Underwater vehicle path planning and obstacle avoidance method based on deep learning

Abstract

The invention relates to the technical field of underwater robots, and discloses an underwater vehicle path planning and obstacle avoidance method based on deep learning, comprising the following steps: S1, acquiring multi-modal data of an underwater environment through an underwater sensor and a Beidou satellite navigation system, and generating a high-precision environment model in real time; and S2, fusing the global positioning information provided by the Beidou satellite with the data of the sonar, the inertial navigation system and the visual SLAM by using a multi-sensor fusion algorithm. By fusing a high-definition sensor and a Beidou satellite navigation system and combining a deep learning model, the environment sensing and path planning capability of an underwater vehicle is remarkably improved, and in a complex and changeable underwater environment, the high-definition sensor can collect accurate environment data in real time, so that the accuracy of path planning is improved. And the deep learning model accurately identifies and classifies obstacles according to the data and plans an optimal obstacle avoidance path, and the navigation capability of the aircraft is improved by the sensing precision and real-time performance of the deep learning model.

Description

Technical Field

[0001] This invention relates to the field of underwater robot technology, specifically to a deep learning-based method for underwater vehicle path planning and obstacle avoidance. Background Technology

[0002] With the development of marine development and underwater exploration technologies, underwater vehicles have become an important tool for marine exploration, rescue, military and scientific research, and have been widely used. However, the underwater environment is complex and uncertain. Factors such as obstacle distribution, water flow changes and underwater topography increase the difficulty of path planning and obstacle avoidance for vehicles. Therefore, autonomous navigation technology for underwater vehicles has become a research focus.

[0003] Traditional path planning and obstacle avoidance methods are mainly based on classical algorithms, which rely on discretized mesh graphs for path planning. For example, CN109470249A discloses an optimal path planning and obstacle avoidance design method for underwater vehicles, which integrates a binary heap acceleration algorithm, a 26-neighborhood NA* algorithm, and a "circular arc-straight line-circular arc" turning strategy. This method considers the physical characteristics of the underwater vehicle and optimizes the path through the "circular arc-straight line-circular arc" turning strategy, effectively reducing inflection points and improving path smoothness. However, it still has the following shortcomings:

[0004] This method relies on the construction of a 3D mesh map to identify obstacles, and it does not integrate the real-time perception capability of high-definition sensors or multimodal data. The poor light and short line of sight in the underwater environment result in low accuracy of environmental perception and it cannot cope with complex underwater terrain and dynamically changing obstacles.

[0005] Traditional NA* algorithms plan paths based on pre-defined heuristic search rules, lacking adaptability. When encountering unknown or complex environments, they cannot adjust their paths and lack autonomous learning mechanisms in dynamically changing underwater environments, resulting in insufficient adaptability and flexibility.

[0006] Although this method optimizes the local path through turning strategies, it has shortcomings in combining global navigation and local obstacle avoidance, and is prone to getting trapped in local optima, causing the vehicle to be unable to effectively cope with changes in the global environment.

[0007] Therefore, those skilled in the art provide a deep learning-based underwater vehicle path planning and obstacle avoidance method to solve the problems mentioned in the background art. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for underwater vehicle path planning and obstacle avoidance, thereby solving the problems mentioned in the background section.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solution: a deep learning-based underwater vehicle path planning and obstacle avoidance method, comprising the following steps: S1, acquiring multimodal data of the underwater environment through underwater sensors and the BeiDou satellite navigation system, and generating a high-precision environment model in real time;

[0010] S2. Utilize multi-sensor fusion algorithms to fuse global positioning information provided by BeiDou satellites with data from sonar, inertial navigation systems, and visual SLAM.

[0011] S3. Process environmental perception data through deep learning models, and use convolutional neural networks to identify and classify obstacles in the 3D environment acquired by high-definition cameras and LiDAR, and predict the navigation path.

[0012] S4. Based on a deep reinforcement learning model, obstacle avoidance paths are planned in real time in dynamic environments. The obstacle avoidance strategy is adjusted through the reward function in reinforcement learning, with the specific formula as follows:

[0013] R=-λ1·d obstacle +λ2·d goal -λ3·C collision ,

[0014] Where R is the reward value, d obstacle d is the distance between the vehicle and the obstacle. goal C is the distance between the vehicle and the target point. collision λ1, λ2, and λ3 are the collision penalty coefficients, and λ1, λ2, and λ3 are the weighting parameters.

[0015] S5. Utilizing the global position information provided by the BeiDou Navigation Satellite System, if the vehicle deviates from the globally planned path, the global path correction module will be automatically activated to replan the path based on BeiDou satellite data.

[0016] S6. Based on the turning radius and physical dimensions of the aircraft, a path smoothing strategy of "circular arc-straight line-circular arc" is adopted to optimize the path output by deep learning.

[0017] S7. Through an online learning mechanism, the aircraft can adaptively adjust itself based on new environmental data during execution and update the weights of the deep learning model.

