliminate false correlations in data, the problem of divergence in traditional filtering is solved;

[0078] (2) Optimize path planning efficiency and dynamic adaptability: The GA-PSO (genetic-particle swarm) fusion algorithm realizes the rapid generation of static paths, reduces the response time of dynamic obstacle replanning, improves the obstacle avoidance success rate, solves the problems of slow path generation and poor dynamic adaptability of traditional algorithms, and reduces the risk of collision.

[0079] (3) Improve navigation correction accuracy: The DQN algorithm based on comprehensive tracking error can dynamically output speed and steering angle correction, avoiding trajectory deviation caused by traditional proportional control, and ensuring that the robot travels along the planned path;

[0080] (4) Enhance the stability of the execution layer control: dual PID control of speed and steering angle further improves the overall stability of the navigation system.

[0081] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0082] Fig. 1 This is an overall block diagram of the intelligent navigation and detection system for a wheeled inspection robot based on multi-sensor fusion, as described in this invention.

[0083] Fig. 2 This is a comparison of the dynamic obstacle replanning response time curves of the present invention and traditional robot navigation as a function of the number of obstacles in an embodiment of the present invention.

[0084] Fig. 3 This is a comparison chart showing the tracking error of the present invention and traditional robot navigation as a function of navigation time, according to an embodiment of the present invention. Detailed Implementation

[0085] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0086] Please see Figs. 1-3 The intelligent navigation and detection system for wheeled inspection robots based on multi-sensor fusion includes a sensor layer, a data fusion layer, a path planning layer, a navigation correction layer, and an execution layer connected in sequence.

[0087] The sensor layer collects environmental 3D point cloud, images, robot posture, motion speed, and environmental temperature and humidity data and transmits them synchronously to the data fusion layer;

[0088] The sensor layer includes LiDAR, vision camera, inertial measurement unit (IMU), odometer, and temperature and humidity sensors;

[0089] LiDAR is used for obstacle recognition and localization: by emitting a laser beam to scan the environment and receiving the reflected light signals, it calculates three-dimensional point cloud data to obtain the robot's global position. and coordinates of environmental obstacles ( , );

[0090] Visual cameras are used for target recognition: after acquiring environmental images, the target coordinates are output through the YOLOv8 target detection algorithm;

[0091] Inertial Measurement Unit (IMU) is used to acquire robot attitude: robot attitude angles are acquired via a three-axis accelerometer and a three-axis gyroscope. By combining acceleration and velocity, the robot's posture can be obtained. For roll angle, The pitch angle, Yaw angle;

[0092] The odometer is integrated into the drive wheels to measure distance and speed. It uses Magene model C706;

[0093] Temperature and humidity sensors are used for environmental monitoring: Texas Instruments (TI) model HDC2010 is used to obtain environmental temperature and humidity data;

[0094] All sensors transmit data via RS485 / Ethernet interfaces with a unified timestamp and in JSON format.

[0095] The data fusion layer uses an improved ensemble Kalman filter algorithm to process multi-source data, fuse multi-sensor data, and output the robot's global position, attitude, and coordinates of environmental obstacles;

[0096] The details of the improved ensemble Kalman filter algorithm for the data fusion layer are as follows:

[0097] Step 1: Initialize the set: Set the fusion state vector , Indicates the robot's global position. The coordinates are eastward. The coordinates are north-facing. For height, Indicates the robot's attitude angle. For roll angle, The pitch angle, Yaw angle This represents the velocity; the initial set state is the sensor's initial measurements superimposed with a Gaussian perturbation, expressed as: , In the formula, Take 30, Let i be the state vector of the i-th member of the set at the initial time. This is the mean vector of sensor measurements at the initial moment. Let be the initial perturbation covariance matrix. The symbol for Gaussian distribution;

[0098] Step 2: Based on the differential motion physical model of the wheeled robot, predict the state of each set member, as shown in the following expression:

[0099] ;

[0100] In the formula, for Time of the first The predicted state vector of each set member, This is a differential motion physical model function for a wheeled robot. for The robot control input vector at time 1. for Time of the first The analytical state vector of each set member, The model error covariance matrix;

[0101] Step 3: Calculate the forecast ensemble mean and forecast error covariance matrix, and calculate the Kalman gain. The specific process is as follows:

