Multi-AUV adaptive path planning method for complex underwater environment

Abstract

The invention discloses a multi-AUV adaptive path planning method for a complex underwater environment, and belongs to the technical field of cooperative control of autonomous underwater vehicles, and the method comprises the following steps: S1, carrying out the grid division of a sea area, and generating a local observation tensor; s2, constructing a Markov decision model according to the local observation tensor and the navigation state tensor; s3, inputting the local observation tensor and the navigation state tensor into a double-branch cross attention network to generate Q value estimation; s4, adjusting Q value estimation by using the dual Q network; s5, distributed collaborative learning is carried out; and S6, generating an initial path according to the adjusted Q value estimation, and correcting the initial path by using a distributed collaborative learning result to obtain a global optimal path. According to the method, the problem of insufficient path planning robustness caused by dynamic interference and static obstacle coupling in a complex underwater environment is solved, and an end-to-end optimization link from local observation to global decision is formed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of underwater autonomous vehicle cooperative control, and particularly relates to a multi-AUV adaptive path planning method for complex underwater environment. BACKGROUND

[0002] With the surge in demand for tasks such as ocean resource exploration, submarine cable inspection and ecological monitoring, multi-AUV cooperative operation technology has become the core direction in the field of deep sea engineering. Multi-AUV system, through distributed sensing and task allocation, has incomparable advantages in spatial coverage capability and task fault tolerance, and is widely used in complex scenes such as submarine hydrothermal sulfide exploration and transoceanic communication cable intelligent inspection. However, the high dynamic and uncertain nature of the underwater environment poses a serious challenge to multi-AUV cooperative path planning.

[0003] Firstly, the time-varying characteristics of the complex underwater environment significantly increase the difficulty of path planning. The underwater environment contains time-varying ocean current field, unstructured terrain obstacles and intermittent communication interruption, etc. Multiple disturbances, traditional path planning methods based on geometric rules are difficult to quantify dynamic interference factors, and are prone to path oscillation in the ocean current mutation scene; group intelligence algorithms (such as genetic algorithm and particle swarm optimization) can generate globally suboptimal paths, but their iterative convergence time grows exponentially with the number of AUVs, which cannot meet the real-time demand.

[0004] Secondly, there is an inherent contradiction between the global optimality and local autonomy of multi-AUV cooperation. Traditional deep reinforcement learning methods construct an end-to-end mapping between the environment and the decision, providing a new paradigm for dynamic system decision-making, but the incompleteness of local observation information easily leads to the accumulation of Q value estimation bias; the design of fixed reward function is difficult to adapt to the dynamic balance of multi-objective tasks such as task completion, energy optimization and path smoothness, and in high-density obstacle scenes, the strategy conflict between obstacle avoidance demand and task efficiency often occurs.

[0005] In addition, the particularity of underwater communication further restricts the efficiency of multi-AUV cooperation. The strong attenuation characteristics of seawater medium to wireless signals result in limited communication bandwidth and significant end-to-end delay fluctuations, and traditional distributed learning relies on high-frequency parameter exchange, which easily leads to loss of feature resolution, and the path planning accuracy is seriously reduced in dense obstacle scenes; at the same time, existing research generally ignores the influence of underwater working conditions on algorithm robustness, such as the lack of strategy degradation mechanism when communication is interrupted, which leads the system to easily fall into local optimum.

[0006] The core contradictions of the current research can be summarized as three points: first, the traditional DRL method is difficult to balance local autonomy and group collaboration, and the global optimal path planning ability is insufficient; second, the efficiency and privacy protection contradiction of the distributed learning framework in the communication limited scene is not effectively solved, and the dynamic adaptability of quantization and sparsification strategy is insufficient; third, the multi-scale spatiotemporal feature extraction technology has limited modeling ability for complex ocean current movement, which limits the accuracy of obstacle identification and path tracking. Although existing researches try to break through by using attention mechanism, differential privacy federated learning and other technologies, the environment coupling of dynamic reward function is insufficient, the quantitative theory support of distributed learning is weak, and the modeling ability of spatiotemporal features for long-term dependence relationship is limited, which still seriously hinders the engineering application of multi-AUV system in real complex scenes.

