Artificial Intelligence-Based AUV Underwater Navigation Method and System
By generating a global reference path through artificial potential field and heuristic search, and combining line-of-sight guidance and SAC reinforcement learning network to optimize AUV path tracking, the independent design problem of path planning and tracking control in AUV underwater navigation is solved, improving navigation accuracy and autonomy, and realizing adaptive decision-making and control in complex environments.
Patent Information
- Application Number
- CN202611239990.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-10
AI Technical Summary
Existing AUV underwater navigation methods are prone to getting trapped in local minima during path planning, making targets unreachable. Furthermore, they suffer from large attitude accuracy errors during path tracking control, making it difficult to cope with time-varying ocean current interference and model uncertainties. They also lack a closed-loop collaborative mechanism for global path planning and local motion correction.
An artificial potential field algorithm is used to construct an environmental cost potential field. A global reference path set is generated by combining heuristic search. The path tracking deviation is calculated by the line-of-sight guidance algorithm and local environmental perception data is fused. The SAC reinforcement learning network is used to output the expected motion correction data. The control input is generated by combining dynamics deduction and model predictive control to drive the thruster thrust distribution to achieve full-link collaborative optimization.
It improves the navigation accuracy, autonomy, and environmental adaptability of AUVs in complex underwater environments, solves the independent design problem of path planning and tracking control in traditional methods, enhances the dynamic environment information perception capability and path tracking accuracy, and realizes the feasibility and optimality of adaptive online decision-making and control solutions.
Smart Images

Figure CN123065467A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of AUV underwater navigation, and in particular relates to an AUV underwater navigation method and system based on artificial intelligence. Background Technology
[0002] With the development of technologies in fields such as marine resource exploration, subsea pipeline inspection, and marine environmental monitoring, Autonomous Underwater Vehicles (AUVs) have become important platforms for performing underwater operations. Their autonomous navigation and path-following control capabilities directly determine the efficiency and safety of mission execution. Currently, AUV underwater navigation methods are mainly divided into two levels: path planning and path-following control. In path planning, traditional methods include artificial potential field methods and graph search methods, among which the artificial potential field method is widely used due to its computational simplicity. In path-following control, line-of-sight guidance algorithms are commonly used for AUV heading control due to their simple structure and ease of implementation; model predictive control is used for motion control because it can handle system constraints.
[0003] However, the aforementioned traditional methods have many shortcomings. At the path planning level, the artificial potential field method is prone to getting trapped in local minima and suffers from the problem of unreachable targets, making it difficult to generate optimal paths in complex obstacle environments. At the path tracking control level, line-of-sight guidance algorithms have large errors in attitude accuracy and are difficult to effectively cope with time-varying ocean current disturbances and model uncertainties; while model predictive control can handle constraints, its control performance deteriorates significantly under strong ocean currents and wave disturbances. In addition, AUV dynamic models have characteristics such as high nonlinearity, underactuation, time-varying model parameters, and unmodeled dynamics, making it difficult for traditional methods to achieve accurate modeling and adaptive control. More importantly, in existing methods, path planning and path tracking control are mostly designed independently, lacking an effective closed-loop coordination mechanism, making it difficult to achieve end-to-end adaptive navigation from global path planning to local motion correction in dynamically changing underwater environments. Summary of the Invention
[0004] Therefore, it is necessary to provide an AI-based AUV underwater navigation method and system that can solve the above problems.
[0005] Firstly, this application provides an artificial intelligence-based AUV underwater navigation method, including:
[0006] Acquire prior environmental data, navigation mission data, AUV real-time status data, local environmental perception data, and AUV dynamic model parameters;
[0007] Based on prior environmental data, an environmental cost potential field is constructed using an artificial potential field algorithm. Then, based on navigation mission data and the environmental cost potential field, a global reference path set is generated using a heuristic search algorithm.
[0008] Based on the AUV's real-time status data and global reference path set, path tracking deviation data is calculated using a line-of-sight guidance algorithm. Then, local environment perception data and path tracking deviation data are fused using an attention mechanism to obtain a joint state vector.
[0009] The joint state vector is input into a pre-trained SAC reinforcement learning network, which outputs the desired motion correction data.
[0010] Based on the AUV's real-time status data, AUV dynamic model parameters, and desired motion correction data, a target state vector is generated through dynamic deduction. Then, based on the AUV's real-time status data and the target state vector, the control input is obtained through a model predictive control algorithm.
[0011] The thruster thrust distribution data is generated based on the control inputs and is used to drive the AUV.
[0012] Secondly, this application also provides an artificial intelligence-based AUV underwater navigation system, comprising:
[0013] The data acquisition module is used to acquire prior environmental data, navigation mission data, AUV real-time status data, local environmental perception data, and AUV dynamic model parameters.
[0014] The path planning module is used to construct an environmental cost potential field based on prior environmental data using an artificial potential field algorithm, and to generate a global reference path set based on navigation mission data and the environmental cost potential field using a heuristic search algorithm.
[0015] The guidance fusion module is used to calculate path tracking deviation data based on AUV real-time status data and global reference path set through line-of-sight guidance algorithm, and to fuse local environment perception data and path tracking deviation data through attention mechanism to obtain joint state vector.
[0016] The learning correction module is used to input the joint state vector into the pre-trained SAC reinforcement learning network and output the desired motion correction data.
[0017] The predictive control module is used to generate a target state vector based on the AUV's real-time state data, AUV dynamic model parameters, and desired motion correction data through dynamic deduction, and to solve for the control input quantity based on the AUV's real-time state data and target state vector through a model predictive control algorithm.
[0018] The thrust distribution module is used to generate thruster thrust distribution data based on control inputs. This thruster thrust distribution data is used to drive the AUV.
[0019] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described artificial intelligence-based AUV underwater navigation method.
[0020] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described artificial intelligence-based AUV underwater navigation method.
[0021] The aforementioned AI-based AUV underwater navigation method and system generate a global reference path set through artificial potential fields and heuristic search, providing globally optimal navigation references for AUVs while addressing the problem of artificial potential field methods easily getting trapped in local minima. By fusing line-of-sight guidance and attention mechanisms to generate a joint state vector, the system enhances the AUV's perception and representation of dynamic environmental information, improving path tracking accuracy and robustness. Inputting the joint state vector into a pre-trained SAC reinforcement learning network, utilizing its entropy maximization and exploration characteristics to output desired motion correction data, enables the AUV to possess adaptive online decision-making and motion adjustment capabilities in complex and unknown environments. Combining dynamic deduction and model predictive control to solve control inputs ensures the feasibility and optimality of the control solution while satisfying system dynamic constraints and time-varying ocean current interference. Driving AUV execution through thruster thrust allocation data achieves end-to-end collaborative optimization from planning to control, improving the AUV's navigation accuracy, autonomy, and environmental adaptability in complex underwater environments. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 This is a flowchart illustrating the steps of the AI-based AUV underwater navigation method of the present invention.
[0024] Figure 2 A diagram illustrating the target state vector generation steps of an AI-based AUV underwater navigation method in one embodiment of the present invention.
[0025] Figure 3This is a structural diagram of the AI-based AUV underwater navigation system of the present invention. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0027] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based underwater navigation method for AUVs is provided. This embodiment illustrates the application of this method to a navigation terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. The implementation environment of this application includes the AUV itself, navigation sensors mounted on the AUV (such as an inertial navigation system, Doppler velocimeter, depth gauge, forward-looking sonar, hydrophone, etc.), and a navigation terminal. The navigation terminal can be mounted on the AUV itself, on shore, or on the mother ship. Typical application scenarios include: when the AUV performs subsea pipeline inspection or marine resource exploration tasks, the navigation terminal acquires pre-loaded environmental prior data and navigation task data, and simultaneously collects AUV attitude, velocity, and information on local obstacles and ocean currents through the navigation sensors; the navigation terminal runs this method to generate thrust distribution commands, driving the AUV to navigate autonomously.
[0028] In this embodiment, the method includes the following steps:
[0029] S01, acquire prior environmental data, navigation mission data, AUV real-time status data, local environmental perception data, and AUV dynamic model parameters.
[0030] Optionally, prior environmental data refers to a dataset of prior environmental features collected before AUV diving operations using sensors such as acoustic Doppler current profilers, temperature, salinity, and depth sensors, side-scan sonar, and seabed topography mapping equipment. This dataset encompasses seabed topography, spatial distribution of static obstacles, and environmental ocean current field information over historical statistical periods. Navigation mission data refers to a set of mission parameters pre-set according to specific operational requirements. AUV real-time status data refers to the current motion status information calculated and output in real-time by the navigation sensors onboard the AUV. Local environmental perception data refers to dynamic obstacle information and real-time ocean current velocity vector information within a local area surrounding the AUV, collected in real-time by environmental perception sensors such as forward-looking sonar, side-scan sonar, and collision avoidance sonar. AUV dynamic model parameters refer to a set of mathematical model parameters used to describe the motion characteristics of the AUV itself.