[0102] S31. Calculate the mean of the forecast ensemble and the covariance matrix of the forecast error to quantify the forecast uncertainty. The expression is as follows:

[0103] ;

[0104] ;

[0105] S32, Introduction Time observation vector Observation matrix H, observation error covariance matrix R;

[0106] S33. Calculate the Kalman gain and determine the correction weights of the observed data for the prediction. The expression is as follows:

[0107] ;

[0108] In the formula, express The mean vector of the time forecast set; Indicates the total number of members in the set; express Time forecast error covariance matrix; express The Kalman gain matrix at each time step;

[0109] The prediction error covariance matrix in step 3 is localized to avoid spurious correlations during the fusion process. The expression is as follows:

[0110] ;

[0111] In the formula, Indicates the localization factor; For the fusion state vector The The component and the first The correlation distance of each component; Indicates the localization radius;

[0112] Localized forecast error covariance matrix The expression is as follows:

[0113] ;

[0114] In the formula, For Schuler products, This represents a localized matrix.

[0115] Step 4: Analyze and update the set members, and output the mean of the merged states, as shown in the following expression:

[0116] ;

[0117] In the formula, express Time of the first The analysis state vector of each set member;

[0118] Step 5: Calculate the mean of the analysis set The first three components are the global position. Points in the processed LiDAR point cloud data that are less than a set threshold are selected as the coordinates of environmental obstacles. , ).

[0119] The path planning layer is based on a fusion algorithm of improved genetic algorithm and particle swarm optimization. It generates a static optimal path by combining the fused obstacle information and triggers real-time replanning when dynamic obstacles appear, outputting a dynamic path.

[0120] The output process of dynamic paths in the path planning layer is as follows:

[0121] A1. Construct an environment map using the grid method, with obstacle grids marked as 1 and free grids marked as 0;

[0122] B1. Generate the static optimal path containing path reference nodes using the PSO algorithm; the specific process is as follows:

[0123] B11. Initialize the particle swarm: Set the number of particles and the maximum number of iterations. Each particle corresponds to a candidate path, and the path consists of a series of continuous grid nodes.

[0124] B12. Iteratively update particle velocity and position, optimize the distribution of path reference nodes, and realize path search. The expression is as follows:

[0125] ;

[0126] ;

[0127] In the formula, Indicates the first During the nth iteration The velocity vector of each particle reflects the speed and direction of the particle's movement in the solution space; Indicates inertia weight; Indicates the first During the nth iteration The velocity vector of each particle; This represents the particle's own cognitive coefficient, which controls the weight of the particle's tendency to move closer to its historical best position; , Both represent random numbers within the interval [0,1], increasing the randomness of the algorithm's search and avoiding getting trapped in local optima; Indicates the first The historical optimal position of each particle; Indicates the first During the nth iteration The current position of each particle; It represents the global social cognition coefficient, which controls the weight of particles to move closer to the best position in the history of the entire particle swarm; Indicates the globally optimal position; Indicates the first During the nth iteration The position of each particle;

[0128] B13. Based on the fitness function Calculate the fitness value of each particle and select the best historical position for the current particle. with the global optimal position Fitness function The expression is as follows:

[0129] ;

[0130] in, , , ; , , These are the path length weight, path smoothness weight, and obstacle avoidance success rate weight, respectively. To normalize the path length, To normalize path smoothness, The success rate of overshoot avoidance is 1 if the path and obstacle grid do not overlap, and 0 if they overlap.

[0131] B14. Output static optimal path: After the maximum number of iterations is set, the path is output at the globally optimal position. The corresponding path is the static optimal path. All nodes on this path are static path reference nodes; this path satisfies the conditions of being collision-free, having the shortest path length, and having optimal smoothness.

[0132] C1. The static optimal path with path reference nodes generated in B1 is optimized using the GA algorithm to generate a dynamic path with updated path reference nodes; the specific process is as follows:

[0133] C11. Construct the initial population; use the static optimal path output by the PSO algorithm. Based on this, an initial population is generated through random perturbations (such as replacing a small number of grid nodes and fine-tuning path segments);

[0134] C12. Optimize the preset number of iterations through selection, crossover, and mutation operations;

[0135] C13. Output dynamic path; after the preset number of iterations, select the path with the highest fitness value as the dynamic path. The nodes on this path are the updated path reference nodes. When new dynamic obstacles appear in the environment, the above-mentioned GA algorithm replanning process is triggered to update the path reference nodes in real time.

