An underwater defect detection robot path planning and tracking control method based on multi-data fusion
By employing a multi-data fusion-based path planning and tracking control method, the problem of positioning and path tracking for underwater defect detection robots in GPS-denied environments was solved, achieving high-precision global coverage and local convergence, thereby improving detection efficiency and accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG UNIV
- Filing Date
- 2026-04-28
- Publication Date
- 2026-07-21
AI Technical Summary
Existing underwater defect detection robots suffer from poor positioning accuracy and unstable path tracking in GPS-denied environments, making it difficult to achieve high-precision positioning and detailed detection of specific defect points.
A path planning and tracking control method based on multi-data fusion is adopted, combined with an improved artificial bee colony algorithm to optimize the BP neural network-assisted Kalman filter for integrated navigation, and a global coverage and local approach strategy is combined to use a predefined time sliding mode controller and disturbance observer for robust tracking control.
It achieves high-precision positioning and stable path tracking in GPS-denied environments, improving detection efficiency and accuracy, especially significantly enhancing the ability to approach specific defect points.
Smart Images
Figure CN122431133A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent control and path planning technology for underwater robots, specifically to a path planning and tracking control method, system and storage medium for underwater defect detection robots based on multi-data fusion. It is particularly suitable for high-precision positioning, coverage scanning and accurate proximity detection of defects (such as cracks, corrosion, cavitation, etc.) in underwater structures such as bridge piers, dams, ship hulls and subsea pipelines in GPS-denied environments. Background Technology
[0002] Underwater structures, during long-term service, are prone to surface defects such as cracks, rust, and cavitation pits due to factors such as water erosion, biological adhesion, chemical corrosion, and mechanical damage. Failure to detect and repair these defects in a timely manner can lead to structural failure or even safety accidents. Traditional underwater defect detection relies primarily on visual inspection by divers or manual operation by remotely operated vehicles (ROVs), which suffers from low efficiency, high subjectivity, underwater operational hazards, and poor positioning accuracy. In recent years, with the development of underwater robotics technology, some research has attempted to combine inertial navigation systems (INS), Doppler velocimeters (DVLs), electronic compasses, and depth gauges for underwater positioning. However, the complex underwater environment presents challenges such as strong interference, low light levels, multipath effects, and complete failure of the Global Positioning System (GPS) signal. This causes positioning errors of a single navigation system to accumulate over time, and DVLs are prone to signal loss near the bottom or in turbid water, failing to meet the path tracking requirements for high-precision defect detection. Existing path planning methods mostly rely on pre-set fixed routes and lack the ability to adaptively correct dynamic deviations. Traditional PID control suffers from slow response, large overshoot, and weak disturbance resistance in underwater robot path tracking, especially when performing "approach-based" fine inspection of specific defect points (such as crack tips or corrosion pit centers), resulting in unsmooth steering and poor stability. Therefore, there is an urgent need for an underwater defect detection robot control method that can achieve high-precision positioning in GPS-denied environments, possesses intelligent path planning that combines global coverage and local approach, and employs an advanced robust nonlinear controller for fast and stable tracking. Summary of the Invention
[0003] This invention aims to overcome the problems of poor positioning accuracy, unstable path tracking, and insufficient ability to approach specific defect points in existing underwater defect detection robots. It provides a path planning and tracking control method, system, and storage medium for underwater defect detection robots based on multi-data fusion, which can achieve high-precision integrated navigation, adaptive path planning, and robust tracking control.
[0004] To achieve the above objectives, the technical solution provided by this invention is as follows:
[0005] A path planning and tracking control method for an underwater defect detection robot based on multi-data fusion, characterized by the following steps:
[0006] Step S1: Based on the improved artificial bee colony algorithm, optimize the BP neural network-assisted Kalman filter combined navigation method, integrate data from the inertial navigation system, Doppler velocimeter, electronic compass and depth gauge, and use it for high-precision positioning of underwater defect detection robot in GPS denied environment;
[0007] Step S2: Combining global coverage and local approach strategies, a bow-shaped round-trip scanning path covering the entire defect area is first generated to ensure no omissions in detection; then, for the marked high-priority defect feature points, a local approach path based on a pure tracking model is dynamically generated, and the forward look distance is adaptively adjusted to smoothly turn, achieving accurate and smooth approach to key locations such as crack tips and corrosion pit centers.