[0007] Therefore, how to design an adaptive path planning method considering dynamic environment perception, multi-objective optimization and limited communication bandwidth adaptation, to realize efficient obstacle avoidance and global optimal path collaborative planning of multi-AUV in complex underwater environment, has become a technical problem to be solved in the current field. SUMMARY

[0008] In order to solve the above problems, the present application provides a multi-AUV adaptive path planning method for complex underwater environment.

[0009] The technical scheme of the present application is: a multi-AUV adaptive path planning method for complex underwater environment comprises the following steps:

[0010] S1, grid division is performed on the sea area to generate a local observation tensor;

[0011] S2, a Markov decision model is constructed according to the local observation tensor and the navigation state tensor;

[0012] S3, based on the Markov decision model, the local observation tensor and the navigation state tensor are input into a double-branch cross-attention network to generate a Q value estimate;

[0013] S4, the Q value estimate is adjusted by using a double Q network;

[0014] S5, distributed collaborative learning is performed;

[0015] S6, an initial path is generated according to the adjusted Q value estimate, and the initial path is corrected by using the distributed collaborative learning result to obtain a global optimal path.

[0016] Further, S1 comprises the following sub-steps:

[0017] S11, the sea area is divided into a plurality of uniform grid cells;

[0018] S12, generating time-varying ocean current based on a plurality of grid cells by superimposing Gaussian vortex model;

[0019] S13, updating the obstacle position of each grid cell;

[0020] S14, determining the local observation range of each AUV;

[0021] S15, generating a local observation tensor according to the local ocean current velocity component of the time-varying ocean current, the updated obstacle position of each grid cell, the local observation range of each AUV and the dynamic target heat map.

[0022] Further, in S2, the Markov decision model The expression is:

[0023] ;

[0024] Wherein, is the state space, is the action space, is the reward function, is the state transition function, is the discount factor.

[0025] Further, the state space includes a local observation tensor and a navigation state tensor; the navigation state tensor includes the pose, velocity, normalized energy reserve and task flag of the AUV;

[0026] The action space is composed of the joint space of the heading angle adjustment amount and the speed adjustment amount;

[0027] The expression of the reward function is:

[0028] ;

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] Wherein, is the task completion reward, is the safety obstacle avoidance penalty, is the energy efficiency penalty, is the path smoothness reward, is the total number of dynamic target points in the scene, is the exponential, is the AUV Euclidean distance to the target, is a control reward decay rate, is a repeated coverage penalty coefficient, is the total number of AUVs, is an AUV Euclidean distance to the target, is a collision penalty coefficient between AUVs, is a penalty gradient coefficient, is a safety distance threshold, is an AUV position coordinates at time, is an AUV position coordinates at time, is an obstacle avoidance penalty term, is the total number of obstacles detected by the AUV, is a safety threshold, is the position coordinates of the obstacle, is a heading angle adjustment amount of the AUV is a speed adjustment amount of the AUV is a maximum speed threshold of the AUV, is a speed adjustment penalty weight, is a speed adjustment energy consumption penalty coefficient, is the maximum number of time steps for a single round, is the heading angle of the AUV at time, is the heading angle of the AUV at time, is a reward weight of task completion reward, is a reward weight of energy efficiency penalty, is a reward weight of path smoothness reward, is a logarithmic function; The state transition function includes the pose, speed, energy reserve and task flag of the AUV. Further, S3 includes the following sub-steps:

[0034] S31, input the local observation tensor into a double-branch cross-attention network, use a convolutional layer to extract local feature maps, use a residual module to enhance the feature representation of the local feature maps, and then pass through a global average pooling layer to obtain an environmental feature vector;

[0035] Further, S3 includes the following sub-steps:

[0036] S31, input the local observation tensor into a double-branch cross-attention network, use a convolutional layer to extract local feature maps, use a residual module to enhance the feature representation of the local feature maps, and then pass through a global average pooling layer to obtain an environmental feature vector; ​​

[0037] S32. Input the navigation state vector into the dual-branch cross-attention network, use the first-level fully connected layer to obtain the intermediate vector, and use the second-level fully connected layer to obtain the navigation feature vector.