[0136] This embodiment compares the dynamic obstacle replanning response time of the traditional navigation curve and the navigation curve of the present invention with the variation of the number of obstacles. Fig. 2 As shown, for traditional navigation curves, the response time increases linearly with the number of obstacles, reaching 4.8s with 5 obstacles. Due to the use of a single path planning algorithm, the replanning calculation efficiency is low. In contrast, the navigation curve of this invention shows a gradual growth trend, with a response time of only 0.9s with 5 obstacles. This is because the GA-PSO fusion algorithm uses static paths as the initial population, which greatly shortens the replanning iteration time.

[0137] The navigation correction layer uses a deep reinforcement learning algorithm to output speed correction and steering angle correction based on the robot's real-time state and the overall tracking error.

[0138] The calculation process for the overall tracking error is as follows:

[0139] A2. Calculate the horizontal position error The deviation of the robot's trajectory in the plane is reflected by the Euclidean distance between the robot's current horizontal position and the path reference node, as expressed in the following expression:

[0140] ;

[0141] in, , They are respectively The east and north coordinates after time merging , These are the east and north coordinates of the path reference node, respectively;

[0142] B2. Calculate heading error The deviation in the driving direction is reflected by the angular deviation between the robot's current heading and the reference heading of the path, as expressed in the following expression:

[0143] ;

[0144] in, for Yaw angle after time-integration The reference heading for the path reference node;

[0145] C2. Calculate the overall tracking error The weighted fusion of horizontal position error and heading error is used as input to the DQN algorithm, and the expression is:

[0146] ;

[0147] in, ; For horizontal position error weights, This represents the heading error weight.

[0148] This embodiment compares the error magnitudes of traditional navigation and the navigation of the present invention, such as... Fig. 3 As shown, the traditional navigation curve shows a continuous upward trend with increasing time, and the error reaches 0.52 at 250s; the navigation curve of the present invention shows a stable low-amplitude fluctuation trend, and the error is only 0.16 at 250s, which can compensate for trajectory deviation in real time.

[0149] The execution layer receives correction instructions and drives the DC motor, servo motor, and electromagnetic brake through the PID controller to achieve precise navigation of the robot along the planned path.

[0150] The execution layer includes a DC motor, a servo motor, and an electromagnetic brake. A PID controller is used to achieve speed and steering angle tracking, as shown in the following expression:

[0151] ;

[0152] in, This is the proportionality coefficient. The integral coefficient is... These are the differential coefficients. For the desired speed, For actual speed, To track speed.