[0008] Step S3: A path tracing control method based on a predefined time sliding mode controller is used to drive the actuator to accurately track the planned path;
[0009] Step S4: Design the sliding surface and convergence conditions of the predefined time sliding mode controller. The sliding surface is as follows:
[0010]
[0011] in, For sliding mode parameters, To track the error; the predefined time convergence condition is: This allows the tracking error to be minimized within a user-defined timeframe. It converges inward to a small neighborhood near the origin;
[0012] Step S5: Design the disturbance observer and adaptive mechanism for the predefined time sliding mode controller. The disturbance observer is as follows:
[0013]
[0014] in, This is the disturbance estimate. To achieve adaptive gain, an upper limit for parameter uncertainty is determined by combining a parameter adaptive mechanism, enabling real-time compensation for unknown disturbances such as ocean waves.
[0015] Furthermore, according to the method described in claim 1, the specific optimization method of the improved artificial bee colony algorithm in the combined navigation method based on the improved artificial bee colony algorithm to optimize the BP neural network-assisted Kalman filter in step S1 is as follows:
[0016] The search method for collecting bees and the selection probability of observation bees are optimized by introducing dynamic adjustment factors and global optimal solution guidance strategies.
[0017] The formula for searching for honeybees is:
[0018]
[0019] in, For the first The bee in the first Dimension's current position To randomly select another bee location, The global optimal solution is the first... dimensional components, For interval Random numbers on the screen For the number of iterations The dynamically changing adjustment factor has the following update rule:
[0020]
[0021] in, This represents the current iteration number. The maximum number of iterations, It is a preset constant and satisfies ;
[0022] The selection probability of the observed bees is determined using an adaptive judgment factor. Perform segment selection:
[0023] Generate an interval random numbers on The probability of selection is:
[0024]
[0025] in, For the first The fitness value of the food source, which corresponds to the first food source. The weight threshold of the BP neural network is defined as the reciprocal of the sum of squared errors of the network output; The total number of food sources; the segmented selection mechanism is used to maintain population diversity and avoid getting trapped in local optima.
[0026] Furthermore, according to the method described in claim 1, the specific implementation of the defect region path planning method combining global "bow"-shaped coverage and local pure tracking approximation in step S2 includes:
[0027] Based on the defect region boundary identified by sonar or optical cameras, a global planning path in the shape of an "arch" is generated, covering the entire defect region. The path point set is as follows: ,in For the scan step, For path segment index, Achieve back-and-forth motion;
[0028] For the marked high-priority defect feature points, a "approaching" local planning path based on a pure tracking model is dynamically generated, and the optimal forward look distance is determined by solving geometric relationships.
[0029] The formula for calculating the steering angle of the pure tracking model is as follows:
[0030]
[0031] in, For the spacing between the two thrusters, This represents the lateral deviation of the target point in the vehicle coordinate system. The forward look-ahead distance is determined through optimization.
[0032]
[0033] in, This is the steering angle penalty weighting coefficient. This represents the position tracking error.
[0034] Furthermore, according to the method of claim 1, the specific implementation of the sliding surface and convergence condition of the predefined time sliding mode controller in step S4 includes control law design:
[0035]
[0036] in, For equivalent control items, To switch control terms; predefined time convergence function satisfy , ,and , The parameter is positive, so that the system state changes over time. It converges precisely within the inner quadrant.
[0037] Furthermore, according to the method of claim 1, the specific implementation of the disturbance observer and adaptive mechanism of the predefined time sliding mode controller in step S5 further includes:
[0038] The adaptive gain update law is:
[0039]
[0040] in, It is a positive constant; the estimation error of the disturbance observer converges to zero within a predefined time, realizing accurate estimation and feedforward compensation of unknown time-varying disturbances.