[0038] S33. Map the environmental feature vector and navigation feature vector to query vector, key vector and value vector respectively, and calculate the fused features by scaling the dot product attention layer;

[0039] S34. Q-value estimation by mapping fused features to discrete action space using fully connected layers.

[0040] Furthermore, in S33, the expression for calculating the fused features by scaling the dot product attention layer is:

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] in, For query vector, For key vectors, For value vectors, For activation function, The dimension of the key vector. For attention mechanisms, For environmental feature vectors, For navigation feature vectors, As the first learnable parameter, As the second learnable parameter, As the third learnable parameter, This represents the matrix transpose operator.

[0046] Furthermore, in S4, the dual-Q network includes a behavioral network and a target network;

[0047] In S4, the action strategy is adjusted through priority experience playback and dynamic learning rate adjustment.

[0048] Furthermore, in the priority experience replay, the priority of each experience is... The expression is:

[0049] ;

[0050] in, For TD error, This is a smoothing term; its value is a positive number close to 0.

[0051] The expression for dynamic learning rate adjustment is:

[0052] ;

[0053] in, The current learning rate, The initial learning rate, To adjust the coefficient, Let the loss function be the variance of the fluctuation. The initial loss variance, For the maximum policy entropy, This represents the current policy entropy.

[0054] Furthermore, in S5, distributed collaborative learning includes multi-AUV experience collaborative sharing, environmental potential field fusion correction, and policy collaborative evolution.

[0055] The beneficial effects of this invention are:

[0056] (1) This invention constructs a full-link technology framework that integrates deep reinforcement learning and distributed collaborative integration. It achieves dynamic feature fusion of environmental perception and navigation status through a dual-branch heterogeneous network architecture. The environmental perception branch uses a residual convolutional network to extract spatial features such as ocean currents and obstacles, while the navigation status branch encodes temporal features such as pose and energy through a fully connected network. The two are then connected through a cross-attention layer to achieve adaptive correlation modeling of heterogeneous information. This framework breaks through the limitations of traditional single-modal feature learning, solves the problem of insufficient robustness of path planning caused by the coupling of dynamic interference and static obstacles in complex underwater environments, and forms an end-to-end optimization link from local observation to global decision-making.

[0057] (2) This invention proposes a dynamic perception and reward adaptation mechanism based on real-time environmental conditions. It constructs a multi-scale environmental representation through hierarchical state space modeling. The global layer uses a discrete grid to divide the ocean current field and obstacle distribution, while the local layer defines a spherical perception region to obtain real-time heatmaps of nearby AUV locations and dynamic targets. A time-varying reward weight is designed by combining obstacle density and target distance to achieve a dynamic balance between obstacle avoidance priority and task orientation. This mechanism solves the problem of policy rigidity in traditional fixed reward functions under high obstacle density or sudden ocean current scenarios, improving the algorithm's adaptability to unstructured environments.

[0058] (3) This invention designs a distributed collaborative optimization and policy evolution mechanism, which realizes group collaborative decision-making through multi-agent experience sharing and gradient aggregation; the local experience of each AUV is uploaded to the central server after priority sampling, and the fusion generates a global coverage heat map to guide sample selection; gradient diversity incentive terms are introduced to suppress policy homogenization, and the group learning efficiency is maintained under limited communication bandwidth by combining a hierarchical parameter synchronization protocol. This mechanism breaks through the communication delay bottleneck of centralized decision-making architecture, solves the problem of collaborative instability of multiple AUVs in bandwidth-limited scenarios, and achieves a dynamic balance between local autonomy and global collaboration.