[0153] Therefore, this invention employs the aforementioned intelligent navigation and detection system for wheeled inspection robots based on multi-sensor fusion, comprising a sensor layer, a data fusion layer, a path planning layer, a navigation correction layer, and an execution layer connected in sequence. The sensor layer synchronously collects environmental and robot state data from multiple devices. The data fusion layer uses an improved ensemble Kalman filter (including localization processing) to fuse multi-source data, outputting accurate position, attitude, and obstacle coordinates. The path planning layer, based on the GA-PSO fusion algorithm, generates static / dynamic paths with reference nodes, and replans in real time when encountering dynamic obstacles. The navigation correction layer uses the DQN algorithm to output speed and steering angle corrections based on the comprehensive tracking error. The execution layer drives the execution components with dual PID control to achieve precise navigation. Through the collaboration of each layer, the system solves the problems of large navigation deviations and slow response in traditional navigation systems.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. 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 still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent navigation and detection system for a wheeled inspection robot based on multi-sensor fusion, characterized in that: It includes a sensor layer, a data fusion layer, a path planning layer, a navigation correction layer, and an execution layer connected in sequence; The sensor layer collects environmental 3D point cloud, images, robot posture, motion speed, and environmental temperature and humidity data and transmits them synchronously to the data fusion layer; The data fusion layer employs an improved ensemble Kalman filter algorithm to process multi-source data, fusing multi-sensor data to output the robot's global position, attitude, and environmental obstacle coordinates. The improved ensemble Kalman filter algorithm includes: initializing the ensemble; predicting the state of each ensemble member based on the differential motion physical model of the wheeled robot; calculating the predicted ensemble mean and prediction error covariance matrix, and calculating the Kalman gain; analyzing and updating the ensemble members, and outputting the fused state mean; finally, outputting the robot's global position, attitude, and environmental obstacle coordinates. The prediction error covariance matrix is ​​localized to avoid spurious correlations during the fusion process; the expression is as follows: ； In the formula, Indicates the localization factor; For the fusion state vector The The component and the first The correlation distance of each component; Indicates the localization radius; Localized forecast error covariance matrix The expression is as follows: ； In the formula, For Schuler products, Represents a localized matrix. express Time forecast error covariance matrix; The path planning layer is based on a fusion algorithm of improved genetic algorithm and particle swarm optimization. It generates a static optimal path by combining the fused obstacle information and triggers real-time replanning when dynamic obstacles appear, outputting a dynamic path. The navigation correction layer uses a deep reinforcement learning algorithm to output speed correction and steering angle correction based on the robot's real-time state and the overall tracking error. The execution layer receives correction instructions and drives the DC motor, servo motor and electromagnetic brake through the PID controller to achieve precise navigation of the robot along the planned path; The output process of dynamic paths in the path planning layer is as follows: A1. Construct an environment map using the grid method, with obstacle grids marked as 1 and free grids marked as 0; B1. Generate the static optimal path with path reference nodes using the PSO algorithm; C1. The static optimal path with path reference nodes generated in B1 is optimized by the GA algorithm to generate a dynamic path with updated path reference nodes. The process of generating the static optimal path with path reference nodes using the PSO algorithm in B1 is as follows: B11. Initialize the particle swarm: Set the number of particles and the maximum number of iterations. Each particle corresponds to a candidate path, and the path consists of a series of continuous grid nodes. B12. Iteratively update particle velocity and position, optimize the distribution of path reference nodes, and realize path search. The expression is as follows: ； ； In the formula, Indicates the first During the nth iteration The velocity vector of each particle reflects the speed and direction of the particle's movement in the solution space; Indicates inertia weight; Indicates the first During the nth iteration The velocity vector of each particle; This represents the particle's own cognitive coefficient, which controls the weight of the particle's tendency to move closer to its historical best position; , Both represent random numbers within the interval [0,1], increasing the randomness of the algorithm's search and avoiding getting trapped in local optima; Indicates the first The historical optimal position of each particle; Indicates the first During the nth iteration The current position of each particle; It represents the global social cognition coefficient, which controls the weight of particles to move closer to the best position in the history of the entire particle swarm; Indicates the globally optimal position; Indicates the first During the nth iteration The position of each particle; B13. Based on the fitness function Calculate the fitness value of each particle and select the best historical position for the current particle. with the global optimal position ; B14. Output static optimal path: After the maximum number of iterations is set, the path is output at the globally optimal position. The corresponding path is the static optimal path. All nodes on this path are static path reference nodes; The calculation process for the overall tracking error is as follows: A2. Calculate the horizontal position error The deviation of the robot's trajectory in the plane is reflected by the Euclidean distance between the robot's current horizontal position and the path reference node, as expressed in the following expression: ； in, , They are respectively The east and north coordinates after time merging , These are the east and north coordinates of the path reference node, respectively; B2. Calculate heading error The deviation in the driving direction is reflected by the angular deviation between the robot's current heading and the reference heading of the path, as expressed in the following expression: ； in, for Yaw angle after time-integration The reference heading for the path reference node; C2. Calculate the overall tracking error The weighted fusion of horizontal position error and heading error is used as input to the DQN algorithm, and the expression is: ； in, ; For horizontal position error weights, This represents the heading error weight.