[0041] Furthermore, a path planning and tracking control method for an underwater defect detection robot based on multi-data fusion is characterized in that the underwater defect detection robot comprises:
[0042] A high-precision integrated positioning module is used to implement the integrated navigation method described in step S1 of claim 1; when the external auxiliary sensor signal is valid, this module uses a BP neural network online training mode to correct the cumulative error of the inertial navigation system, and the correction formula is:
[0043]
[0044] in, These are the corrected velocity and position estimates, respectively. These are the original INS solution values. This is the error compensation amount for neural network prediction; when the signal fails, it switches to prediction mode to directly predict and compensate for velocity and position errors;
[0045] A defect area path planning module is used to implement the path planning method described in step S2 of claim 1;
[0046] The adaptive path tracking control module incorporates the predefined time sliding mode controller described in claims 4 and 5; this module calculates in real time the lateral position deviation between the robot's current position and the planned path. and heading angle deviation After inputting a robust nonlinear controller, it outputs control commands to form a closed-loop control circuit. The closed-loop control equation is as follows:
[0047]
[0048] in, Here is the control law function for a robust nonlinear controller. The system state transition function; and the actuator, which drives the robot to travel along the planned path in response to the control commands output by the adaptive path tracking control module.
[0049] The underwater defect detection robot path planning and tracking control method proposed in this invention, based on multi-data fusion, significantly improves the performance of underwater structure defect detection systems, mainly in the following aspects:
[0050] 1. Improved Positioning Accuracy and Reliability: A combined navigation method based on an improved artificial bee colony algorithm and optimized BP neural network-assisted Kalman filter, integrating data from the inertial navigation system, Doppler velocimeter, electronic compass, and depth gauge, achieves high-precision positioning in GPS-denied environments. When external auxiliary sensor signals such as DVL are valid, the BP neural network is trained online to correct the accumulated INS error; when the signal fails, the neural network switches to prediction mode to directly compensate for velocity and position errors, effectively solving the error divergence problem of a single underwater navigation system. The synergy between hard parameter sharing and Kalman filtering enables deep fusion of multi-sensor data, significantly improving the continuity and reliability of positioning.
[0051] 2. Enhancing Path Planning Efficiency and Intelligence: Combining global "bow"-shaped coverage with local pure tracking approach defect area path planning methods, a full-coverage scanning path is generated based on the defect area boundaries identified by sonar or optical cameras. Simultaneously, "approaching" local paths are dynamically generated for high-priority defect feature points. A hard parameter sharing mechanism allows global and local planning to share the same state estimation backbone network. The positional information provided by global coverage guides the generation of local approach paths, while the fine turning angles of local approach paths in turn optimize the step size of the global path, forming a collaborative enhancement between tasks and significantly improving detection coverage efficiency and defect approach accuracy.
[0052] 3. Optimized Tracking Control Performance and Ease of Engineering Deployment: A path tracking control method based on a robust nonlinear controller, selected from data-driven nonlinear model predictive controllers, Lyapunov model-based predictive guidance controllers, predefined time sliding mode controllers, deep reinforcement learning-based end-to-end controllers, or diffusion-prior Lyapunov Actor-Critic framework controllers, achieves fast and stable tracking of the planned path. A single end-to-end control model directly outputs actuator control commands, eliminating the need for complex multi-controller switching logic, significantly reducing system deployment difficulty and inference latency. The combined use of Lyapunov stability constraints and event-triggered mechanisms effectively addresses the issues of large overshoot and slow response when underwater robots face water flow disturbances, making it particularly suitable for underwater edge computing scenarios.