[0031] Optionally, the navigation terminal can obtain prior environmental data, navigation mission data, AUV real-time status data, local environmental perception data, and AUV dynamic model parameters through a data interface, and preprocess them as the basic input for subsequent path planning and motion control.
[0032] S02. Based on prior environmental data, an environmental cost potential field is constructed using an artificial potential field algorithm. Then, based on the navigation mission data and the environmental cost potential field, a global reference path set is generated using a heuristic search algorithm.
[0033] Optionally, the artificial potential field algorithm is a calculation method that virtualizes obstacles and target points in the environmental space as repulsive and gravitational fields, respectively; the environmental cost potential field is a scalar function space used to characterize the navigation cost at each location within the space represented by the environmental prior data; the heuristic search algorithm is a graph search algorithm that uses heuristic functions to guide node expansion to search for the optimal path; the global reference path set refers to a set of several feasible paths that satisfy navigation constraints and connect the starting point and the destination.
[0034] Specifically, the navigation terminal can extract the locations of static obstacles and the distribution characteristics of ocean currents based on prior environmental data. Using an artificial potential field algorithm, the influence of static obstacles is constructed as a repulsive potential field component, and the influence of ocean currents on navigation safety is constructed as a flow field risk potential field component. These potential field components are then weighted and fused to obtain an environmental cost potential field reflecting the navigation safety and passage difficulty at various locations in space. Based on this, the navigation terminal can use the starting point and ending point in the navigation mission data as search boundary conditions, and use the potential field value of the environmental cost potential field as a cost mapping for node expansion. Combining this with a preset heuristic function, the cost from the current node to the ending point is estimated. Through a heuristic search algorithm, nodes are expanded and backtracked from the starting point to the ending point in a discretized spatial grid, generating a global reference path set covering multiple feasible reference paths, providing AUVs with reference navigation paths from a global perspective.
[0035] S03. Based on the AUV's real-time status data and the global reference path set, the path tracking deviation data is calculated using a line-of-sight guidance algorithm. The local environment perception data and the path tracking deviation data are then fused using an attention mechanism to obtain a joint state vector.
[0036] Optionally, the Line of Sight (LOS) algorithm is a path tracking guidance method based on geometric principles. It transforms the path tracking problem into a heading angle tracking control problem by calculating the line-of-sight direction between the AUV's current position and a virtual target point on the reference path. Path tracking deviation data refers to the spatial deviation of the AUV's actual flight path relative to the global reference path. The attention mechanism is a data processing strategy that simulates human visual attention. By assigning differentiated weights to different parts of the input data, it enables the network to focus on the information features most critical to the current decision. The joint state vector is a comprehensive state representation vector formed by concatenating or fusing multi-source heterogeneous perception and state data according to a unified dimension, and is used as input to the subsequent reinforcement learning decision network.
[0037] Specifically, the navigation terminal can calculate path tracking deviation data using a line-of-sight guidance algorithm based on the AUV's real-time status data and a global reference path set. The navigation terminal can input both local environmental perception data and path tracking deviation data into the attention mechanism module. The attention mechanism differentiates and weights the spatial features in the local environmental perception data, strengthening the feature representation of near-field obstacles and adverse flow field components that significantly impact navigation safety. Simultaneously, it assigns corresponding attention weights to each deviation component in the path tracking deviation data according to its importance to path tracking control. The weighted local environmental perception features and path tracking deviation features are then concatenated or fused according to a preset dimensional order to generate a joint state vector. This joint state vector comprehensively represents the AUV's current tracking status relative to the reference path and the risk situation of the surrounding local environment, providing a complete, compact, and information-enhanced state input for the subsequent reinforcement learning decision-making module.
[0038] Optionally, path tracking deviation data is calculated using a line-of-sight guidance algorithm, including:
[0039] Select the reference path point closest to the current position of the AUV from the global reference path set as the matching point, and obtain the position coordinates of the matching point in the global coordinate system. and heading angle ;
[0040] Based on the current position in the AUV real-time status data and the heading angle in the current attitude angle Calculate the line of sight angle :
[0041] ;
[0042] in, This represents the lateral tracking deviation from the current position of the AUV to the reference path segment where the matching point is located. The preset forward sight distance is 2-5 times the AUV length; lateral tracking deviation. The following method is used to calculate the result: Project the position difference vector between the matching point and the current position of the AUV onto the normal direction of the reference path segment where the matching point is located. Path tracking deviation data includes the following components: lateral tracking deviation Longitudinal tracking deviation Depth tracking deviation Heading angle deviation and the aforementioned line of sight angle Among them, longitudinal tracking deviation The depth tracking deviation is the projection component of the position difference vector between the matching point and the current position of the AUV onto the tangent direction of the reference path segment where the matching point is located. This is the difference between the current depth of the AUV and the depth of the matching point.
[0043] Optionally, the local environment perception data and the path tracking deviation data are fused through an attention mechanism to obtain a joint state vector, including:
[0044] The local environment perception data is encoded into an environmental feature vector. ,in For environmental features, the path tracking deviation data is encoded into a deviation feature vector. ,in The deviation feature dimension is used; a cross-attention mechanism is employed to calculate the attention weights of the deviation feature vector on the environmental feature vector.
[0045] ;
[0046] in, To query the weight matrix, The key weight matrix is... For the dimension of attention head, The softmax activation function is used. Let be the attention weight coefficient of the i-th environmental feature component; based on the attention weight coefficient, the environmental feature vector is weighted and aggregated to obtain the weighted environmental features. :
[0047] ;
[0048] in, Environmental feature vector The i-th component; the weighted environmental features The deviation feature vector The current velocity vector and current attitude angle from the AUV's real-time status data are concatenated according to a preset dimensional order to obtain the joint state vector. :
[0049] ;
[0050] Where v is the current velocity vector, The current attitude angle is given, and the dimension of the joint state vector is the sum of the dimensions of its constituent parts.
[0051] S04 inputs the joint state vector into the pre-trained SAC reinforcement learning network and outputs the desired motion correction data.
[0052] Optionally, the SAC reinforcement learning network (Soft Actor-Critic network) is a deep reinforcement learning model based on the maximum entropy reinforcement learning framework. Its key feature is the introduction of an entropy regularization term into the policy optimization objective, encouraging the policy to maximize action entropy while accumulating expected rewards, thereby enhancing the policy's exploratory nature and robustness. Expected motion correction data refers to the set of correction values output by the SAC reinforcement learning network based on the current joint state vector, used to adjust the preset motion commands of the AUV.
[0053] Specifically, the navigation terminal can input the joint state vector into a pre-trained SAC reinforcement learning network. The Actor module (policy network) in this network receives the joint state vector and, through a nonlinear mapping of multiple hidden layers in a neural network, outputs an action probability distribution parameter that conforms to the Gaussian distribution assumption. Based on the optimization objective of maximizing entropy-regularized rewards, the SAC reinforcement learning network balances expected reward with action entropy during policy updates, samples the optimal action from the action probability distribution parameter, and simultaneously outputs value assessment information to guide the next step of online learning. The navigation terminal then processes the sampled action through scaling and amplitude limiting as expected motion correction data. By introducing the pre-trained SAC reinforcement learning network, the AUV can adaptively generate motion corrections online in complex, time-varying underwater environments based on path tracking deviations and local environmental perception information contained in the joint state vector, thus overcoming the shortcomings of traditional line-of-sight guidance algorithms and fixed-parameter controllers in decision-making under dynamic disturbances.
[0054] Optionally, the SAC reinforcement learning network employs a maximum entropy reinforcement learning framework, and its reward function during training... Defined as:
[0055] Where s is the joint state vector, and a is the action vector output by the SAC reinforcement learning network. to The preset regularization coefficients are used. The lateral tracking deviation, The heading angle deviation is... For speed increments, For the heading angle increment, This is a safety indication function. It takes a value of 1 when the distance between the AUV and the obstacle is greater than the safety threshold, and a value of 0 otherwise.
[0056] The SAC reinforcement learning network's Actor module consists of four fully connected hidden layers, with 256, 256, 128, and 128 neurons in each layer. Each hidden layer uses the ReLU activation function, and the output layer uses the Tanh activation function to map action values to the [-1, 1] interval. The input layer dimension of the Actor module is equal to the dimension of the joint state vector, and the output layer dimension is 2, corresponding to the velocity increments. and bow angle increment Pre-training is performed in a simulation environment, which is constructed based on the prior environmental data and AUV dynamics model parameters. A soft policy iterative method is used to alternately update the Actor module and the Critic module.