[0059] (4) The optimization reinforcement learning network update and experience reuse mechanism of the present invention improves the convergence stability of the algorithm through dual Q-value estimation and priority experience replay; adopts dual deep Q network to decouple action selection and value evaluation, introduces Huber loss and gradient pruning to suppress outlier interference; designs a hybrid priority sampling strategy based on TD error and environment coverage to enhance the learning focus on key decision nodes; this mechanism solves the problems of overestimation bias and low sample utilization efficiency of traditional Q learning, and improves the direction consistency and convergence speed of policy gradient update in complex dynamic environments. Attached Figure Description

[0060] Figure 1 A flowchart for an adaptive path planning method for multiple AUVs in complex underwater environments;

[0061] Figure 2 This is a schematic diagram of the structure of a two-branch cross-attention network. Detailed Implementation

[0062] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0063] like Figure 1 As shown, this invention provides a multi-AUV adaptive path planning method for complex underwater environments, comprising the following steps:

[0064] S1. Divide the sea area into grids and generate local observation tensors;

[0065] S2. Construct a Markov decision model based on the local observation tensor and the navigation state tensor;

[0066] S3. Based on the Markov decision model, the local observation tensor and the navigation state tensor are input into a two-branch cross-attention network to generate Q-value estimates;

[0067] S4. Adjust the Q-value estimate using a dual Q-network;

[0068] S5. Perform distributed collaborative learning;

[0069] S6. Based on the adjusted Q-value estimation, generate an initial path and use the results of distributed collaborative learning to correct the initial path to obtain the globally optimal path.

[0070] In this embodiment of the invention, S1 includes the following sub-steps:

[0071] S11. Divide the sea area into several uniform grid units;

[0072] S12. Based on several grid cells, time-varying ocean currents are generated by superimposing Gaussian vortex models;

[0073] S13, Update the obstacle positions in each grid cell;

[0074] S14. Determine the local observation range of each AUV;

[0075] S15. Generate a local observation tensor based on the local ocean current velocity components of the time-varying ocean current, the updated obstacle positions of each grid cell, the local observation range of each AUV, and the dynamic target heat map.

[0076] Divide the sea area into A uniform grid of cells, each cell recording the ocean current velocity vector. , Directional components and obstacle presence indicators (1 for obstacle presence, 0 otherwise). Dynamic ocean current fields are simulated using a superimposed Gaussian vortex model; the number of vortices is also considered. ,strength and scale parameters They each follow a uniform distribution and When the environment is reset, obstacle locations are randomly generated, and the obstacle distribution matrix of the global mesh is updated. .

[0077] In this embodiment of the invention, in S2, the Markov decision model The expression is:

[0078] ;

[0079] in, For state space, For the action space, For the reward function, This is the state transition function. This is the discount factor.

[0080] In this embodiment of the invention, the state space includes a local observation tensor and a navigation state tensor; the navigation state tensor includes the AUV's pose, velocity, normalized energy reserve, and mission flags.

[0081] state space From the local observation tensor With navigation state vector constitute, It includes local ocean current velocity components, obstacle distribution matrix, nearby AUV location matrix, and dynamic target heat map (generated by Gaussian kernel function); Pose including AUV ,speed Normalized energy reserves and mission markers ;in, The local observation grid height (referring to the number of rows of the environmental grid within the local sensing range of the AUV). The width of the local observation grid (referring to the number of columns of the environmental grid within the local sensing range of the AUV). No. The x-axis coordinates of each AUV For the first The y-axis coordinate of each AUV For the first The heading angle of an AUV, For the first The velocity components of an AUV in the x-axis direction For the first The velocity component of an AUV in the y-axis direction.

[0082] The action space is composed of the combined space of heading angle adjustment and speed adjustment;

[0083] reward function The expression is:

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] in, Rewards for completing the task. Penalty for obstacle avoidance in safety Penalty for energy efficiency As a reward for path smoothness, This represents the total number of dynamic target points in the scene. For index, AUV With the goal Euclidean distance, To control the rate of reward decay, The penalty coefficient for repeated coverage. The total number of AUVs AUV With the goal Euclidean distance, This is the collision penalty coefficient between AUVs. To penalize the gradient coefficient, For safe distance threshold, AUV exist Position coordinates at that moment AUV exist Position coordinates at that moment As an obstacle avoidance penalty item, This represents the total number of obstacles detected by the AUV. As a safety threshold, The coordinates of the obstacle's position. AUV The amount of heading angle adjustment, AUV Speed ​​adjustment amount, The maximum speed threshold for AUVs. Adjust the penalty weight to speed. Adjust the energy consumption penalty coefficient to speed. The maximum number of time steps in a single round. AUV exist The heading angle at any moment, AUV exist The heading angle at any moment, The reward weight for task completion. The reward weighting for energy efficiency penalties. The reward weight for path smoothness reward, It is a logarithmic function; the speed adjustment penalty weight is used to balance the energy consumption effects of changes in heading and speed.