[0053] 4. Enhanced Generalization and Robustness: Through multi-sensor fusion and multi-controller adaptive selection mechanisms, high-precision positioning and stable tracking are maintained under different water qualities, depths, and surface conditions of various structures. The complementary advantages of position and heading errors in the joint loss function solve the control deviation problem caused by underwater robot attitude disturbances. The multi-center generalization verification mechanism uses independent underwater test scenarios to evaluate generalization error, ensuring that the model maintains high accuracy under different scenarios such as bridge piers, dams, and subsea pipelines, significantly improving the practical deployment value of the system. Attached Figure Description
[0054] Figure 1 This is a flowchart of a path planning and tracking control method for an underwater defect detection robot based on multi-data fusion proposed in this invention;
[0055] Figure 2 For global "bow" shaped path planning and tracking trajectory comparison;
[0056] Figure 3 This is for local pure tracing to approximate path effects;
[0057] Figure 4 Comparison of tracking performance for a predefined time sliding mode controller (PTC-SMC). Detailed Implementation
[0058] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These fall within the scope of protection of the present application.
[0059] Example 1
[0060] To achieve the above objectives, the technical solution provided by this invention is as follows:
[0061] A path planning and tracking control method for an underwater defect detection robot based on multi-data fusion, characterized by the following steps:
[0062] Step S1: Based on the improved artificial bee colony algorithm, optimize the BP neural network-assisted Kalman filter combined navigation method, integrate data from the inertial navigation system, Doppler velocimeter, electronic compass and depth gauge, and use it for high-precision positioning of underwater defect detection robot in GPS denied environment;
[0063] Step S2: Combining global coverage and local approach strategies, a bow-shaped round-trip scanning path covering the entire defect area is first generated to ensure no omissions in detection; then, for the marked high-priority defect feature points, a local approach path based on a pure tracking model is dynamically generated, and the forward look distance is adaptively adjusted to smoothly turn, achieving accurate and smooth approach to key locations such as crack tips and corrosion pit centers.
[0064] Step S3: A path tracing control method based on a predefined time sliding mode controller is used to drive the actuator to accurately track the planned path;
[0065] Step S4: Design the sliding surface and convergence conditions of the predefined time sliding mode controller. The sliding surface is as follows:
[0066]
[0067] in For sliding mode parameters, To track the error; the predefined time convergence condition is: This allows the tracking error to be minimized within a user-defined timeframe. It converges inward to a small neighborhood near the origin;
[0068] Step S5: Design the disturbance observer and adaptive mechanism for the predefined time sliding mode controller. The disturbance observer is as follows:
[0069]
[0070] in, This is the disturbance estimate. To achieve adaptive gain, an upper limit for parameter uncertainty is determined by combining a parameter adaptive mechanism, enabling real-time compensation for unknown disturbances such as ocean waves.
[0071] Furthermore, according to the method described in claim 1, the specific optimization method of the improved artificial bee colony algorithm in the combined navigation method based on the improved artificial bee colony algorithm to optimize the BP neural network-assisted Kalman filter in step S1 is as follows:
[0072] The search method for collecting bees and the selection probability of observation bees are optimized by introducing dynamic adjustment factors and global optimal solution guidance strategies.
[0073] The formula for searching for honeybees is:
[0074]
[0075] in, For the first The bee in the first Dimension's current position To randomly select another bee location, The global optimal solution is the first... dimensional components, For interval Random numbers on the screen For the number of iterations The dynamically changing adjustment factor has the following update rule:
[0076]
[0077] in, This represents the current iteration number. The maximum number of iterations, It is a preset constant and satisfies ;
[0078] The selection probability of the observed bees is determined using an adaptive judgment factor. Perform segment selection:
[0079] Generate an interval random numbers on The probability of selection is:
[0080]
[0081] in, For the first The fitness value of the food source, which corresponds to the first food source. The weight threshold of the BP neural network is defined as the reciprocal of the sum of squared errors of the network output; The total number of food sources; the segmented selection mechanism is used to maintain population diversity and avoid getting trapped in local optima.
[0082] Furthermore, according to the method described in claim 1, the specific implementation of the defect region path planning method combining global "bow"-shaped coverage and local pure tracking approximation in step S2 includes:
[0083] Based on the defect region boundary identified by sonar or optical cameras, a global planning path in the shape of an "arch" is generated, covering the entire defect region. The path point set is as follows: ,in For the scan step, For path segment index, Achieve back-and-forth motion;
[0084] For the marked high-priority defect feature points, a "approaching" local planning path based on a pure tracking model is dynamically generated, and the optimal forward look distance is determined by solving geometric relationships.