[0057] S05. Based on the AUV's real-time state data, AUV dynamic model parameters, and desired motion correction data, a target state vector is generated through dynamic deduction. Then, based on the AUV's real-time state data and the target state vector, the control input is obtained through a model predictive control algorithm.
[0058] Optionally, dynamic deduction refers to the process of predicting the motion state of an AUV at future times using a dynamic model of the AUV, given the current state and control input conditions, through numerical integration or state recursion. The target state vector refers to the desired state vector at a future time, used as the control objective, within the forward predicted state sequence obtained through dynamic deduction. Model predictive control (MMC) is an advanced control method based on a rolling optimization strategy. It obtains the optimal control sequence that satisfies system dynamic constraints and control quantity limits by solving a constrained optimal control problem in a finite time domain at each sampling time.
[0059] Specifically, the navigation terminal can perform dynamic deduction based on the AUV's real-time state data and desired motion correction data, using the AUV's dynamic model parameters. It then recursively solves for the AUV's state response within a preset prediction time domain using numerical integration, generating a forward predicted state sequence containing state vectors at multiple time steps. The target state vector is then extracted from this sequence. The navigation terminal can use the deviation between the target state vector and the AUV's real-time state data as the optimization objective to construct a model predictive control optimization problem. The AUV's dynamic model parameters are used as the prediction model, and the thruster's physical characteristics are transformed into control amplitude and rate-of-change constraints. This constrained optimization problem is solved online within each control cycle to obtain the optimal control sequence that satisfies the constraints and minimizes the cost function. The first control variable in this sequence is then extracted as the current control input output. By combining dynamic deduction with model predictive control, it is ensured that the generated control input accurately tracks the desired motion correction command while satisfying the AUV's dynamic constraints and thruster physical limitations, achieving a dynamic feasibility mapping from the decision-making layer to the execution layer.
[0060] S06, Generate thruster distribution data based on control inputs. The thruster distribution data is used to drive the AUV.
[0061] Optionally, the thruster thrust allocation data refers to the allocation result of mapping the control input obtained by the model predictive control algorithm to the thrust commands of each AUV thruster. This data includes the thrust amplitude and direction commands that each thruster needs to output.
[0062] Specifically, the navigation terminal can obtain the thruster efficiency data and thruster output limit data of the AUV configuration. Based on the installation position, installation angle and thrust direction of each thruster, it constructs a thrust allocation matrix and maps the control input to the thruster thrust allocation data of each thruster through a control allocation algorithm (such as pseudo-inverse method or quadratic programming method). The navigation terminal can output the thruster thrust allocation data to the AUV's underlying drive system to drive the thrusters to execute the corresponding thrust output and propel the AUV to sail according to the desired motion state.
[0063] The aforementioned AI-based AUV underwater navigation method acquires multi-source information, including prior environmental data, navigation mission data, AUV real-time status data, local environmental perception data, and AUV dynamic model parameters. It then uses an artificial potential field and heuristic search to generate a global reference path set, providing globally optimal navigation reference for the AUV. By calculating path tracking deviation through line-of-sight guidance and fusing local environmental perception data using an attention mechanism, a joint state vector is formed, enhancing the representation of dynamic environmental information and improving path tracking accuracy and disturbance resistance. The joint state vector is input into a pre-trained SAC reinforcement learning network, utilizing its maximum entropy exploration characteristics to output desired motion corrections, endowing the AUV with the ability to make adaptive online decisions in complex and unknown environments, thus compensating for the shortcomings of traditional fixed-parameter controllers in dealing with time-varying disturbances. Through dynamic deduction and model predictive control to solve for control inputs, the optimality and feasibility of the control solution are ensured while satisfying system dynamic constraints, thruster saturation, and ocean current interference. Finally, thrust allocation is generated to drive execution, achieving end-to-end collaborative optimization across the entire link from global planning to local correction and control execution, thereby improving the AUV's navigation accuracy, autonomy, and environmental adaptability in complex underwater environments.
[0064] In one embodiment, an environmental cost potential field is constructed based on prior environmental data using an artificial potential field algorithm, including:
[0065] S11, based on prior environmental data, extract static obstacle location data and environmental ocean current data;
[0066] S12, based on prior environmental data, calculates the spatial variation gradient of ocean current velocity within the spatial range represented by the prior environmental data, and sets the fine-grained spatial division according to the magnitude of the spatial variation gradient.
[0067] S13, Based on the fine-grained spatial division, the spatial range represented by the prior environmental data is discretized to obtain multiple partitioned spaces;
[0068] S14. Based on static obstacle location data and environmental ocean current data, calculate the relative orientation vector of each partitioned space with respect to each static obstacle, and calculate the ocean current velocity vector of the environmental ocean current data in each partitioned space.
[0069] S15, based on the relative azimuth vector and ocean current velocity vector, calculates the directional consistency metric value, and based on the directional consistency metric value, maps the flow field risk coefficient of each partitioned space through a piecewise mapping function;
[0070] S16, Based on the risk coefficients of each flow field, the flow field risk potential field is constructed;
[0071] S17. Based on static obstacle location data, combined with spatial division and preset risk action radius, a static obstacle repulsive potential field is constructed.
[0072] S18, through a preset weighted fusion strategy, the potential field values of the static obstacle repulsive potential field and the flow field risk potential field are weighted and fused to obtain the environmental cost potential field.
[0073] For example, static obstacle location data refers to data representing the coordinate position information of fixed obstacles such as seabed topographic protrusions, shipwrecks, and reefs in navigation space. Environmental ocean current data refers to historical statistics or prior observation data describing the spatial distribution of ocean current velocity magnitude and direction within the navigation area. Spatial variation gradient refers to the vector field composed of the partial derivatives of the ocean current velocity vector along each coordinate axis within the spatial range represented by the environmental prior data, used to measure the drastic change in ocean current velocity in space. Spatial partitioning fine-grainedness refers to the quantified value of the side length or resolution of each grid cell in each dimension of space when discretizing the spatial range represented by the environmental prior data. Spatial partitioning refers to the individual grid cells obtained after discretizing the continuous spatial range represented by the environmental prior data according to the spatial partitioning fine-grainedness.
[0074] A relative azimuth vector is a vector pointing from a defined spatial area to the location of a static obstacle, used to characterize the spatial orientation of the defined space relative to the static obstacle. An ocean current velocity vector is the ocean current velocity vector given by environmental ocean current data at a defined spatial area. A directional consistency metric is a scalar index used to quantify the degree of consistency between the relative azimuth vector and the ocean current velocity vector. A piecewise mapping function is a nonlinear mapping rule that maps input values to corresponding output values according to their numerical intervals. A flow field risk coefficient is a quantitative index characterizing the degree of risk posed by ocean currents to AUV navigation safety at each defined spatial area. A flow field risk potential field is a scalar potential field formed by organizing the flow field risk coefficients of each defined space according to their spatial positional relationships, used to describe the distribution of spatial navigation risks caused by ocean currents.
[0075] The static obstacle repulsive potential field refers to the potential field constructed based on static obstacle location data, used to characterize the repulsive effect of static obstacles on AUVs. The weighted fusion strategy refers to the rule of summing the potential field values of multiple potential fields in the same partitioned space according to preset weight coefficients to obtain a comprehensive potential field value.
[0076] Optionally, in S11, the navigation terminal can parse and extract prior environmental data, extract information describing the location of static obstacles to obtain static obstacle location data, and at the same time extract information describing the spatial distribution of environmental ocean currents to generate environmental ocean current data.
[0077] Optionally, in S12, the navigation terminal can calculate the rate of change of the ocean current velocity vector at each location within the spatial range characterized by the prior environmental data along each coordinate axis in space, based on the environmental ocean current data. That is, it can solve for the Jacobian matrix or partial derivative of the ocean current velocity vector field to obtain the spatial gradient of ocean current velocity. The amplitude of the gradient is calculated, and the spatial division fineness is adaptively set according to the magnitude of the amplitude, so that a finer division granularity is used in areas where ocean current velocity changes drastically, and a coarser division granularity is used in areas where the change is gradual.
[0078] Optionally, in S13, the navigation terminal can discretize the entire spatial range represented by the prior environmental data in each dimension of the three-dimensional space according to the fine-grained spatial division, dividing it into several non-overlapping grid units, with each grid unit serving as a partition space.