[0090] To address the policy conflict between obstacle avoidance requirements and task efficiency that often arises in traditional fixed-weight designs in densely obstacle-saturated scenarios, an adaptive reward weight adjustment mechanism based on real-time environment perception is proposed. The safe obstacle avoidance penalty function is designed such that when the distance between any two AUVs is less than a safe distance threshold... Or the distance between the AUV and the obstacle is less than the safety threshold. At that time, a negative reward function that decreases with distance is applied.

[0091] , ,in, For the first AUV in The speed of time For the first AUV in The speed of time.

[0092] The state transition function includes the AUV's pose, velocity, energy reserve, and mission flags;

[0093] The pose state of the AUV includes position. and heading angle , by action Driver update, among which, For AUV in x-axis coordinate at time [time] For AUV in The y-coordinate at time [time]. This is the heading angle adjustment amount. This is the speed adjustment amount.

[0094] Heading angle update:

[0095] ;

[0096] in, Let be the heading angle of the AUV at time t+1. Let be the heading angle of the AUV at time t. This is the amount of heading angle adjustment during the maneuver.

[0097] Location update: based on current speed and heading angle Calculate the x-component of velocity. and y component Its expression is:

[0098] ;

[0099] ;

[0100] Time step The displacement within is:

[0101] ;

[0102] ;

[0103] The expression for updating the pose is:

[0104] ;

[0105] ;

[0106] in, For AUV in x-axis coordinate at time [time] For AUV in The y-coordinate at time [time]. For AUV in x-axis coordinate at time [time] For AUV in The y-coordinate at time [time]. for Time's up The x-axis displacement increment at time t. for Time's up The y-axis displacement increment at time t. At the current speed, For heading angle, For time step;

[0107] Speed ​​adjustment during the movement via heading rotation matrix Update speed after direction correction:

[0108] ;

[0109] Among them, the heading rotation matrix Used for direction correction.

[0110] The speed update expression is:

[0111] ;

[0112] in, for The velocity vector at time t, for The velocity vector at time t, For speed adjustment amount, This is the heading rotation matrix;

[0113] The update expression for energy reserves is:

[0114] ;

[0115] in, for Normalized energy reserves at any given moment for Normalized energy reserves at any given moment As the first energy consumption weighting coefficient, For course adjustment, This is the second energy consumption weighting coefficient;

[0116] The update expression for the task flag is:

[0117] ;

[0118] in, for The task marker at any moment, for The task marker at any moment, Let x be the x-coordinate of the target point. The ordinate of the target point, This represents the effective radius. The task flag is 0 or 1.

[0119] In this embodiment of the invention, S3 includes the following sub-steps:

[0120] S31. Input the local observation tensor into the dual-branch cross-attention network, extract the local feature map using the convolutional layer, enhance the feature representation of the local feature map using the residual module, and then obtain the environmental feature vector through the global average pooling layer.

[0121] S32. Input the navigation state vector into the dual-branch cross-attention network, use the first-level fully connected layer to obtain the intermediate vector, and use the second-level fully connected layer to obtain the navigation feature vector.

[0122] S33. Map the environmental feature vector and navigation feature vector to query vector, key vector and value vector respectively, and calculate the fused features by scaling the dot product attention layer;

[0123] S34. Q-value estimation by mapping fused features to discrete action space using fully connected layers.