[0018] S8. Based on the planned path and the results of real-time obstacle avoidance output, generate motion control commands for the vehicle and control the vehicle's thrusters and servos to perform motion.

[0019] S9. During the operation of the aircraft, continuously acquire environmental perception data and update the model in real time;

[0020] S10. Monitor the deviation between the vehicle's position and the target path through the Beidou navigation system, and activate the emergency obstacle avoidance mode in case of emergency.

[0021] Preferably, in step S2, the algorithm formula for correcting the positioning error is as follows:

[0022] P t+1 =P t +α·(P BDS -P t )+β·(P INS -P t )+γ·(P SLAM -P t ),

[0023] Among them, P t+1 For the corrected position, P BDS P INS P SLAM The location information provided for BeiDou navigation, inertial navigation and visual SLAM, where α, β and γ are weighting coefficients.

[0024] Preferably, the multi-sensor fusion algorithm is used to combine data from the BeiDou Navigation Satellite System, the Inertial Navigation System, and Visual SLAM to enhance the positioning accuracy of underwater vehicles. The following is the formula for this fusion algorithm:

[0025] P fused =α·P BDS +β·P INS +γ·P SLAM ,

[0026] Among them, P fused The final localization result after fusion, P BDS Positioning information obtained through the BeiDou satellite navigation system, P INS Positioning information obtained through an inertial navigation system, P SLAM The positioning information obtained through visual SLAM, and the weighting coefficients of the α, β, and γ system data.

[0027] Preferably, the path planning uses a convolutional neural network combined with LiDAR data for environmental perception and dynamically adjusts the planned path. The following is the algorithm formula for this path planning:

[0028] The environmental perception and path planning formula is as follows:

[0029]

[0030] Among them, P optimal The optimal path optimized by convolutional neural network and LiDAR perception, candidate points of path P, d goal The distance d from the path point to the target pointobstacle The distance C from the path point to surrounding obstacles smooth The smoothness penalty function of path P, λ1, λ2, and λ3 control the weight coefficients of various factors.

[0031] Preferably, the global path correction module of the BeiDou navigation assistance is used to automatically replan the path when the vehicle deviates from the predetermined path. The following are the relevant algorithm formulas:

[0032] Global path correction formula:

[0033] P new =P current +α·(P BDS -P current )+β·(P planned -P current ),

[0034] Among them, P new The corrected new waypoint, P current The location of the aircraft, P BDS P uses precise global location information obtained through the BeiDou Navigation Satellite System. planned The target points on the predetermined planning path, and the weighting coefficients α and β.

[0035] Preferably, the path smoothing and optimization uses a "circular arc-straight line-circular arc" strategy to generate a smooth path for the vehicle's physical characteristics. The following is the algorithm formula for calculating the smooth path:

[0036] Circular arc-straight line-circular arc path optimization formula:

[0037] Calculation of circular arc segments: Where, θ arc The angle of the arc segment, L arc The arc length of the circular segment, r is the turning radius of the aircraft;

[0038] Calculation of straight segment length:

[0039] Among them, L line The straight-line distance between arc segments; the coordinates of the start and end points of the straight-line segments x1, y1 and x2, y2;

[0040] Path smoothness optimization function:

[0041] C smooth =λ1·θ arc +λ2·L line -λ3·(1 / r),

[0042] Among them, C smooth The path smoothness cost function, with weight coefficients λ1, λ2, and λ3.

[0043] Preferably, the dynamic self-learning adjustment updates the weights of the deep learning model in real time through an online learning mechanism. The following is the algorithm formula for this process:

[0044] Online learning weight update formula:

[0045]

[0046] Among them, w t+1 The model weights, w, updated at time t+1 t The model weights at time t, and the learning rate η. Loss function L with respect to weight w t The gradient represents the model's performance on the input data x. t and real label y t The error derivative on;

[0047] Real-time environment input:

[0048] x t Input data collected by sensors or vision systems in this environment;

[0049] y t The actual obstacle avoidance path or successful navigation result is used as a label in supervised learning to correct the model's predictions;

[0050] Loss function:

[0051] Where, f(w) t x t Deep learning models for input x t The predicted output, y t Realistic navigation or obstacle avoidance results.

[0052] Preferably, the controller performs dynamic adjustments to the underwater vehicle through multi-channel control signals. The controller algorithm formula is as follows:

[0053] Thruster speed control formula: v new =v current +K v ·(v desired -v current ),

[0054] Among them, v new Adjusted thruster speed, v current Thruster speed, v desired Desired thruster speed, K v Speed ​​adjustment gain coefficient;

[0055] Heading control formula (based on proportional-integral-derivative control, PID):

[0056]

[0057] Where, θ new Adjusted heading angle, θ current Heading angle, e(t) heading error, K p K i K d These are the proportional, integral, and derivative coefficients of the PID controller.