[0079] Optionally, in S14, for each partitioned space, the navigation terminal can calculate the relative azimuth vector from the partitioned space to each static obstacle using the spatial coordinates of the partitioned space and the position coordinates of each static obstacle; at the same time, the navigation terminal can determine the ocean current velocity vector at each partitioned space by spatial interpolation or direct table lookup based on the environmental ocean current data.
[0080] Optionally, in S15, the navigation terminal can calculate the angle or direction cosine between the relative azimuth vector and the ocean current velocity vector for each combination of partitioned space and each static obstacle to obtain a direction consistency metric value; input the direction consistency metric value to a piecewise mapping function, which presets risk coefficient values corresponding to different metric value intervals; by judging the interval in which the direction consistency metric value falls, the flow field risk coefficient of each partitioned space for each static obstacle is mapped; and the flow field risk coefficients of the same partitioned space for different static obstacles are combined to obtain the final flow field risk coefficient of each partitioned space.
[0081] Optionally, in S16, the navigation terminal can organize the flow field risk coefficients of each partitioned space according to the spatial location of each partitioned space to form a scalar potential field covering the entire spatial range represented by the prior environmental data, which is the flow field risk potential field.
[0082] Optionally, in S17, the navigation terminal can, based on the static obstacle position data, take each static obstacle position as the center and a preset risk action radius as the range of action, calculate the repulsive potential field value for each partitioned space falling within the range of action according to its distance from the static obstacle, and set the repulsive potential field value to zero for partitioned spaces outside the range of action, thus constructing the static obstacle repulsive potential field.
[0083] Optionally, in S18, the navigation terminal can obtain the potential field value of each partitioned space in the static obstacle repulsion potential field and the potential field value in the flow field risk potential field. Then, according to a preset weighted fusion strategy, the potential field value of the static obstacle repulsion potential field and the potential field value in the flow field risk potential field are multiplied by their respective weight coefficients and summed to obtain the comprehensive environmental cost potential field value of the partitioned space, forming an environmental cost potential field covering the entire space range.
[0084] Optionally, the environmental cost potential field is obtained by weighting and fusing the potential field values of the static obstacle repulsive potential field and the flow field risk potential field using a preset weighted fusion strategy, specifically including:
[0085] Potential field value of the i-th partitioned space in the repulsive potential field of a static obstacle The potential value of the i-th partitioned space in the flow field risk potential field Normalization was performed on each of the three methods to obtain the normalized potential field values. and :
[0086] ;
[0087] ;
[0088] Where j represents the number of partitions traversed. and These represent taking the minimum and maximum values of the potential field across all partitioned spaces, respectively.
[0089] The comprehensive potential field value U_{total}^{(i)} of the i-th partitioned space in the environmental cost potential field is calculated by the following formula:
[0090] ;
[0091] in, These are the weighting coefficients for the repulsive potential field of static obstacles. Let be the weighting coefficients of the flow field risk potential field, and satisfy . , , ; and Based on the mission priority parameters in the navigation mission data: when the safety priority in the navigation mission data is higher than the energy consumption priority, increase... The value of is increased when the energy consumption priority in the navigation mission data is higher than the safety priority. The value of .
[0092] In one embodiment, the joint state vector is input into a pre-trained SAC reinforcement learning network, which outputs desired motion correction data, including:
[0093] S21, Based on local environmental perception data, extract the velocity component of the ocean current velocity vector in the current heading direction of the AUV, and the velocity gradient rate of the ocean current velocity vector in the direction perpendicular to the current heading direction of the AUV.
[0094] S22, the joint state vector is input into the Actor module of the SAC reinforcement learning network, and the initial action distribution parameters are output through hidden layer mapping. The initial action distribution parameters include the action mean and the action standard deviation.
[0095] S23, by using the gradient-normalized weighted exploration amplitude modulation rule, the motion standard deviation is anisotropically scaled according to the rate of change of the flow velocity gradient to obtain the exploration noise amplitude;
[0096] S24. Based on the flow velocity component, the mean motion value, and the exploration noise amplitude, the motion parameters to be corrected are obtained through reparameterized sampling.
[0097] S25. Based on the time series of path tracking deviation data in the joint state vector, the deviation change acceleration is calculated. Based on the deviation change acceleration and the current acceleration in the AUV real-time state data, an inertial delay compensation term is constructed. The inertial delay compensation term is then superimposed on the motion parameters to be corrected to obtain the desired motion correction data.
[0098] Specifically, the velocity gradient change rate refers to the spatial gradient of the ocean current velocity vector in the direction perpendicular to the current bow of the AUV at the AUV's current position, used to characterize the strength of the crossflow shearing effect. Initial action distribution parameters refer to the statistics describing the action probability distribution obtained by the Actor module of the SAC reinforcement learning network based on the joint state vector mapping, including the action mean and action standard deviation. The exploration amplitude modulation rule is a calculation rule that dynamically adjusts the exploration noise amplitude according to environmental characteristics, changing the amplitude of the exploration noise in each dimension by anisotropically scaling the action standard deviation. Anisotropic scaling refers to applying different scaling coefficients to different dimensional components of the action standard deviation vector, causing the exploration noise to have different amplitudes in different action dimensions.
[0099] Reparameterization sampling refers to a sampling method that samples independent noise from a standard normal distribution and then generates motion samples through a linear transformation of the motion mean and scaled standard deviation. The motion parameters to be corrected refer to the initial motion corrections obtained through reparameterization sampling, before inertial delay compensation. Deviation change acceleration refers to the second derivative of the path tracking deviation data over time, used to measure the acceleration or deceleration trend of deviation changes. The inertial delay compensation term is a correction vector constructed based on the deviation change acceleration and the current acceleration of the AUV, used to compensate for the delay in the AUV's dynamic inertial response.
[0100] Optionally, in S21, the navigation terminal can extract the ocean current velocity vector from the local environmental perception data and project the ocean current velocity vector onto the current heading direction and its vertical direction of the AUV respectively: the projection component in the heading direction is the velocity component, and the rate of change of the projection component in the vertical direction is obtained along space, that is, the spatial derivative of the vertical component in the direction perpendicular to the heading direction is calculated to obtain the velocity gradient rate of change.
[0101] Optionally, in S22, the navigation terminal can input the joint state vector into the Actor module in the pre-trained SAC reinforcement learning network. The Actor module consists of multiple fully connected hidden layers. The joint state vector is sequentially transformed by linear transformations and mapped by nonlinear activation functions in the multiple fully connected hidden layers, and then mapped into the action mean vector and action standard deviation vector in the output layer to obtain the initial action distribution parameters.
[0102] Optionally, in S23, the navigation terminal can anisotropically scale the motion standard deviation based on the rate of change of the flow velocity gradient using an exploration amplitude modulation rule based on gradient normalization weighting. Specifically, the navigation terminal can normalize the rate of change of the flow velocity gradient to obtain a normalized gradient representation value, determine the scaling factor for each motion dimension based on the gradient representation value, apply a larger scaling factor to motion dimensions sensitive to crossflow, and apply a smaller scaling factor to other dimensions, and multiply the scaling factor by the corresponding elements of the motion standard deviation to obtain the exploration noise amplitude.
[0103] Optionally, in S24, the navigation terminal can sample independent and identically distributed noise vectors from the standard normal distribution, and combine the action mean with the exploration noise amplitude according to the reparameterization sampling method: specifically, the navigation terminal can multiply the exploration noise amplitude with the standard normal noise vector element by element and then add it to the action mean element by element to obtain the action parameters to be corrected.
[0104] Optionally, in S25, the navigation terminal can extract the time series of path tracking deviation data from the joint state vector, perform second-order numerical difference, calculate the deviation change acceleration, and simultaneously obtain the current acceleration from the AUV real-time state data. The deviation change acceleration and the current acceleration are multiplied by their respective preset compensation gain coefficients and then summed to construct an inertial delay compensation term. The navigation terminal can then add the inertial delay compensation term element by element to the motion parameters to be corrected to obtain the desired motion correction data.
[0105] In one embodiment, the AUV dynamic model parameters include a mass matrix, a Coriolis centripetal force matrix, a damping matrix, a restoring force vector, hydrodynamic coefficients, and a thruster saturation limit. The hydrodynamic coefficients include a hydrodynamic added mass coefficient and a hydrodynamic damping coefficient. The desired motion correction data includes velocity increments and bow angle increments. The AUV real-time status data includes the current velocity vector, current attitude angle, current position, and current acceleration.