[0124] In this embodiment of the invention, in S33, the expression for calculating the fused features by scaling the dot product attention layer is:

[0125] ;

[0126] ;

[0127] ;

[0128] ;

[0129] in, For query vector, For key vectors, For value vectors, For activation function, The dimension of the key vector. For attention mechanisms, For environmental feature vectors, For navigation feature vectors, As the first learnable parameter, As the second learnable parameter, As the third learnable parameter, This represents the matrix transpose operator.

[0130] In this embodiment of the invention, in S4, the dual Q network includes a behavior network and a target network;

[0131] In S4, the action strategy is adjusted through priority experience playback and dynamic learning rate adjustment.

[0132] In this embodiment of the invention, the priority of each experience is determined during priority experience playback. The expression is:

[0133] ;

[0134] in, For TD error, This is a smoothing term; its value is a positive number close to 0.

[0135] TD error ( (Discount factor); sampling probability ( ), and use importance sampling weights. ( It grows linearly from 0.4 to 1.0. Correction deviation for experience pool capacity; where, For instant rewards, The Q-value function of the target network. For the next state, For the first Priority of each sample The power of α is the priority intensity parameter.

[0136] The expression for dynamic learning rate adjustment is:

[0137] ;

[0138] in, The current learning rate, The initial learning rate, To adjust the coefficient, Let the loss function be the variance of the fluctuation. The initial loss variance, For the maximum policy entropy, This represents the current policy entropy.

[0139] In this embodiment of the invention, S5, distributed collaborative learning includes multi-AUV experience collaborative sharing, environmental potential field fusion correction, and policy collaborative evolution.

[0140] In this embodiment of the invention, the expression for multi-AUV experience collaborative sharing is:

[0141] ;

[0142] ;

[0143] ;

[0144] in, For the sample Mixed priority sampling probability, The current empirical sample number to be sampled. These are the global weighting coefficients. Samples for the global experience pool Priority weights, The first in the global experience pool The priority weight of each sample, For other sample index labels, The priority weights for the local experience pool. The first in the local experience pool The priority weight of each sample, For TD error, For TD error priority intensity parameters, For state Global coverage heatmap coefficient at the location, For the sample The corresponding environment state vector, To cover the gain intensity parameter, For activation function, To coordinate the number of AUVs, The length of the trajectory time window. The time decay factor, For index, For the first AUV in The state vector at time t, Current state The position x-axis coordinate, Current state The y-axis coordinate of the position. Radius of spatial influence;

[0145] The expression for the environmental potential field fusion correction is:

[0146] ;

[0147] ;

[0148] ;

[0149] in, To determine the joint Q-value after integrating the environmental potential field, This represents the raw Q-value output by the current network. It is an exponential function. For the first The local obstacle avoidance potential field of an AUV The AUV is in a state The local potential field value at that time, This indicates the current status of the AUV. For the neighboring AUV set, The index number of the adjacent AUV. The repulsion coefficient of neighboring AUVs, The current state The Euclidean distance between the AUV and neighboring AUVs / obstacles For the first The location coordinates of the nearest AUV The obstacle repulsion coefficient has not been verified. For the first A newly detected unverified obstacle by an AUV. In the state Next action The joint Q value at time, The motion value of the AUV. These are the learnable parameters of the current network. For the current network state Next action The Q value output at that time. The global situation field coupling coefficient, The coupling coefficient is the local potential field. For global navigation potential field;

[0150] , , Obstacle detection confidence weights (based on the number of correct detections by the AUV) Number of false alarms calculate), The verified obstacle repulsion constant is taken as 0.5. Let be the gravitational constant at the target point (taken as 0.3). Let be the set of target points.

[0151] Loss function of co-evolution of strategies The expression is:

[0152] ;

[0153] ;

[0154] in, For the joint loss function, The time-varying consensus loss coefficient, To train time steps, For the strategy consensus loss item, For the total number of coordinated AUVs, For the first The policy network parameters for each AUV, For the first The policy network parameters for each AUV.