[0058] Attitude control formula: φ new =φ current +K φ ·(φ desired -φ current ),

[0059] Where, φ new Adjusted attitude angle, φ current Attitude angle, φ desired Desired attitude angle, K φ Attitude control gain coefficient;

[0060] Total control signal formula: u control =[v new θ new , φ new ],

[0061] The formula for the total control signal represents the final control signal u. control It is determined by the thruster speed v new Heading angle θ new and attitude angle φ new The combined multi-channel control signals are used to adjust the motion and attitude of the aircraft.

[0062] Preferably, the data feedback and real-time optimization monitors the vehicle's sensor data and path execution status in real time, and dynamically optimizes path planning and obstacle avoidance strategies through a feedback mechanism. The following is the algorithm formula for real-time data feedback and optimization:

[0063] Real-time optimization feedback formula (based on feedback control theory):

[0064] ΔP adjust =K f ·(P desired -P current ),

[0065] Where, ΔP adjust The correction value, P, is used to adjust the path planning. desired Desired path point, P current Actual path point, Kf Feedback gain coefficient;

[0066] Path error calculation formula:

[0067]

[0068] Among them, e P (t) Path error, x desired y desired z desired The three-dimensional coordinates of the desired location, x current y current z current The three-dimensional coordinates of the actual location;

[0069] Real-time optimization and adjustment formula:

[0070] P new =P current +ΔP adjust -η·e P (t),

[0071] Among them, P new The optimized new path point, η is the learning rate.

[0072] Preferably, the path tracking and monitoring is based on high-precision positioning data provided by the BeiDou Navigation Satellite System, which monitors the deviation between the vehicle and the target path in real time and makes timely adjustments. The following is the algorithm formula for path tracking and monitoring:

[0073] Path tracking error formula:

[0074]

[0075] Among them, e track (t) Tracking error between the vehicle and the target waypoint, x target y target z target The desired three-dimensional coordinates of the target point, x BDS y BDS z BDS Three-dimensional coordinates of the aircraft obtained through the BeiDou Navigation Satellite System;

[0076] Path correction formula: P adjust =P BDS +K p ·e track (t),

[0077] Among them, P adjust Adjusted path point, P BDS The BeiDou positioning coordinates of the aircraft, K p Proportional gain;

[0078] Heading adjustment formula: θ new =θ current +K θ ·(θ target -θ current ),

[0079] Where, θ new Adjusted heading angle, θ current The heading angle of the aircraft, θ target Desired heading angle, K θ Proportional gain coefficient for heading adjustment;

[0080] Overall control update formula: u control =[P adjust θ new ],

[0081] Among them, u control Adjustment signals used to control the aircraft.

[0082] This invention provides a deep learning-based method for underwater vehicle path planning and obstacle avoidance. It offers the following advantages:

[0083] 1. This invention significantly improves the environmental perception and path planning capabilities of underwater vehicles by integrating high-definition sensors with the BeiDou satellite navigation system and combining them with deep learning models. In complex and ever-changing underwater environments, high-definition sensors can collect accurate environmental data in real time, while deep learning models can accurately identify and classify obstacles based on the data and plan the optimal obstacle avoidance path. The accuracy and real-time performance of its perception enhance the navigation capabilities of the vehicle, especially in low visibility and complex terrain conditions.

[0084] 2. This invention introduces a dynamic adaptive learning mechanism, which enables the deep learning model to continuously optimize path planning and obstacle avoidance strategies based on the environmental data acquired in real time during the operation of the vehicle. This method can learn and update model weights autonomously without relying on fixed algorithms or preset environmental information, and has a strong environmental adaptability, ensuring that the vehicle can maintain stable planning capabilities and obstacle avoidance performance even in unknown or changing underwater environments.

[0085] 3. This invention integrates global high-precision positioning information provided by BeiDou satellites with local path planning intelligence from deep learning to ensure that the vehicle has higher stability in global path navigation and flexible obstacle avoidance capabilities in local environments. Its intelligent combination of global and local paths avoids the limitations of traditional algorithms that get stuck in local optima, ensuring that the vehicle always runs along the optimal route and can effectively cope with dynamically changing obstacles and environments. Attached Figure Description

[0086] Figure 1This is a flowchart of the present invention. Detailed Implementation

[0087] To enable those skilled in the art to understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0088] The present invention will now be described in detail with reference to the accompanying drawings:

[0089] Example:

[0090] Please see the appendix Figure 1 This invention provides a deep learning-based underwater vehicle path planning and obstacle avoidance method, including the following steps: S1, acquiring multimodal data of the underwater environment through underwater sensors and the BeiDou satellite navigation system, and generating a high-precision environment model in real time;

[0091] S2. Utilize multi-sensor fusion algorithms to fuse global positioning information provided by BeiDou satellites with data from sonar, inertial navigation systems, and visual SLAM.

[0092] S3. Process environmental perception data through deep learning models, and use convolutional neural networks to identify and classify obstacles in the 3D environment acquired by high-definition cameras and LiDAR, and predict the navigation path.