[0106] like Figure 2 As shown, based on the AUV's real-time state data, AUV dynamic model parameters, and desired motion correction data, a target state vector is generated through dynamic deduction, including:
[0107] S31, based on the velocity increment, bow angle increment, current velocity vector and current attitude angle, the generalized force vector is calculated;
[0108] S32, based on the mass matrix, Coriolis centripetal force matrix, damping matrix, restoring force vector and generalized force vector, construct the state equation of the AUV dynamic model to obtain the first state equation;
[0109] S33, based on the hydrodynamic added mass coefficient, update the corresponding matrix elements in the mass matrix of the first state equation, and based on the hydrodynamic damping coefficient, update the corresponding matrix elements in the damping matrix of the first state equation to obtain the second state equation.
[0110] S34. Based on the ocean current velocity vector in the local environmental perception data and the current position in the AUV real-time status data, an ocean current disturbance term is added to the second state equation to obtain the third state equation.
[0111] S35. The third state equation is numerically integrated using the fourth-order Runge-Kutta method to obtain the forward predicted state sequence, which includes state vectors at multiple time steps. The state vector of the last time step is extracted from the forward predicted state sequence as the target state vector.
[0112] Specifically, the AUV dynamic model parameters include the mass matrix, Coriolis centripetal force matrix, damping matrix, restoring force vector, hydrodynamic coefficients, and thruster saturation limit. The mass matrix is a positive definite matrix describing the AUV's inertial mass and its distribution characteristics, comprising rigid body inertial mass and additional mass. The Coriolis centripetal force matrix characterizes the inertial coupling effect caused by the rotating coordinate system during AUV motion. The damping matrix describes the relationship between the fluid damping force and the velocity of the AUV during underwater motion, mainly including linear and nonlinear damping terms. The restoring force vector is the restoring torque vector generated by the difference between gravity and buoyancy, which depends on the AUV's attitude angle and hydrodynamic center of gravity position. The hydrodynamic coefficients include the hydrodynamic additional mass coefficient and the hydrodynamic damping coefficient, used to quantitatively characterize the additional inertial and damping effects exerted by the surrounding fluid on the AUV during underwater motion. The thruster saturation limit is the physical limit value of the maximum thrust amplitude and maximum thrust change rate that each thruster can output. The expected motion correction data includes velocity increment and heading angle increment, representing the correction amounts of the SAC reinforcement learning network output to the AUV's linear velocity vector and heading angle, respectively. The AUV's real-time state data includes the current velocity vector, current attitude angle, current position, and current acceleration, which can be obtained in real time from sensors such as the inertial navigation system, Doppler velocimeter, and depth gauge.
[0113] Optionally, in S31, the navigation terminal can superimpose the velocity increment and heading angle increment in the desired motion correction data onto the current velocity vector and current attitude angle in the AUV's real-time status data to obtain the corrected desired velocity vector and desired attitude angle. The navigation terminal can convert the deviation between the desired velocity vector and the current velocity vector, and the deviation between the desired attitude angle and the current attitude angle, into a generalized force vector in the body coordinate system using inverse dynamics. This generalized force vector contains longitudinal force, lateral force, and heading moment components, which are used to characterize the net external force and net moment required by the AUV to achieve the desired motion correction amount for each degree of freedom.
[0114] Optionally, in S32, the navigation terminal can construct the dynamic state equation of the AUV in the body coordinate system based on the mass matrix, Coriolis centripetal force matrix, damping matrix, restoring force vector and generalized force vector in the AUV dynamic model parameters, and obtain the first state equation. The first state equation uses the current velocity vector and the current attitude angle as state variables to describe the response relationship of the AUV body coordinate system acceleration under the action of generalized force.
[0115] Optionally, in S33, the navigation terminal can update the matrix elements related to the added mass effect in the mass matrix of the first state equation according to the hydrodynamic added mass coefficient, that is, replace the corresponding inertial mass elements with the sum of rigid body mass and added mass; at the same time, according to the hydrodynamic damping coefficient, update the coefficient terms related to linear damping and nonlinear damping in the damping matrix of the first state equation to obtain the second state equation, so that the dynamic model can more accurately reflect the added inertial effect and damping effect of underwater fluid on AUV.
[0116] Optionally, in S34, the navigation terminal can extract the ocean current velocity vector at the current location from the local environmental perception data. Based on the principle of relative velocity, an ocean current disturbance term is added to the volume coordinate acceleration expression of the second state equation. That is, the absolute velocity vector is replaced by the velocity vector of the AUV relative to the ocean current to calculate the damping force and Coriolis force, so that the damping matrix and Coriolis centripetal force matrix in the state equation are calculated based on the relative velocity vector, thus obtaining a third state equation that can take into account time-varying ocean current disturbances.
[0117] Optionally, in S35, the navigation terminal can use the current velocity vector and current attitude angle in the real-time state data of the AUV at the current moment as initial conditions, and the preset navigation sampling time as the integration step size, and use the fourth-order Runge-Kutta method to numerically integrate and recursively derive the third state equation. In the prediction time domain, the state vector of each time step is calculated sequentially to form a forward prediction state sequence containing state vectors of multiple time steps. The state vector of the last time step is extracted from the forward prediction state sequence as the target state vector. The target state vector integrates the combined effects of dynamic response, hydrodynamic correction and ocean current disturbance, providing an accurate expected reference state for subsequent model predictive control.
[0118] Optionally, the third state equation is numerically integrated using the fourth-order Runge-Kutta method to obtain the forward predicted state sequence, including:
[0119] Let the state vector of the third state equation at time t be... Where v(t) is the current velocity vector, For the current attitude angle, the third state equation is uniformly expressed as: Where u(t) is the control input at the current moment, and w(t) is the ocean current disturbance term;
[0120] Set the integration step size ,in The navigation sampling time is denoted by m, which is a preset number of substeps and takes the value of a positive integer. For the current moment , with the current state vector As the initial value, in the prediction time domain The state vector for each substep is calculated iteratively using the following recursive formula:
[0121] ;
[0122] ;
[0123] ;
[0124] ;
[0125] ;
[0126] in, For the time of the j-th sub-step, Let j be the state vector of the j-th sub-step. These are the four slope vectors of the fourth-order Runge-Kutta method;
[0127] After completing m sub-steps of recursion, record the current state vector as the predicted state for one time step, and repeat the above process until the prediction time domain is reached. The endpoint is reached, and the forward predicted state sequence is obtained. .
[0128] In one embodiment, the control input is obtained by solving a model predictive control algorithm based on the AUV's real-time state data and the target state vector, including:
[0129] S41, based on the target state vector and the real-time state data of the AUV, the state deviation vector is calculated;
[0130] S42. Based on the navigation mission data, the navigation sampling time is extracted, and based on the mass matrix, Coriolis centripetal force matrix and damping matrix in the AUV dynamic model parameters, as well as the navigation sampling time, the third state equation is discretized to obtain the first discrete state space model.
[0131] S43, based on the current acceleration in the AUV real-time status data and the ocean current velocity vector in the local environment perception data, the unmodeled dynamic compensation amount and the external ocean current disturbance compensation amount are estimated by the extended state observer.
[0132] S44. Based on the state deviation vector and the velocity increment and heading angle increment in the expected motion correction data, the terminal constraint conditions of the first discrete state space model are constructed. Based on the terminal constraint conditions, the unmodeled dynamic compensation amount and the external ocean current disturbance compensation amount, the first discrete state space model is fed forward to obtain the second discrete state space model.
[0133] S45, based on the current velocity vector and current position in the AUV real-time status data, and the static obstacle position data in the local environment perception data, calculate the relative obstacle distance and relative approach speed;
[0134] S46, using a preset time-domain mapping table, the prediction time domain is set according to the relative obstacle distance, and the control time domain is set according to the relative approach speed; the state deviation vector is input into the second discrete state-space model, and the rolling optimization prediction sequence is derived by combining the prediction time domain and the control time domain;
[0135] S47. Based on the rolling optimization prediction sequence and the state deviation vector, a quadratic term of state deviation is constructed; based on the velocity increment and heading angle increment in the expected motion correction data, a quadratic term of expected motion deviation is constructed; the quadratic cost function is obtained by weighted summation of the quadratic term of state deviation and the quadratic term of expected motion deviation.
[0136] S48, based on the saturation limit of the AUV propeller, sets control quantity amplitude constraints and control quantity change rate constraints; in the control time domain, based on the control quantity change rate constraints and control quantity amplitude constraints, solves the quadratic cost function to obtain the control input quantity.
[0137] For example, the state deviation vector is the difference vector obtained by subtracting the target state vector from the corresponding state variables in the AUV's real-time state data. The extended state observer is a state estimator that does not rely on an exact mathematical model. It estimates the unmodeled dynamics of the system and external disturbances as extended states in real time, outputting compensation for unmodeled dynamics and external ocean current disturbances. The terminal constraint is a state-feasibility domain constraint applied at the end of the prediction time domain of the first discrete state-space model. Feedforward compensation is the superposition of the terminal constraint, the compensation for unmodeled dynamics, and the compensation for external ocean current disturbances into the prediction equation of the discrete state-space model.