[0155] Real-time guidance through dynamic rewards: During path execution, the weights of the reward function (task completion, obstacle avoidance, energy consumption, etc.) are dynamically adjusted based on the current environmental state (target distance, obstacle density). For example, when the AUV approaches the target, the task completion reward weight increases from 20% to 70%, forcing the path to converge towards the target; when entering a high-obstacle area, the obstacle avoidance penalty weight increases from 0.4 to 0.8, prioritizing collision avoidance. This mechanism ensures a dynamic balance between exploration and utilization, and between global objectives and local safety.

[0156] Real-time response to potential field correction: In response to sudden ocean currents or new obstacles, the path is double-corrected through the global navigation potential field (updated every 100 steps) and the local obstacle avoidance potential field (updated every 1 step) constructed by S5: the global potential field guides the path toward the target and avoids known obstacles, while the local potential field pushes away new obstacles or nearby AUVs in real time (when the distance is <2.5m).

[0157] Global optimization of multi-AUV collaboration: Through experience sharing and parameter synchronization in S5, each AUV shares high-value experience (such as obstacle avoidance strategies in high-density obstacle areas) and synchronizes parameters every 100 steps (20% global strategy + 80% local experience) to avoid strategy homogenization; at the same time, through conflict resolution (repulsion force of neighboring AUVs) and task allocation (penalty for repeated coverage), it ensures that the group path does not overlap and the target coverage is efficient.

[0158] In embodiments of the present invention, such as Figure 2 As shown, the network's dual-branch heterogeneous feature fusion architecture is generally divided into four parts: the environment perception branch, the navigation state branch, the cross-attention fusion layer, and the output layer. The environment perception branch is shown on the upper left, with the local environment observation tensor represented by a green cuboid. The input is 20×20×5 dimensionality. First, it passes through an orange convolutional layer labeled "3×3 kernel" and "ReLU", outputting a 20×20×32 feature map. Then, it is reduced to 10×10×32 using a "2×2 pooling kernel". Next, it enters the core residual module. The main path contains two cascaded orange convolutional layers with "3×3 kernel" and "ReLU", expanding the number of channels from 32 to 64 and outputting a 10×10×64 feature map. Simultaneously, the left branch uses a blue "1×1 convolution" module to upscale the pooled 10×10×32 feature map to 64 channels, which is then residually connected to the main path output via a "⊕" symbol, where "⊕" represents element-wise addition in the residual connection. Afterward, it is reduced to 5×5×64 using a "2×2 pooling kernel", and then compressed into a 64×1 environmental feature vector using a green "global average pooling" module. The bottom left side displays the navigation status branch, with each circle containing multi-dimensional navigation status information such as "v, e, g, φ". The input is a 7×1 vector, which is expanded to 128×1 by the blue "fully connected layer", activated by LeakyReLU, passed through the orange "regularization" module (containing 30% Dropout), and then compressed into a 64×1 navigation feature vector by the blue "fully connected layer". The pink "cross-attention fusion layer" in the middle receives... and ,Will Mapped to query vector , Mapped to key vector Sum value vector 64-dimensional fused features are generated through a scaled dot product attention mechanism. Finally, the blue "fully connected layer" on the right outputs a 9-dimensional Q-value. to ), corresponding to the discrete action space to guide AUV decision-making.

[0159] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A multi-AUV adaptive path planning method for complex underwater environments, characterized in that, Includes the following steps: S1. Divide the sea area into grids and generate local observation tensors; S2. Construct a Markov decision model based on the local observation tensor and the navigation state tensor; S3. Based on the Markov decision model, the local observation tensor and the navigation state tensor are input into a two-branch cross-attention network to generate Q-value estimates; S4. Adjust the Q-value estimate using a dual Q-network; S5. Perform distributed collaborative learning; S6. Based on the adjusted Q-value estimation, generate an initial path and use the results of distributed collaborative learning to correct the initial path to obtain the globally optimal path.

2. The multi-AUV adaptive path planning method for complex underwater environments according to claim 1, characterized in that, S1 includes the following sub-steps: S11. Divide the sea area into several uniform grid units; S12. Based on several grid cells, time-varying ocean currents are generated by superimposing Gaussian vortex models; S13, Update the obstacle positions in each grid cell; S14. Determine the local observation range of each AUV; S15. Generate a local observation tensor based on the local ocean current velocity components of the time-varying ocean current, the updated obstacle positions of each grid cell, the local observation range of each AUV, and the dynamic target heat map.