[0093] S4. Based on a deep reinforcement learning model, obstacle avoidance paths are planned in real time in dynamic environments. The obstacle avoidance strategy is adjusted through the reward function in reinforcement learning, with the specific formula as follows:

[0094] R=-λ1·d obstacle +λ2·d goal -λ3·C collision ,

[0095] Where R is the reward value, d obstacle d is the distance between the vehicle and the obstacle. goal C is the distance between the vehicle and the target point. collision λ1, λ2, and λ3 are the collision penalty coefficients, and λ1, λ2, and λ3 are the weighting parameters.

[0096] S5. Utilizing the global position information provided by the BeiDou Navigation Satellite System, if the vehicle deviates from the globally planned path, the global path correction module will be automatically activated to replan the path based on BeiDou satellite data.

[0097] S6. Based on the turning radius and physical dimensions of the aircraft, a path smoothing strategy of "circular arc-straight line-circular arc" is adopted to optimize the path output by deep learning.

[0098] S7. Through an online learning mechanism, the aircraft can adaptively adjust itself based on new environmental data during execution and update the weights of the deep learning model.

[0099] S8. Based on the planned path and the results of real-time obstacle avoidance output, generate motion control commands for the vehicle and control the vehicle's thrusters and servos to perform motion.

[0100] S9. During the operation of the aircraft, continuously acquire environmental perception data and update the model in real time;

[0101] S10. Monitor the deviation between the vehicle's position and the target path through the Beidou navigation system, and activate the emergency obstacle avoidance mode in case of emergency.

[0102] The benefits of S1 include generating high-precision environmental models in real-time dynamic underwater environments, improving the vehicle's environmental awareness and reaction speed; the benefits of S2 include enhancing the reliability of navigation and obstacle avoidance in complex environments, especially in areas with unstable signals or many obstacles, ensuring continuous and accurate navigation; the benefits of S3 include quickly identifying and classifying obstacles, automatically planning obstacle avoidance paths, and accurately predicting future navigation paths; the benefits of S4 include enabling the vehicle to learn to avoid obstacles and move towards the target in complex environments, effectively reducing the risk of collisions; the benefits of S5 include effectively adjusting direction in complex or unpredictable environments to avoid getting lost; and S6... The benefits of S7 are: reducing sharp turns and discontinuities in the path, improving the vehicle's turning stability and path feasibility; the benefit of S8 is that the dynamic learning mechanism enhances the vehicle's ability to cope with unknown environments and dynamic changes, enabling it to continuously optimize path planning and obstacle avoidance strategies; the benefit of S9 is that it provides a closed loop from perception to control, allowing the vehicle to flexibly cope with complex underwater environments and autonomously perform obstacle avoidance operations; the benefit of S10 is that it ensures the vehicle always makes decisions based on the latest environmental information during long-term operation, maintaining efficient navigation performance; the benefit of S10 is that once a deviation or emergency is detected, an emergency obstacle avoidance mode can be immediately activated to quickly adjust the vehicle's heading and speed. In step S2, the algorithm formula for correcting the positioning error is:

[0103] P t+1 =P t +α·(P BDS -P t )+β·(P INS -P t )+γ·(P SLAM -P t ),

[0104] Among them, P t+1 For the corrected position, P BDS P INS P SLAM The location information provided for BeiDou navigation, inertial navigation and visual SLAM, where α, β and γ are weighting coefficients.

[0105] The core function of the formula is to integrate the advantages of each sensor by fusing data from multiple sensors, thereby reducing the errors caused by a single sensor. BeiDou navigation provides high-precision global positioning, but it cannot provide stable signals in some underwater environments. Inertial navigation systems can provide stable positioning information in a short time, but they are prone to drift errors in long-term use. Visual SLAM provides real-time environmental perception through data from cameras and lidar, but it is limited in low-visibility environments. Through the fusion formula, the vehicle can integrate position information from different sources, compensate for their respective deficiencies, and thus improve the overall positioning accuracy. By fusing multi-source data, the formula can effectively reduce positioning errors caused by the failure or insufficient accuracy of a single sensor, enhancing the robustness and stability of the vehicle. By weighted fusing data from different sensors through this formula, the vehicle can maintain high positioning accuracy in complex environments, thereby enhancing the accuracy of path planning and obstacle avoidance.

[0106] Multi-sensor fusion algorithms are used to combine data from the BeiDou Navigation Satellite System, inertial navigation systems, and visual SLAM to enhance the positioning accuracy of underwater vehicles. The following is the formula for this fusion algorithm:

[0107] P fused =α·P BDS +β·P INS +γ·P SLAM ,

[0108] Among them, P fused The final localization result after fusion, P BDS Positioning information obtained through the BeiDou satellite navigation system, P INS Positioning information obtained through an inertial navigation system, P SLAM The positioning information obtained through visual SLAM, and the weighting coefficients of the α, β, and γ system data.