[0138] The relative obstacle distance is the Euclidean distance between the current position of the AUV and the position of the static obstacle; the relative approach speed is the projection component of the current velocity vector in the direction from the current position of the AUV to the position of the static obstacle. The time-domain mapping table is a preset nonlinear mapping relationship that maps the relative obstacle distance and relative approach speed to the prediction time-domain length and the control time-domain length. The rolling optimization prediction sequence is a set of future state prediction values obtained by recursively deriving the second discrete state-space model with the current state deviation vector as the initial condition within each control cycle.
[0139] The quadratic term of the state deviation is a cost term consisting of a weighted quadratic form of the state deviation vectors at each time step in the rolling optimization prediction sequence; the quadratic term of the desired motion deviation is a cost term consisting of a weighted quadratic form of the deviations between the velocity increment and heading angle increment in the desired motion correction data and the corresponding control variables. The quadratic cost function is a scalar function obtained by weighted summation of the quadratic terms of the state deviation and the desired motion deviation. The control variable amplitude constraint is a limitation on the upper and lower limits of the output thrust amplitude of each thruster based on the thruster saturation limit; the control variable rate of change constraint is a limitation on the upper limit of the control variable change amplitude within adjacent control cycles.
[0140] Optionally, in S41, the navigation terminal can subtract each state component in the target state vector from the corresponding current state component in the AUV real-time state data to obtain the state deviation vector.
[0141] Optionally, in S42, the navigation terminal can extract the navigation sampling time from the navigation mission data, and using the navigation sampling time as the step size, based on the mass matrix, Coriolis centripetal force matrix and damping matrix, use the forward Euler method to discretize the third state equation into a difference equation to obtain the first discrete state space model.
[0142] Optionally, in S43, the navigation terminal can take the current acceleration and ocean current velocity vector as input, estimate the total system disturbance online through the recursive equation of the extended state observer, and separate the unmodeled dynamic compensation and the external ocean current disturbance compensation.
[0143] Optionally, the unmodeled dynamic compensation and external ocean current disturbance compensation are estimated using an extended state observer, including:
[0144] The acceleration response equation of the AUV dynamic model in volume coordinates is rewritten in an extended state-space form that includes the total disturbance:
[0145] ;
[0146] in, The state variable corresponding to the velocity vector in the AUV's real-time status data. These are extended state variables used to characterize the total disturbance of the system. The preset control gain coefficient is the reciprocal of the main diagonal element of the quality matrix, and u is the control input at the current moment. Let y be the rate of change of the total disturbance, and y be the system output, where the total disturbance is... It includes the sum of the unmodeled dynamic compensation amount and the external ocean current disturbance compensation amount;
[0147] Construct the following linearly extended state observer:
[0148] ;
[0149] in, Let x_1 be an estimated value. Total disturbance The estimated value, and The observer gain coefficient; the observer gain coefficient and Based on observer bandwidth Sure:
[0150] ;
[0151] Among them, the observer bandwidth The value of is 3-5 times the minimum eigenvalue of the damping matrix in the AUV dynamic model, which is used to ensure that the convergence speed of the observer is greater than the dynamic response speed of the system.
[0152] The estimated value of the total disturbance Decomposed into unmodeled dynamic compensation quantities Compensation for external ocean current disturbances The sum, of which the external ocean current disturbance compensation is calculated using the forward Eulerian method based on the ocean current velocity vector in the local environmental perception data, and the unmodeled dynamic compensation is calculated by... The result is obtained by subtracting the compensation amount for external ocean current interference.
[0153] Optionally, in S44, the navigation terminal can construct a mathematical expression for the terminal constraints based on the state deviation vector, velocity increment, and heading angle increment; and superimpose the terminal constraints, unmodeled dynamic compensation, and external ocean current disturbance compensation in feedforward form into the prediction update equation of the first discrete state space model to obtain the second discrete state space model.
[0154] Optionally, in S45, the navigation terminal can calculate the Euclidean distance between the current position and the positions of each static obstacle as the relative obstacle distance, and project the current velocity vector onto the unit vector pointing from the current position to the obstacle to obtain the relative approach velocity.
[0155] Optionally, in S46, the navigation terminal can input the relative obstacle distance and relative approach speed into the time domain mapping table to obtain the prediction time domain length and control time domain length; using the state deviation vector as the initial condition, the second discrete state space model is recursively derived in the prediction time domain to generate a rolling optimization prediction sequence.
[0156] Optionally, in S47, the navigation terminal can obtain a quadratic term of state deviation by summing the weighted quadratic form of the state deviation vectors at each moment in the rolling optimization prediction sequence, obtain a quadratic term of expected motion deviation by summing the weighted quadratic form of the deviations between the control quantity and the expected motion correction quantity at each moment, and obtain a quadratic cost function by summing the weighted values according to preset weights.
[0157] Optionally, the specific expression for the quadratic cost function J(k) is:
[0158] ;
[0159] Where k is the discrete-time index of the current control cycle; To predict the time domain, For the control time domain, and ; The state vector at time k+i predicted at time k is obtained recursively from the second discrete state-space model. The reference state vector at time k+i is determined based on the velocity increment and heading angle increment in the desired motion correction data and the real-time state data of the AUV. The control input quantity to be applied at time k+i predicted at time k; This refers to the change in control quantity between adjacent control cycles; Let be the state deviation weighting matrix, and be a positive definite symmetric matrix. The dimension of the state vector; To control the incremental weighting matrix, it is a positive definite symmetric matrix. To control the dimensions of the input, The control amplitude weighting matrix is a positive semi-definite symmetric matrix;
[0160] in, Q is a positive definite symmetric matrix obtained by solving the discrete algebraic Riccati equation, where Q is the terminal cost weighting matrix and Q is the state deviation weighting matrix. This represents the weighted quadratic norm of vector z under the weighted matrix W; The state deviation is the quadratic term. To control the incremental quadratic term, To control the quadratic term of the amplitude, For terminal cost items;
[0161] The quadratic term of the state deviation and the quadratic term of the desired motion deviation are associated through the weighting matrices Q and R: the velocity increment and heading angle increment in the desired motion correction data are... The setting indirectly affects the value of the quadratic term of the state deviation;
[0162] The weighting matrices Q, R, and S are set according to the performance preference parameters in the navigation mission data: when the tracking accuracy priority is higher than the energy consumption priority in the navigation mission data, the matrix element values of Q are increased; otherwise, the matrix element values of R and S are increased.
[0163] Optionally, in S48, the navigation terminal can set the upper and lower limits of the control quantity amplitude and the upper limit of the rate of change of adjacent cycles based on the thruster saturation limit. In the control time domain, the optimal control sequence is solved by using a quadratic cost function as the optimization objective and two types of constraints as constraints. The first control quantity is extracted as the control input quantity output for the current cycle.
[0164] In one embodiment, the first state equation is calculated as follows:
[0165] ;
[0166] in, Let M be the acceleration vector in body coordinates, and M be the inertial mass matrix. It is a generalized force vector. Let v be the Coriolis centripetal force matrix, and v be the current velocity vector. Here is the damping matrix. This is the restoring force vector.
[0167] Specifically, τ is the acceleration vector in the body coordinate system, representing the linear and angular accelerations of the AUV in the three axes of the body coordinate system; M is the inertial mass matrix, a positive definite symmetric matrix that includes the rigid body inertial mass and hydrodynamic added mass, used to describe the mass and inertial distribution characteristics of the AUV; τ is the generalized force vector obtained by inverse dynamic transformation of the desired motion correction, containing three components: longitudinal force, lateral force, and bow torque. is the Coriolis centripetal force matrix, whose elements are linear functions of the current velocity vector v, representing the inertial coupling effect caused by angular velocity when the AUV moves in the rotating coordinate system; v is the current velocity vector, which includes the longitudinal velocity, lateral velocity and bow angular velocity of the AUV in the volume coordinate system. The damping matrix contains linear and nonlinear damping terms. Its matrix elements are related to the magnitude of the current velocity vector v and are used to describe the fluid damping force experienced by the AUV when it moves underwater. The restoring force vector is the restoring torque vector generated by the difference between gravity and buoyancy, and its value depends on the current attitude angle of the AUV. .
[0168] Specifically, the navigation terminal can obtain the inertial mass matrix M and the Coriolis centripetal force matrix from the AUV dynamic model parameters based on the generalized force vector τ. Damping matrix and restoring force vector And obtain the current velocity vector v; substitute it into In this process, we obtain the equation The acceleration vector in the body coordinate system is given by this formula, which fully describes the physical equilibrium relationship between the generalized force, Coriolis inertial force, damping force, restoring force and body coordinate acceleration. It is the basis for subsequent hydrodynamic coefficient correction, ocean current disturbance addition and numerical integration derivation.