[0085] The formula for calculating the steering angle of the pure tracking model is as follows:
[0086]
[0087] in, For the spacing between the two thrusters, This represents the lateral deviation of the target point in the vehicle coordinate system. The forward look-ahead distance is determined through optimization.
[0088]
[0089] in, This is the steering angle penalty weighting coefficient. This represents the position tracking error.
[0090] Furthermore, according to the method of claim 1, the specific implementation of the sliding surface and convergence condition of the predefined time sliding mode controller in step S4 includes control law design:
[0091]
[0092] in, For equivalent control items, To switch control terms; predefined time convergence function satisfy , ,and , The parameter is positive, so that the system state changes over time. It converges precisely within the inner quadrant.
[0093] Furthermore, according to the method of claim 1, the specific implementation of the disturbance observer and adaptive mechanism of the predefined time sliding mode controller in step S5 further includes:
[0094] The adaptive gain update law is:
[0095]
[0096] in, It is a positive constant; the estimation error of the disturbance observer converges to zero within a predefined time, realizing accurate estimation and feedforward compensation of unknown time-varying disturbances.
[0097] Furthermore, a path planning and tracking control method for an underwater defect detection robot based on multi-data fusion is characterized in that the underwater defect detection robot comprises:
[0098] A high-precision integrated positioning module is used to implement the integrated navigation method described in step S1 of claim 1; when the external auxiliary sensor signal is valid, this module uses a BP neural network online training mode to correct the cumulative error of the inertial navigation system, and the correction formula is:
[0099]
[0100] in, These are the corrected velocity and position estimates, respectively. These are the original INS solution values. This is the error compensation amount for neural network prediction; when the signal fails, it switches to prediction mode to directly predict and compensate for velocity and position errors;
[0101] A defect area path planning module is used to implement the path planning method described in step S2 of claim 1;
[0102] The adaptive path tracking control module incorporates the predefined time sliding mode controller described in claims 4 and 5; this module calculates in real time the lateral position deviation between the robot's current position and the planned path. and heading angle deviation After inputting a robust nonlinear controller, it outputs control commands to form a closed-loop control circuit. The closed-loop control equation is as follows:
[0103]
[0104] in, Here is the control law function for a robust nonlinear controller. The system state transition function; and the actuator, which drives the robot to travel along the planned path in response to the control commands output by the adaptive path tracking control module.
[0105] The following example demonstrates the effectiveness and feasibility of the underwater defect detection robot path planning and tracking control method based on multi-data fusion disclosed in this application, proving its effectiveness and feasibility under the combined approach of integrated navigation, path planning, and robust tracking control. The specific experimental setup is as follows: The simulation uses a six-DOF autonomous underwater vehicle model with a mass of 50 kg, a length of 1.2 m, a thruster spacing of 0.6 m, and a maximum forward-looking distance of 5.0 m; the sensor configuration includes an INS with an output frequency of 100 Hz, a DVL sensor with an output frequency of 10 Hz, an electronic compass with an output frequency of 1 Hz, and a depth gauge with an output frequency of 10 Hz; the simulated corrosion defect area on the bridge pier surface is a rectangular region. Two high-priority defect feature points are internally marked. (Center of the corrosion pit) and (Crack tip); Apply ocean wave disturbance The entire simulation process does not provide GPS signals, relying solely on a combination of INS / DVL / electronic compass / depth meter for positioning; the predefined time sliding mode controller parameters are set to... Convergence time Perturbation observer parameters Path planning uses a bow-shaped global step size. Pure tracking of local approximation forward distance optimization weights Improved artificial bee colony algorithm for bee colony size in integrated navigation. Maximum number of iterations Regulatory factors The BP neural network structure is (Input layer 12-dimensional, hidden layer 25-dimensional, output layer 6-dimensional), learning rate 0.01; total simulation time 60 seconds, robot starts from initial position Starting from the beginning, a "bow"-shaped global coverage path is executed for the first 40 seconds, and then the path approaches sequentially for the next 20 seconds. and Control cycle .