2. The intelligent navigation and detection system for a wheeled inspection robot based on multi-sensor fusion as described in claim 1, characterized in that: The sensor layer includes LiDAR, vision camera, inertial measurement unit (IMU), odometer, and temperature and humidity sensors; LiDAR is used for obstacle recognition and localization: by emitting a laser beam to scan the environment and receiving the reflected light signals, it calculates three-dimensional point cloud data to obtain the robot's global position. and coordinates of environmental obstacles ( , ); Visual cameras are used for target recognition: after acquiring environmental images, the target coordinates are output through the YOLOv8 target detection algorithm; Inertial Measurement Unit (IMU) is used to acquire robot attitude: robot attitude angles are acquired via a three-axis accelerometer and a three-axis gyroscope. By combining acceleration and velocity, the robot's posture can be obtained. For roll angle, The pitch angle, Yaw angle; The odometer is integrated into the drive wheels to measure distance and speed. ; Temperature and humidity sensors are used for environmental monitoring to obtain ambient temperature and humidity data. All sensors transmit data via RS485 / Ethernet interfaces with a unified timestamp and in JSON format.

3. The intelligent navigation and detection system for a wheeled inspection robot based on multi-sensor fusion according to claim 2, characterized in that: The details of the improved ensemble Kalman filter algorithm for the data fusion layer are as follows: Step 1: Initialize the set: Set the fusion state vector , Indicates the robot's global position. The coordinates are eastward. The coordinates are north-facing. For height, Indicates the robot's attitude angle. For roll angle, The pitch angle, Yaw angle This represents the velocity; the initial set state is the sensor's initial measurements superimposed with a Gaussian perturbation, expressed as: , In the formula, Let i be the state vector of the i-th member of the set at the initial time. This is the mean vector of sensor measurements at the initial moment. Let be the initial perturbation covariance matrix. The symbol for Gaussian distribution is . Indicates the total number of members in the set; Step 2: Based on the differential motion physical model of the wheeled robot, predict the state of each set member, as shown in the following expression: ； In the formula, for Time of the first The predicted state vector of each set member, This is a differential motion physical model function for a wheeled robot. for The robot control input vector at time 1. for Time of the first The analytical state vector of each set member, The model error covariance matrix; Step 3: Calculate the forecast ensemble mean and forecast error covariance matrix, and calculate the Kalman gain. The specific process is as follows: S31. Calculate the mean of the forecast ensemble and the covariance matrix of the forecast error to quantify the forecast uncertainty. The expression is as follows: ； ； S32, Introduction Time observation vector Observation matrix H, observation error covariance matrix R; S33. Calculate the Kalman gain and determine the correction weights of the observed data for the prediction. The expression is as follows: ； In the formula, express The mean vector of the time forecast set; Indicates the total number of members in the set; express Time forecast error covariance matrix; express The Kalman gain matrix at each time step; Step 4: Analyze and update the set members, and output the mean of the merged states, as shown in the following expression: ； In the formula, express Time of the first The analysis state vector of each set member; Step 5: Calculate the mean of the analysis set The first three components are the global position. Points in the processed LiDAR point cloud data that are less than a set threshold are selected as the coordinates of environmental obstacles. ).

4. The intelligent navigation and detection system for a wheeled inspection robot based on multi-sensor fusion according to claim 3, characterized in that: In C1, the GA algorithm optimizes the static optimal path with path reference nodes generated in B1, and the process of generating the dynamic path with updated path reference nodes is as follows: C11. Construct the initial population; use the static optimal path output by the PSO algorithm. Based on this, an initial population is generated through random perturbation; C12. Optimize the preset number of iterations through selection, crossover, and mutation operations; C13. Output dynamic path; after the preset number of iterations, select the path with the highest fitness value as the dynamic path. The nodes on this path are the updated path reference nodes. When new dynamic obstacles appear in the environment, the above-mentioned GA algorithm replanning process is triggered to update the path reference nodes in real time.

5. The intelligent navigation and detection system for a wheeled inspection robot based on multi-sensor fusion according to claim 4, characterized in that: Fitness function of GA-PSO fusion algorithm for: ； in, , , ; , , These are the path length weight, path smoothness weight, and obstacle avoidance success rate weight, respectively. To normalize the path length, To normalize path smoothness, The success rate of overshoot avoidance is 1 if the path and obstacle grid do not overlap, and 0 if they overlap.

6. The intelligent navigation and detection system for a wheeled inspection robot based on multi-sensor fusion according to claim 5, characterized in that: The execution layer includes a DC motor, a servo motor, and an electromagnetic brake. A PID controller is used to achieve speed and steering angle tracking, as shown in the following expression: ； in, This is the proportionality coefficient. The integral coefficient is... These are the differential coefficients. For the desired speed, For actual speed, To track speed.

Images