[0169] In one embodiment, generating thruster thrust distribution data based on control input includes:
[0170] S51: Obtain the thruster efficiency data and thruster output limit data of the AUV, and calculate the initial thrust allocation vector based on the control input and thruster efficiency data using the pseudo-inverse method.
[0171] S52, based on the initial thrust distribution vector and thruster output limit data, identifies thrust components that exceed the AUV thruster saturation limit, and redistributes the thrust components through quadratic programming to obtain thruster thrust distribution data.
[0172] For example, thruster efficiency data refers to a set of quantitative parameters characterizing the energy conversion efficiency of each AUV thruster when converting command signals into actual thrust output, usually stored in the form of an efficiency matrix; thruster output limit data refers to a set of data describing the maximum thrust amplitude and upper limit of thrust change rate that each thruster can output under physical and safety constraints, including maximum forward thrust, maximum reverse thrust, and maximum thrust change rate per unit time; the pseudo-inverse method is a mathematical method that obtains the minimum norm least squares solution of a system of linear equations by solving the Moore-Penrose generalized inverse of a matrix; the initial thrust allocation vector refers to the initial solution vector of thrust commands for each thruster obtained directly by the pseudo-inverse method without saturation constraint treatment; quadratic programming is a mathematical programming method for solving the optimal solution of a quadratic objective function under linear constraints; thrust components refer to the thrust command values corresponding to a certain thruster in the initial thrust allocation vector.
[0173] Optionally, in S51, the navigation terminal can read pre-calibrated thruster efficiency data and thruster output limit data from the configuration parameter database. The thruster efficiency data is represented as an efficiency matrix B, where the number of rows represents the number of control degrees of freedom, the number of columns represents the total number of thrusters, and each element represents the product of the contribution coefficient and efficiency coefficient of the corresponding thruster unit command to a specific degree of freedom. The thruster output limit data includes the upper limit vector of thrust amplitude saturation for each thruster. and lower bound vector and rate of change limit vector The navigation terminal uses the current control input, i.e., the desired generalized force vector τ, output by the model predictive control algorithm as the total thrust demand to be allocated. Based on the control input τ and the efficiency matrix B, it calculates the initial thrust allocation vector using the pseudo-inverse method. The specific calculation formula is as follows: The pseudo-inverse solution satisfies the control input quantity. Under the premise of equality constraints, the sum of squares of each thrust command is minimized, thereby obtaining the initial thrust allocation vector.
[0174] Optionally, the thrust components are redistributed through quadratic programming, specifically including:
[0175] The truncated initial thrust distribution vector Based on the reference, the thrust command of each thruster Let be the decision variables, where Given the total number of thrusters, construct the following quadratic programming problem:
[0176]
[0177]
[0178]
[0179]
[0180] in, For Hessian matrix, The thruster efficiency matrix is... The preset regularization coefficients are used. The vector of coefficients for linear terms. The actual thrust allocation vector for the previous control cycle. and These are constraints on the minimum and maximum thrust changes per step for each thruster. This refers to the remaining thrust demand vector that still needs to be reallocated after truncation; the constraints in the... and These are the lower and upper limits of the thruster saturation limit. The thruster response speed parameters are preset based on the navigation mission data;
[0181] When the quadratic programming problem has a feasible solution, the optimal solution obtained will be... As the thrust distribution data of the aforementioned thruster;
[0182] When the quadratic programming problem has no feasible solution, the relaxation factor method is used to relax the equality constraints. Replace with inequality constraints ,in Using a preset tolerance threshold, the quadratic programming problem is solved again to obtain a suboptimal feasible solution as the thruster thrust allocation data.
[0183] Optionally, in S52, the navigation terminal can... Compared with the upper limit of amplitude saturation in the thruster output limit data and lower limit By comparing elements one by one, thrust components exceeding the corresponding limits are identified. These over-limit components are truncated to their saturation limits, and the remaining thrust demand resulting from the truncation is taken as the deviation to be redistributed. Using the thrust commands of all thrusters as decision variables, the optimization objective is to minimize the quadratic deviation between the truncated initial allocation vector and the decision variables. A quadratic programming problem is constructed with the amplitude saturation constraints and rate of change constraints of each thruster as constraints. The effective set method or interior point method is used to solve the problem. Under the premise of satisfying all output limits, the over-limit thrust components are optimally transferred to other unsaturated thrusters, so that the redistributed thrust commands are all within the limits and the error between the combined thrust and the control input is minimized. The optimized solution is output as thruster thrust allocation data to the underlying drive system to drive each thruster to execute the corresponding thrust output.
[0184] The aforementioned AI-based AUV underwater navigation method acquires prior environmental data, navigation mission data, AUV real-time status data, local environmental perception data, and dynamic model parameters, providing a multi-source information foundation for subsequent global planning and local control. By constructing an environmental cost potential field through an artificial potential field algorithm and fusing it with heuristic search to generate a global reference path set, it solves the problem of traditional artificial potential field methods easily getting trapped in local minima, while providing the AUV with a globally optimal navigation reference path. Then, based on the real-time status and the global reference path, a line-of-sight guidance algorithm is used to calculate the path tracking deviation, and an attention mechanism is used to fuse local environmental perception and deviation data into a joint state vector. This enhances the ability to perceive and represent dynamic environmental information differently, significantly improving path tracking accuracy and disturbance resistance. The joint state vector is then input into a pre-trained SAC reinforcement learning network. The system utilizes the maximum entropy exploration characteristic of the network to output desired motion correction data, enabling the AUV to have adaptive online decision-making capabilities in complex time-varying environments. This effectively compensates for the shortcomings of traditional fixed-parameter controllers in dealing with model uncertainties and ocean current disturbances. Subsequently, based on the dynamic model parameters and desired corrections, dynamic deduction is performed to generate the target state vector. Combined with model predictive control, the control input that satisfies the thruster saturation constraints is solved, which can achieve optimal control while ensuring the dynamic feasibility of the system, overcoming the performance degradation of traditional model predictive control under strong disturbances. Finally, thruster thrust allocation data is generated based on the control input to drive the AUV's navigation. This achieves end-to-end collaborative optimization across the entire link, from global path planning and local motion correction to low-level control execution, thereby improving the overall navigation accuracy, autonomy, and environmental adaptability of the AUV in complex underwater environments.
[0185] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0186] Based on the same inventive concept, this application also provides an AI-based AUV underwater navigation system for implementing the AI-based AUV underwater navigation method described above. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations of one or more AI-based AUV underwater navigation system embodiments provided below can be found in the limitations of the AI-based AUV underwater navigation method described above, and will not be repeated here.
[0187] In one exemplary embodiment, such as Figure 3 As shown, an artificial intelligence-based AUV underwater navigation system is provided, including:
[0188] The data acquisition module 101 can be used to acquire prior environmental data, navigation mission data, AUV real-time status data, local environmental perception data, and AUV dynamic model parameters.
[0189] The path planning module 102 can be used to construct an environmental cost potential field based on prior environmental data and an artificial potential field algorithm, and generate a global reference path set based on navigation mission data and the environmental cost potential field through a heuristic search algorithm.
[0190] The guidance fusion module 103 can be used to calculate path tracking deviation data based on AUV real-time status data and global reference path set through line-of-sight guidance algorithm, and fuse local environment perception data and path tracking deviation data through attention mechanism to obtain joint state vector.
[0191] The learning correction module 104 can be used to input the joint state vector into a pre-trained SAC reinforcement learning network and output the desired motion correction data.
[0192] The predictive control module 105 can generate a target state vector based on the AUV's real-time state data, AUV dynamic model parameters, and desired motion correction data through dynamic deduction, and solve for the control input quantity based on the AUV's real-time state data and target state vector through a model predictive control algorithm.
[0193] The thrust distribution module 106 can be used to generate thruster distribution data based on control inputs, which is used to drive the AUV.
[0194] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the AI-based AUV underwater navigation method as described above.
[0195] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.
[0196] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0197] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.