[0109] By weighted and fused data from different sensors, integrating positioning information from BeiDou satellite navigation, inertial navigation, and visual SLAM systems, positioning accuracy is significantly improved. The complementary advantages of each sensor minimize errors caused by partial failure or insufficient accuracy of a single sensor, ensuring accurate positioning of the vehicle in complex environments. BeiDou navigation may fail in certain deep-sea areas or when obstructed by obstacles, inertial navigation may drift over time, and visual SLAM may be limited in low light or unclear environments. This fusion algorithm effectively reduces the deficiencies of a single sensor, enabling the vehicle to continuously provide accurate position.

[0110] The path planning uses a convolutional neural network combined with LiDAR data for environmental perception and dynamically adjusts the planned path. The following is the algorithm formula for this path planning:

[0111] Environmental perception and path planning formula:

[0112]

[0113] Among them, P optimal The optimal path optimized by convolutional neural network and LiDAR perception, candidate points of path P, d goal The distance d from the path point to the target point obstacle The distance C from the path point to surrounding obstacles smooth The smoothness penalty function of path P, λ1, λ2, and λ3 control the weight coefficients of various factors.

[0114] By using convolutional neural networks to process 3D environmental data acquired by LiDAR, obstacles can be automatically identified and environmental complexity can be assessed. The resulting path planning is more intelligent than traditional algorithms. Convolutional neural networks can efficiently perceive key features in complex underwater environments, enabling the algorithm to maintain stable navigation capabilities even in dynamically changing environments.

[0115] The global path correction module of BeiDou navigation assistance is used to automatically replan the path when the vehicle deviates from the predetermined path. The following are the relevant algorithm formulas:

[0116] Global path correction formula:

[0117] P new =P current +α·(P BDS -P current )+β·(P planned -P current ),

[0118] Among them, P new The corrected new waypoint, P current The location of the aircraft, P BDSP uses precise global location information obtained through the BeiDou Navigation Satellite System. planned The target points on the predetermined planning path, and the weighting coefficients α and β.

[0119] When the vehicle deviates from the predetermined path, it automatically performs global path correction. By using the global precise positioning provided by the BeiDou Navigation Satellite System and combining it with the original planned path, the vehicle can adjust its path in a timely manner after detecting the deviation, ensuring that it always travels along the globally optimal route and avoiding path deviation caused by local obstacles or environmental changes. It effectively integrates the real-time positioning information of BeiDou navigation and the predetermined planned path, and performs path correction by combining the information of the two. BeiDou navigation provides the vehicle's precise real-time position, while the predetermined path provides the ideal navigation target. Through weighting coefficients, it ensures that the vehicle always moves towards the target in a dynamic environment and maintains the continuity of the path.

[0120] Path smoothing and optimization uses a "circular arc-straight line-circular arc" strategy to generate smooth paths for the vehicle's physical characteristics. The following is the algorithm formula for calculating the smooth path:

[0121] Circular arc-straight line-circular arc path optimization formula:

[0122] Calculation of circular arc segments: Where, θ arc The angle of the arc segment, L arc The arc length of the circular segment, r is the turning radius of the aircraft;

[0123] Calculation of straight segment length:

[0124] Among them, L l ine The straight-line distance between arc segments; the coordinates of the start and end points of the straight-line segments x1, y1 and x2, y2;

[0125] Path smoothness optimization function:

[0126] C smooth =λ1·θ arc +λ2·L l ine -λ3·(1 / r),

[0127] Among them, C smooth The path smoothness cost function, with weight coefficients λ1, λ2, and λ3.

[0128] In actual navigation, the weights of the deep learning model are updated in real time. During operation, the vehicle acquires new environmental data through sensors, makes predictions based on path planning and obstacle avoidance decisions, and calculates the error by comparing the model's predictions with the actual results. The model weights are then adjusted based on the gradient descent algorithm. This process ensures that the model can quickly learn and adapt to new environmental changes during actual operation. By updating the model weights through gradient descent and minimizing the prediction error, the model can accurately predict the vehicle's path planning and obstacle avoidance decisions, thereby improving the overall accuracy and performance of the deep learning model.

[0129] Dynamic self-learning adjustment updates the weights of the deep learning model in real time through an online learning mechanism. The following is the algorithm formula for this process:

[0130] Online learning weight update formula:

[0131]

[0132] Among them, w t+1 The model weights, w, updated at time t+1 t The model weights at time t, and the learning rate η. Loss function L with respect to weight w t The gradient represents the model's performance on the input data x. t and real label y t The error derivative on;

[0133] Real-time environment input:

[0134] x t Input data collected by sensors or vision systems in this environment;

[0135] y t The actual obstacle avoidance path or successful navigation result is used as a label in supervised learning to correct the model's predictions;

[0136] Loss function:

[0137] Where, f(w) t x t Deep learning models for input x t The predicted output, y t Realistic navigation or obstacle avoidance results.