3. The multi-AUV adaptive path planning method for complex underwater environments according to claim 1, characterized in that, In S2, the Markov decision model The expression is: ； in, For state space, For the action space, For the reward function, This is the state transition function. This is the discount factor.

4. The multi-AUV adaptive path planning method for complex underwater environments according to claim 3, characterized in that, The state space includes a local observation tensor and a navigation state tensor; the navigation state tensor includes the AUV's pose, velocity, normalized energy reserve, and mission flags. The action space is composed of the combined space of heading angle adjustment and speed adjustment; The reward function The expression is: ； ； ； ； ； in, Rewards for completing the task. Penalty for obstacle avoidance in safety Penalty for energy efficiency As a reward for path smoothness, This represents the total number of dynamic target points in the scene. For index, AUV With the goal Euclidean distance, To control the rate of reward decay, The penalty coefficient for repeated coverage. The total number of AUVs AUV With the goal Euclidean distance, This is the collision penalty coefficient between AUVs. To penalize the gradient coefficient, For safe distance threshold, AUV exist Position coordinates at that moment AUV exist Position coordinates at that moment As an obstacle avoidance penalty item, This represents the total number of obstacles detected by the AUV. As a safety threshold, The coordinates of the obstacle's position. AUV The amount of heading angle adjustment, AUV Speed ​​adjustment amount, The maximum speed threshold for AUVs. Adjust the penalty weight to speed. Adjust the energy consumption penalty coefficient to speed. The maximum number of time steps in a single round. AUV exist The heading angle at any moment, AUV exist The heading angle at any moment, The reward weight for task completion. The reward weighting for energy efficiency penalties. The reward weight for path smoothness reward, It is a logarithmic function; The state transition function includes the AUV's pose, velocity, energy reserve, and mission flags.

5. The multi-AUV adaptive path planning method for complex underwater environments according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Input the local observation tensor into the dual-branch cross-attention network, extract the local feature map using the convolutional layer, enhance the feature representation of the local feature map using the residual module, and then obtain the environmental feature vector through the global average pooling layer. S32. Input the navigation state vector into the dual-branch cross-attention network, use the first-level fully connected layer to obtain the intermediate vector, and use the second-level fully connected layer to obtain the navigation feature vector. S33. Map the environmental feature vector and navigation feature vector to query vector, key vector and value vector respectively, and calculate the fused features by scaling the dot product attention layer; S34. Q-value estimation by mapping fused features to discrete action space using fully connected layers.

6. The multi-AUV adaptive path planning method for complex underwater environments according to claim 5, characterized in that, In step S33, the expression for calculating the fused features by scaling the dot product attention layer is as follows: ； ； ； ； in, For query vector, For key vectors, For value vectors, For activation function, The dimension of the key vector. For attention mechanisms, For environmental feature vectors, For navigation feature vectors, As the first learnable parameter, As the second learnable parameter, As the third learnable parameter, This represents the matrix transpose operator.

7. The multi-AUV adaptive path planning method for complex underwater environments according to claim 1, characterized in that, In S4, the dual-Q network includes a behavior network and a target network; In S4, the action strategy is adjusted by prioritizing experience playback and dynamically adjusting the learning rate.

8. The multi-AUV adaptive path planning method for complex underwater environments according to claim 7, characterized in that, In the priority experience playback, the priority of each experience is... The expression is: ； in, For TD error, For smoothing terms; The expression for the dynamic learning rate adjustment is: ； in, The current learning rate, The initial learning rate, To adjust the coefficient, Let the loss function be the variance of the fluctuation. The initial loss variance, For the maximum policy entropy, This represents the current policy entropy.

9. The multi-AUV adaptive path planning method for complex underwater environments according to claim 1, characterized in that, In S5, distributed collaborative learning includes multi-AUV experience collaborative sharing, environmental potential field fusion correction, and policy collaborative evolution.

Images