Claims
1. An AUV underwater navigation method based on artificial intelligence, characterized in that, The method includes: Acquire prior environmental data, navigation mission data, AUV real-time status data, local environmental perception data, and AUV dynamic model parameters; Based on the prior environmental data, an environmental cost potential field is constructed using an artificial potential field algorithm. Then, based on the navigation mission data and the environmental cost potential field, a global reference path set is generated using a heuristic search algorithm. Based on the AUV real-time status data and the global reference path set, path tracking deviation data is calculated using a line-of-sight guidance algorithm. The local environment perception data and the path tracking deviation data are then fused using an attention mechanism to obtain a joint state vector. The joint state vector is input into a pre-trained SAC reinforcement learning network, which outputs the desired motion correction data. Based on the AUV real-time status data, the AUV dynamic model parameters, and the desired motion correction data, a target state vector is generated through dynamic deduction. Then, based on the AUV real-time status data and the target state vector, the control input is obtained through a model predictive control algorithm. Based on the control input, thruster thrust distribution data is generated, which is used to drive the AUV.
2. The method according to claim 1, characterized in that, The step of constructing an environmental cost potential field based on the prior environmental data using an artificial potential field algorithm includes: Based on the aforementioned prior environmental data, static obstacle location data and environmental ocean current data are extracted. Based on the environmental prior data, the spatial variation gradient of ocean current velocity within the spatial range represented by the environmental prior data is calculated, and the fineness of spatial division is set according to the magnitude of the spatial variation gradient. Based on the fine-grained spatial division, the spatial range represented by the prior environmental data is discretized to obtain multiple partitioned spaces; Based on the static obstacle location data and the environmental ocean current data, the relative orientation vector of each of the partitioned spaces with respect to each of the static obstacles is calculated, and the ocean current velocity vector of the environmental ocean current data in each of the partitioned spaces is calculated. Based on the relative azimuth vector and the ocean current velocity vector, the directional consistency metric is calculated, and based on the directional consistency metric, the flow field risk coefficient of each of the partitioned spaces is obtained by mapping through a piecewise mapping function. Based on the aforementioned flow field risk coefficients, the flow field risk potential field is constructed. Based on the static obstacle location data, combined with the spatial division and the preset risk radius, a static obstacle repulsive potential field is constructed. By using a preset weighted fusion strategy, the potential field values of the static obstacle repulsive potential field and the flow field risk potential field are weighted and fused to obtain the environmental cost potential field.
3. The method according to claim 2, characterized in that, The step of inputting the joint state vector into a pre-trained SAC reinforcement learning network and outputting desired motion correction data includes: Based on the local environmental perception data, the velocity component of the ocean current velocity vector in the current heading direction of the AUV and the velocity gradient rate of the ocean current velocity vector in the direction perpendicular to the current heading direction of the AUV are extracted. The joint state vector is input into the Actor module of the SAC reinforcement learning network, and the initial action distribution parameters are output through hidden layer mapping. The initial action distribution parameters include the action mean and the action standard deviation. By using an exploration amplitude modulation rule based on gradient normalization weighting, the action standard deviation is anisotropically scaled according to the rate of change of the flow velocity gradient to obtain the exploration noise amplitude. Based on the flow velocity component, the mean action value, and the exploration noise amplitude, the action parameters to be corrected are obtained through reparameterized sampling. Based on the time series of the path tracking deviation data in the joint state vector, the deviation change acceleration is calculated, and based on the deviation change acceleration and the current acceleration in the AUV real-time state data, an inertial delay compensation term is constructed. The inertial delay compensation term is then superimposed on the motion parameters to be corrected to obtain the desired motion correction data.
4. The method according to claim 3, characterized in that, The AUV dynamic model parameters include the mass matrix, Coriolis centripetal force matrix, damping matrix, restoring force vector, hydrodynamic coefficients, and thruster saturation limit. The hydrodynamic coefficients include the hydrodynamic added mass coefficient and the hydrodynamic damping coefficient. The desired motion correction data includes velocity increment and heading angle increment. The AUV real-time status data includes the current velocity vector, current attitude angle, current position, and current acceleration. The step of generating a target state vector through dynamic deduction based on the AUV's real-time state data, the AUV's dynamic model parameters, and the desired motion correction data includes: Based on the velocity increment, the heading angle increment, the current velocity vector, and the current attitude angle, the generalized force vector is calculated; Based on the mass matrix, the Coriolis centripetal force matrix, the damping matrix, the restoring force vector, and the generalized force vector, the state equation of the AUV dynamic model is constructed, and the first state equation is obtained. Based on the hydrodynamic added mass coefficient, update the corresponding matrix elements in the mass matrix of the first state equation, and based on the hydrodynamic damping coefficient, update the corresponding matrix elements in the damping matrix of the first state equation to obtain the second state equation. Based on the ocean current velocity vector in the local environmental perception data and the current position in the AUV real-time status data, an ocean current interference term is added to the second state equation to obtain the third state equation; The third state equation is numerically integrated using the fourth-order Runge-Kutta method to obtain a forward predicted state sequence, wherein the forward predicted state sequence includes state vectors at multiple time steps; the state vector of the last time step is extracted from the forward predicted state sequence as the target state vector.
5. The method according to claim 4, characterized in that, The step of obtaining the control input quantity based on the AUV's real-time state data and the target state vector using a model predictive control algorithm includes: Based on the target state vector and the real-time state data of the AUV, the state deviation vector is calculated. Based on the navigation mission data, the navigation sampling time is extracted, and based on the mass matrix, the Coriolis centripetal force matrix and the damping matrix in the AUV dynamic model parameters, as well as the navigation sampling time, the third state equation is discretized to obtain the first discrete state space model. Based on the current acceleration in the AUV real-time status data and the ocean current velocity vector in the local environment perception data, the unmodeled dynamic compensation amount and the external ocean current disturbance compensation amount are estimated by the extended state observer. Based on the state deviation vector and the velocity increment and heading angle increment in the expected motion correction data, the terminal constraint conditions of the first discrete state space model are constructed. Based on the terminal constraint conditions, the unmodeled dynamic compensation amount and the external ocean current interference compensation amount, the first discrete state space model is fed forward to obtain the second discrete state space model. Based on the current velocity vector and current position in the AUV real-time status data, and the static obstacle position data in the local environment perception data, the relative obstacle distance and relative approach speed are calculated. Using a preset time-domain mapping table, a prediction time domain is set based on the relative obstacle distance, and a control time domain is set based on the relative approach speed; the state deviation vector is input into the second discrete state-space model, and the rolling optimization prediction sequence is derived by combining the prediction time domain and the control time domain. Based on the rolling optimization prediction sequence and the state deviation vector, a quadratic term of state deviation is constructed; based on the velocity increment and the heading angle increment in the expected motion correction data, a quadratic term of expected motion deviation is constructed; the quadratic cost function is obtained by weighted summation of the quadratic term of state deviation and the quadratic term of expected motion deviation. Based on the saturation limit of the AUV thruster, control magnitude constraints and control rate of change constraints are set; within the control time domain, based on the control rate of change constraints and the control magnitude constraints, the quadratic cost function is solved to obtain the control input.
6. The method according to claim 4, characterized in that, The formula for calculating the first state equation is: ; in, Let M be the acceleration vector in body coordinates, and M be the inertial mass matrix. Let be the generalized force vector. Let v be the Coriolis centripetal force matrix, and v be the current velocity vector. Let be the damping matrix. Let be the restoring force vector.
7. The method according to claim 1, characterized in that, The step of generating thruster thrust distribution data based on the control input includes: The thruster efficiency data and thruster output limit data of the AUV are obtained, and the initial thrust allocation vector is calculated by the pseudo-inverse method based on the control input and the thruster efficiency data. Based on the initial thrust distribution vector and the thruster output limit data, thrust components exceeding the AUV thruster saturation limit are identified, and the thrust components are redistributed through quadratic programming to obtain the thruster thrust distribution data.
8. An AUV underwater navigation system based on artificial intelligence, characterized in that, The system includes: The data acquisition module is used to acquire prior environmental data, navigation mission data, AUV real-time status data, local environmental perception data, and AUV dynamic model parameters. The path planning module is used to construct an environmental cost potential field based on the prior environmental data using an artificial potential field algorithm, and to generate a global reference path set based on the navigation mission data and the environmental cost potential field using a heuristic search algorithm. The guidance fusion module is used to calculate path tracking deviation data based on the real-time status data of the AUV and the global reference path set through a line-of-sight guidance algorithm, and to fuse the local environment perception data and the path tracking deviation data through an attention mechanism to obtain a joint state vector. The learning correction module is used to input the joint state vector into the pre-trained SAC reinforcement learning network and output the desired motion correction data. The predictive control module is used to generate a target state vector by dynamic deduction based on the real-time state data of the AUV, the dynamic model parameters of the AUV, and the expected motion correction data, and to solve for the control input quantity by model predictive control algorithm based on the real-time state data of the AUV and the target state vector. The thrust distribution module is used to generate thruster distribution data based on the control input, and the thruster distribution data is used to drive the AUV.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.