[0138] The weights of the deep learning model are updated in real time during actual navigation. The vehicle acquires new environmental data through sensors during operation and makes predictions based on path planning and obstacle avoidance decisions. By comparing the model's predictions with actual results, the system ensures that the model can quickly learn and adapt to new environmental changes during actual operation. In dynamic environments, environmental conditions are constantly changing, such as the sudden appearance of new obstacles or changes in target positions. Through an online learning mechanism, the vehicle can continuously update the model weights and dynamically adjust obstacle avoidance strategies during real-time operation.

[0139] The controller executes dynamic adjustments to the underwater vehicle using multi-channel control signals. The controller algorithm formula is as follows:

[0140] Thruster speed control formula: v new =v current +K v ·(v desired -v current ),

[0141] Among them, v new Adjusted thruster speed, v current Thruster speed, v desired Desired thruster speed, K v Speed ​​adjustment gain coefficient;

[0142] Heading control formula (based on proportional-integral-derivative control, PID):

[0143]

[0144] Where, θ new Adjusted heading angle, θ current Heading angle, e(t) heading error, K p K i K d These are the proportional, integral, and derivative coefficients of the PID controller.

[0145] Attitude control formula: φ new =φ current +K φ ·(φ desired -φ current ),

[0146] Where, φ new Adjusted attitude angle, v current Attitude angle, φ desired Desired attitude angle, K φ Attitude control gain coefficient;

[0147] Total control signal formula: u control =[v new θnew , φ new ],

[0148] The formula for the total control signal represents the final control signal u. control It is determined by the thruster speed v new Heading angle θ new and attitude angle φ new The combined multi-channel control signals are used to adjust the motion and attitude of the aircraft.

[0149] Speed ​​control ensures the vehicle travels at the optimal speed, preventing excessive speed fluctuations and optimizing energy consumption. Heading control ensures the vehicle travels along the planned path, adjusting the heading through PID control to prevent deviation from the target route. Attitude control maintains the vehicle's stable attitude by dynamically adjusting attitude angles, ensuring motion stability in complex underwater environments. The overall control signal integrates all control signals to ensure the vehicle can flexibly and smoothly adjust its speed, direction, and attitude, enabling it to operate efficiently in dynamically changing underwater environments and achieve autonomous obstacle avoidance and path planning.

[0150] Data feedback and real-time optimization: The system monitors the vehicle's sensor data and path execution in real time, and dynamically optimizes path planning and obstacle avoidance strategies through a feedback mechanism. The following is the algorithm formula for real-time data feedback and optimization:

[0151] Real-time optimization feedback formula (based on feedback control theory):

[0152] ΔP adjust =K f ·(P desired -P current ),

[0153] Where, ΔP adjust The correction value, P, is used to adjust the path planning. desired Desired path point, P current Actual path point, K f Feedback gain coefficient;

[0154] Path error calculation formula:

[0155]

[0156] Among them, e P (t) Path error, x desired y desired z desired The three-dimensional coordinates of the desired location, x current y current z current The three-dimensional coordinates of the actual location;

[0157] Real-time optimization and adjustment formula:

[0158] P new =P current +ΔP adjust -η·e P (t),

[0159] Among them, P new The optimized new path point, η is the learning rate.

[0160] Through a feedback control mechanism, the formula can adjust the path planning in a timely manner according to the actual operation of the vehicle, ensuring that the vehicle can quickly correct itself when it deviates from the expected path. The dynamic adaptive adjustment mechanism can cope with path deviations caused by uncertainties in complex underwater environments, ensuring that the vehicle can navigate along the optimal path. The deviation between the vehicle's position and the expected path point is clearly measured. The larger the error value, the farther the vehicle deviates from the planned path. The system makes corresponding adjustments based on the error, ensuring that the navigation system can always monitor the path execution in real time and prevent serious deviations.

[0161] Path tracking and monitoring is based on high-precision positioning data provided by the BeiDou Navigation Satellite System. It monitors the deviation between the vehicle and the target path in real time and makes timely adjustments. The following is the algorithm formula for path tracking and monitoring:

[0162] Path tracking error formula:

[0163]

[0164] Among them, e track (t) Tracking error between the vehicle and the target waypoint, x target y target z target The desired three-dimensional coordinates of the target point, x BDS y BDS z BDS Three-dimensional coordinates of the aircraft obtained through the BeiDou Navigation Satellite System;

[0165] Path correction formula: P adjust =P BDS +K p ·e track (t),

[0166] Among them, P adjust Adjusted path point, P BDS The BeiDou positioning coordinates of the aircraft, K p Proportional gain;

[0167] Heading adjustment formula: θ new =θ current +K θ ·(θ target -θ current ),

[0168] Where, θ new Adjusted heading angle, θ current The heading angle of the aircraft, θ target Desired heading angle, K θ Proportional gain coefficient for heading adjustment;

[0169] Overall control update formula: u control =[P adjust θ new ],

[0170] Among them, u control Adjustment signals used to control the aircraft.