[0106] Simulation results are as follows Figures 2 to 4 The corresponding performance table is shown below. Figure 2 This is a comparison diagram of global "bow"-shaped path planning and tracking trajectory. The black dashed line represents the ideal planned path, and the red solid line represents the robot's actual movement trajectory. The path point set satisfies... The adjacent rows are smoothly transitioned by arcs, achieving 100% coverage of the defect area. The actual trajectory highly overlaps with the planned path, with an average lateral deviation of... , the maximum deviation This verifies the effectiveness of the bow-shaped path generation method described in this invention.
[0107] Figure 3 The result of local pure tracking approaching path (based on high-priority feature points) (For example). In the diagram, the blue dot represents the robot's current position, and the green star represents the target point. The red curve represents the planned approach path. (Look-forward distance) According to the optimization formula Real-time calculation of steering angle The size decreases smoothly as the robot approaches the target point. Ultimately, the robot moves at approximately... The accuracy of docking at Nearby, a fine-grained approximation was achieved, resulting in improved accuracy compared to traditional pure tracking methods. .
[0108] Figure 4 Tracking performance graph for a predefined time sliding mode controller (PTC-SMC). Figure 4 (a) represents the lateral position error. Curve changing over time: Initial error At a predefined time It converges to within 0.05 m, and then stabilizes. . Figure 4 (b) represents the heading angle error. Curve: Initial Deviation It also converges to within 5 seconds. Within this range. In contrast, a traditional PID controller has a convergence time of approximately 12 seconds under the same disturbance, and also exhibits… The steady-state oscillations are shown above. These results fully demonstrate that the predefined time sliding mode controller in this invention possesses fast convergence and strong robustness.
[0109] The performance metrics are summarized in the table below:
[0110]
[0111] In summary, this method demonstrates the following advantages in path planning and tracking control tasks for underwater defect detection robots: 1) High-precision integrated navigation: The improved artificial bee colony algorithm and optimized BP neural network-assisted Kalman filter control the positioning error after DVL failure to within 0.15 m; 2) Efficient path planning: The combination of global bow-shaped coverage and local pure tracking approach achieves 100% coverage of the defect area and a high-priority point approach accuracy of 0.05 m; 3) Robust tracking control: The predefined time sliding mode controller completes path tracking within 5 seconds with a steady-state error of only 0.03 m, and the overshoot is reduced by 68% compared to PID. These results fully demonstrate the effectiveness and feasibility of the method proposed in this application.
[0112] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions that are any combination of the above technical features. It should be noted that for those skilled in the art, various improvements and modifications can be made without departing from the principle of this invention, and these modifications are also considered within the scope of protection of this invention.
Claims
1. A path planning and tracking control method for an underwater defect detection robot based on multi-data fusion, characterized in that, Includes the following steps: Step S1: Based on the improved artificial bee colony algorithm, optimize the BP neural network-assisted Kalman filter combined navigation method, integrate data from the inertial navigation system, Doppler velocimeter, electronic compass and depth gauge, and use it for high-precision positioning of underwater defect detection robot in GPS denied environment; Step S2: Combining global coverage and local approach strategies, firstly, a "bow"-shaped round-trip scanning path covering the entire defect area is generated to ensure no omissions in detection; then, for the marked high-priority defect feature points, a local approach path based on a pure tracking model is dynamically generated, and the forward look distance is adaptively adjusted to smoothly turn, achieving accurate and smooth approach to key locations such as crack tips and corrosion pit centers. Step S3: A path tracing control method based on a predefined time sliding mode controller is used to drive the actuator to accurately track the planned path; Step S4: Design the sliding surface and convergence conditions of the predefined time sliding mode controller. The sliding surface is as follows: ； in, For sliding mode parameters, To track the error; the predefined time convergence condition is: This allows the tracking error to be minimized within a user-defined timeframe. It converges inward to a small neighborhood near the origin; Step S5: Design the disturbance observer and adaptive mechanism for the predefined time sliding mode controller. The disturbance observer is as follows: ； in, This is the disturbance estimate. To achieve adaptive gain, an upper limit for parameter uncertainty is determined by combining a parameter adaptive mechanism, enabling real-time compensation for unknown disturbances such as ocean waves.