[0171] The path tracking error formula monitors the distance deviation between the vehicle and the target waypoint in real time, ensuring that the vehicle can detect deviations from the planned route in a timely manner. The error value provides feedback information, indicating the degree of path correction required. The path correction formula generates adjusted waypoints based on the error, dynamically correcting the vehicle's position to ensure it returns to the correct route. Through proportional gain, the correction intensity can be controlled to avoid over-correction or slow response. The heading adjustment formula ensures that the vehicle's direction is always towards the target. If the vehicle's heading differs from the desired heading, the system adjusts the heading angle using this formula to keep the vehicle's direction correct, ensuring the accuracy of path tracking. The overall control update formula integrates path position and heading adjustment to generate a multi-channel control signal. This signal synchronously controls the vehicle's position and heading, enabling the vehicle to accurately execute the planned path and continuously maintain the correct navigation status through real-time adjustments.

[0172] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based underwater vehicle path planning and obstacle avoidance method, characterized in that: The method comprises the following steps: S1, acquiring multimodal data of underwater environment through underwater sensors and Beidou satellite navigation system, and generating a high-precision environment model in real time; S2, using multi-sensor fusion algorithms to fuse the global positioning information provided by Beidou satellites with data from sonar, inertial navigation systems, and visual SLAM; S3, process environmental perception data through deep learning models, use convolutional neural networks to identify and classify obstacles in the three-dimensional environment obtained by high-definition cameras and lidar, and predict navigation paths; S4. Based on the deep reinforcement learning model, the obstacle avoidance path is planned in real time in a dynamic environment. The obstacle avoidance strategy is adjusted by the reward function in reinforcement learning. The specific formula is: R=-λ1·d obstacle +λ2·d goal -λ3·C collision , Among them, R is the reward value, d obstacle is the distance between the aircraft and the obstacle, d goal is the distance between the aircraft and the target point, C collision is the collision penalty coefficient, λ1, λ2, λ3 are weight parameters; S5. Using the global position information provided by the Beidou navigation system, if the aircraft deviates from the global planned path, the global path correction module is automatically started to re-plan the path based on the Beidou satellite data; S6. Based on the turning radius and physical size of the aircraft, the path smoothing strategy of "arc-straight line-arc" is used to optimize the path output by deep learning; S7. Through the online learning mechanism, the vehicle can make adaptive adjustments based on new environmental data during execution and update the weights of the deep learning model; S8, generating motion control instructions for the aircraft based on the results of the planned path and real-time obstacle avoidance output, and controlling the propeller and steering gear of the aircraft to execute the motion; S9. During the operation of the aircraft, continuously obtain environmental perception data and update the model in real time; S10. Monitor the deviation between the position of the aircraft and the target path through the Beidou navigation system, and enable the emergency obstacle avoidance mode in an emergency.

2. The underwater vehicle path planning and obstacle avoidance method based on deep learning according to claim 1, characterized in that: In step S2, the algorithm formula for correcting the positioning error is: P t+1 =P t +α·(P BDS -P t )+β·(P INS -P t )+γ·(P SLAM -P t ), Among them, P t+1 is the corrected position, P BDS , P INS , P SLAM It is the location information provided by Beidou navigation, inertial navigation and visual SLAM, and α, β, and γ are weighting coefficients.

3. The underwater vehicle path planning and obstacle avoidance method based on deep learning according to claim 1, characterized in that: The multi-sensor fusion algorithm is used to combine the data of the Beidou navigation system, the inertial navigation system and the visual SLAM to enhance the positioning accuracy of the underwater vehicle. The following is the formula of the fusion algorithm: P fused =α·P BDS +β·P INS +γ·P SLAM , Among them, P fused The final positioning result after fusion, P BDS The positioning information obtained through the Beidou satellite navigation system, P INS The positioning information obtained by the inertial navigation system, P SLAM The positioning information obtained through visual SLAM, the weighting coefficients of the α, β, and γ system data.

4. The underwater vehicle path planning and obstacle avoidance method based on deep learning according to claim 1, characterized in that: The path planning uses a convolutional neural network combined with lidar data for environmental perception and dynamically adjusts the planned path. The following is the algorithm formula in the path planning: The environment perception and path planning formula: Among them, P optimal The optimal path after convolutional neural network and lidar perception optimization, candidate points of P path, d goal The distance from the path point to the target point, d obstacle The distance from the path point to the surrounding obstacles, C smooth The smoothness penalty function of the P path, λ1, λ2, and λ3 control the weight coefficients of each factor.

5. The underwater vehicle path planning and obstacle avoidance method based on deep learning according to claim 1, characterized in that: The Beidou navigation-assisted global path correction module is used to automatically re-plan the path when the aircraft deviates from the predetermined path. The following is the algorithm formula involved: Global path correction formula: P new =P current +α·(P BDS -P current )+β·(P planned -P current ), Among them, P new The corrected new path point, P current The position of the aircraft, P BDS The precise global location information obtained by the Beidou navigation system, P planned The target point on the predetermined planning path, α, β weight coefficients.

6. The underwater vehicle path planning and obstacle avoidance method based on deep learning according to claim 1, characterized in that: The path smoothing and optimization uses the "arc-straight-arc" strategy to generate a smooth path of the physical characteristics of the spacecraft. The following is the algorithm formula for calculating the smooth path: Arc-straight-arc path optimization formula: Arc segment calculation: Among them, θ arc The angle of the arc segment, L arc The length of the arc segment, r is the turning radius of the vehicle; Calculation of straight line segment length: Among them, L line The straight-line distance between arc segments; the coordinates of the start and end points of the x1, y1 and x2, y2 straight-line segments; Path smoothness optimization function: C smooth =λ1·θ arc +λ2·L line -λ3·(1 / r), Among them, C smooth Path smoothness cost function, λ1, λ2, λ3 weight coefficients.

7. The underwater vehicle path planning and obstacle avoidance method based on deep learning according to claim 1, characterized in that: The dynamic self-learning adjustment updates the weights of the deep learning model in real time through an online learning mechanism. The following is the algorithm formula of the process: Online learning weight update formula: Among them, w t+1 The updated model weight at time t+1, w t The model weight at time t, η learning rate, Loss function L for weight w t The gradient of the model indicates that the model is t and the true label y t The error derivative on ; Real-time environment input: x t : Input data collected by the sensor or vision system in this environment; y t : The actual obstacle avoidance path or successful navigation result, used as a label in supervised learning to correct the model's prediction; Loss function: Among them, f(w t , x t ) The deep learning model takes input x t The predicted output, y t Actual navigation or obstacle avoidance results.

8. The underwater vehicle path planning and obstacle avoidance method based on deep learning according to claim 1, characterized in that: The controller performs dynamic adjustment of the underwater vehicle through multi-channel control signals. The following is the controller algorithm formula: Thruster speed control formula: v new =v current +K v ·(v desired -v current ), Among them, v new Adjusted thruster speed, v current Propeller speed, v desired Desired thruster speed, K v Speed ​​adjustment gain factor; Heading control formula (based on proportional-integral-derivative control, PID): Among them, θ new Adjusted heading angle, θ current Heading angle, e(t) Heading error, K p , K i , K d are the proportional, integral and differential coefficients of the PID controller; Attitude control formula: φ new =φ current +K φ ·(φ desired -φ current ), Among them, φ new Adjusted attitude angle, φ current Attitude angle, φ desired Desired attitude angle, K φ Attitude control gain coefficient; Total control signal formula: u control =[v new ,θ new ,φ new ], The total control signal formula represents the final control signal u control , is given by the propeller speed v new , heading angle θ new and attitude angle φ new The multi-channel control signals together are used to adjust the motion and attitude of the spacecraft.

9. The underwater vehicle path planning and obstacle avoidance method based on deep learning according to claim 1, characterized in that: The data feedback and real-time optimization monitors the sensor data and path execution of the aircraft in real time, and dynamically optimizes the path planning and obstacle avoidance strategy through the feedback mechanism. The following is the algorithm formula for real-time data feedback and optimization: Real-time optimization feedback formula (based on feedback control theory): ΔP adjust =K f ·(P desired -P current ), Among them, ΔP adjust Correction value used to adjust path planning, P desired Expected path point, P current Actual path point, K f Feedback gain coefficient; Path error calculation formula: Among them, e P (t) Path error, x desired ,y desired 、z desired The three-dimensional coordinates of the desired position, x current ,y current 、z current The three-dimensional coordinates of the actual position; Real-time optimization adjustment formula: P new =P current +ΔP adjust -η·e P (t), Among them, P new The new path point after optimization, η is the learning rate.

10. The underwater vehicle path planning and obstacle avoidance method based on deep learning according to claim 1, characterized in that: The path tracking and monitoring is based on the high-precision positioning data provided by the Beidou navigation system, which monitors the deviation between the aircraft and the target path in real time and makes timely adjustments. The following is the algorithm formula for path tracking and monitoring: Path tracking error formula: Among them, e track (t) Tracking error between the vehicle and the target path point, x target ,y target 、z target The three-dimensional coordinates of the desired target point, x BDS ,y BDS 、z BDS The three-dimensional coordinates of the aircraft obtained through the Beidou navigation system; Path correction formula: P adjust =P BDS +K p ·e track (t), Among them, P adjust The adjusted path point, P BDS Beidou positioning coordinates of the aircraft, K p Proportional gain; Course adjustment formula: θ new =θ current +K θ ·(θ target -θ current ), Among them, θ new Adjusted heading angle, θ current Aircraft heading angle, θ target Desired heading angle, K θ Proportional gain coefficient for heading adjustment; Overall control update formula: u control =[P adjust ,θ new ], Among them, u control Adjustment signals used to control the aircraft.

Images