2. The method according to claim 1, characterized in that, In step S1, the specific optimization method of the improved artificial bee colony algorithm for optimizing BP neural network-assisted Kalman filtering is as follows: The search method for collecting bees and the selection probability of observation bees are optimized by introducing dynamic adjustment factors and global optimal solution guidance strategies. The formula for searching for honeybees is: ； in, For the first The bee in the first Dimension's current position To randomly select another bee location, The global optimal solution is the first... dimensional components, For interval Random numbers on the screen For the number of iterations The dynamically changing adjustment factor has the following update rule: ； in, This represents the current iteration number. The maximum number of iterations, It is a preset constant and satisfies ; The selection probability of the observed bees is determined using an adaptive judgment factor. Perform segment selection: Generate an interval random numbers on The probability of selection is: ； in, For the first The fitness value of the food source, which corresponds to the first food source. The weight threshold of the BP neural network is defined as the reciprocal of the sum of squared errors of the network output; The total number of food sources; the segmented selection mechanism is used to maintain population diversity and avoid getting trapped in local optima.
3. The method according to claim 1, characterized in that, The specific implementation of the defect region path planning method combining global "bow"-shaped coverage and local pure tracking approximation described in step S2 includes: Based on the defect region boundary identified by sonar or optical cameras, a global planning path in the shape of an "arch" is generated, covering the entire defect region. The path point set is as follows: ,in For the scan step, For path segment index, Achieve back-and-forth motion; For the marked high-priority defect feature points, a "approaching" local planning path based on a pure tracking model is dynamically generated, and the optimal forward look distance is determined by solving geometric relationships. The formula for calculating the steering angle of the pure tracking model is as follows: ； in, For the spacing between the two thrusters, This represents the lateral deviation of the target point in the vehicle coordinate system. The forward look-ahead distance is determined through optimization. ； in, This is the steering angle penalty weighting coefficient. This represents the position tracking error.
4. The method according to claim 1, characterized in that, The specific implementation of the sliding surface and convergence conditions of the predefined time sliding mode controller mentioned in step S4 includes the design of the control law: ； in, For equivalent control items, To switch control terms; predefined time convergence function satisfy , ,and , The parameter is positive, so that the system state changes over time. It converges precisely within the inner quadrant.
5. The method according to claim 1, characterized in that, The specific implementation of the disturbance observer and adaptive mechanism of the predefined time sliding mode controller mentioned in step S5 also includes: The adaptive gain update law is: ； in, It is a positive constant; the estimation error of the disturbance observer converges to zero within a predefined time, realizing accurate estimation and feedforward compensation of unknown time-varying disturbances.
6. A path planning and tracking control method for an underwater defect detection robot based on multi-data fusion, characterized in that, The underwater defect detection robot includes: A high-precision integrated positioning module is used to implement the integrated navigation method described in step S1 of claim 1; when the external auxiliary sensor signal is valid, this module uses a BP neural network online training mode to correct the cumulative error of the inertial navigation system, and the correction formula is: ； in, These are the corrected velocity and position estimates, respectively. These are the original INS solution values. This is the error compensation amount for neural network prediction; when the signal fails, it switches to prediction mode to directly predict and compensate for velocity and position errors; A defect area path planning module is used to implement the path planning method described in step S2 of claim 1; The adaptive path tracking control module incorporates the predefined time sliding mode controller described in claims 4 and 5; this module calculates in real time the lateral position deviation between the robot's current position and the planned path. and heading angle deviation After inputting a robust nonlinear controller, it outputs control commands to form a closed-loop control circuit. The closed-loop control equation is as follows: ； in, Here is the control law function for a robust nonlinear controller. The system state transition function; and the actuator, which drives the robot to travel along the planned path in response to the control commands output by the adaptive path tracking control module.