A fault-tolerant method to improve the robustness of underwater robot networking

Abstract

The present invention discloses a fault-tolerant method for improving the robustness of underwater robot networking, comprising: adopting target feature information and utilizing an improved ant colony algorithm to generate an optimal networking scheme for the role division and spatial position distribution of robot nodes; the target feature information is transmitted to a leader node through an underwater acoustic communication link, and the target information and task requirements are broadcast to neighboring nodes to trigger a collaborative operation mechanism; in the collaborative operation process, a dynamic partitioning collaborative strategy based on Thiessen polygons is adopted, the leader node performs preliminary fusion and processing of the received data, introduces a multi-classification fault diagnosis model for determining the type and severity of the fault, and adopts a dynamic path reconstruction algorithm to re-plan the position and path of the task node; in the multi-robot collaborative operation process, each node estimates and models the probability distribution of the target through Bayesian reasoning and evidence theory methods; after the operation task is completed, all data is transmitted to a data processing center on a mother ship.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to a fault-tolerant method for improving the robustness of underwater robot networking. Background Art

[0002] When performing tasks, underwater robots often need to move and work together in a vast underwater area. Due to the complex and changeable underwater environment, there are currents, obstacles, water temperature, water flow, visibility changes and electromagnetic interference factors, which pose severe challenges to the networking and communication of robots. Robots need to share their position, speed, water depth, and surrounding environment information in real time, and coordinate to adjust task allocation to adapt to the changing environment. At the same time, due to limited underwater communication conditions, data transmission between robots will be affected by delays and packet loss, affecting the overall task execution efficiency. In addition, during long-term underwater operations, some robots will exit the network due to energy exhaustion, damaged waterproof performance or failure. How to dynamically adjust the network topology to ensure the redundancy and reliability of data transmission is a key issue. Robots need to dynamically join or exit the network according to changes in their own status and surrounding environment, and adjust the communication path and task allocation in time to maintain the high fault tolerance and robustness of the entire networking system. In this process, efficient networking protocols and scheduling algorithms maximize the role of each robot and achieve overall coordination and evolution, which is a complex technical challenge.

[0003] The present invention proposes a solution to the above shortcomings: collect target feature information and underwater environmental parameters, generate an optimal networking solution in combination with the performance parameters of heterogeneous underwater robots, determine the fault tolerance level of each node, and when a robot node detects a new target, transmit the information to the leader node for evaluation, schedule the node to perform tasks, and trigger the collaborative operation mechanism, allocate sub-areas according to the fault tolerance level of each robot node, collect data and transmit it in real time, if the leader node fails, the backup node takes over, and a detailed three-dimensional model is generated, if a node fails, the backup node is requested to take over the task, and the emergency procedure is started, during the collaborative operation of each node, the probability distribution of the target is estimated and modeled through reasoning and learning methods, and the recognition model is adjusted in real time. After the task is completed, all data is transmitted to the data processing center of the mother ship for comprehensive processing and three-dimensional visualization, and is transmitted to the shore-based research center in real time via satellite communication. Summary of the invention

[0004] In order to solve the above-mentioned technical problems, the present invention provides a fault-tolerant method for improving the robustness of underwater robot networking.

[0005] The technical solution of the present invention is implemented as follows: A fault-tolerant method for improving the robustness of underwater robot networking, comprising:

[0006] In S101, sonar and underwater cameras are used to collect target feature information, including type, size and location, and underwater environmental parameters such as water temperature, water flow and visibility are obtained. According to the energy state, waterproof performance, data cache and computing load of the heterogeneous underwater robots, combined with the depth of underwater operations and regional terrain characteristics, an improved ant colony algorithm is used to generate the optimal networking scheme of the role division and spatial position distribution of the robot nodes. A fuzzy comprehensive evaluation method is used to consider the performance parameters of each robot node to determine the fault tolerance level of each node at different underwater operation depths.

[0007] In S102, when a robot node detects a new target, it transmits the target feature information to the leader node through the underwater acoustic communication link. The leader node evaluates the value of the target and the urgency of the task based on historical data, knowledge base and underwater environmental parameters. If the target value is high and the task is urgent, the leader node will prioritize the nodes with good waterproof performance and high fault tolerance level to perform the task, and broadcast the target information and task requirements to the neighboring nodes to trigger the collaborative operation mechanism.

[0008] In S103, during the collaborative operation, a dynamic partitioning collaborative strategy based on Thiessen polygons is adopted, and the performance parameters of each robot node and the characteristics of the underwater environment are comprehensively considered to divide the area into several sub-areas. Robots with high waterproof levels and large thrust of thrusters are allocated to areas with high water flow speed, and robots with strong obstacle avoidance capabilities are allocated to areas with dense obstacles. The formation and obstacle avoidance behavior of the robot cluster are determined by the artificial potential field method, and the path planning algorithm is used to adjust the robot motion trajectory in real time to avoid obstacles, and the position is dynamically adjusted according to the change of water flow.

[0009] In S104, each robot node collects data in its assigned sub-area and transmits it to the leader node in real time through the underwater acoustic communication link. The leader node initially integrates and processes the received data. If the leader node fails, the backup node with excellent waterproof performance and the highest fault tolerance level takes over its role. The leader node uses a distributed joint Kalman filter algorithm and an iterative closest point algorithm to generate a fine three-dimensional model of the target, dynamically adjust the model parameters, and record the impact of water flow and temperature changes on the environment.

[0010] In S105, according to the operation process of S104, if a robot node fails, the fault type and severity are determined by a multi-classification fault diagnosis model. For minor faults, the task priority of the node is reduced, and it is assigned to continue to perform low-priority tasks. For serious faults, the node sends a help signal to the neighboring nodes, requesting the neighboring nodes to take over the task, and at the same time starts the emergency response procedure, broadcasts the fault information through the underwater acoustic communication link, updates the network topology, and uses a dynamic path reconstruction algorithm to re-plan the location and path of the task node;

[0011] In S106, during the multi-robot collaborative operation, each node estimates and models the probability distribution of the target by combining Bayesian reasoning and evidence theory, and uses transfer learning and incremental learning methods to update and optimize the target recognition model in real time. The robot in the deep water area adjusts the output power of the propeller, and the robot in the shallow water area improves the propeller efficiency and data transmission frequency.

[0012] In S107, after the operation task is completed, all data are transmitted to the data processing center on the mother ship. The data processing center uses big data analysis and virtual reality technology to comprehensively process and three-dimensionally visualize the collected data, generate a panoramic three-dimensional model and virtual reality scene of the area, and use the water flow and temperature data collected by the robot to create a dynamic environmental model, and intuitively display the environment and conditions of underwater operations through virtual reality technology.

[0013] Beneficial Effects

[0014] The present invention collects target feature information and combines it with underwater environmental parameters such as water temperature, water flow and visibility to generate the optimal networking scheme for the role division and spatial position distribution of robot nodes, and determines the fault tolerance level of each node at different underwater operating depths, thereby optimizing the networking scheme. When a robot node detects a new target, the system prioritizes nodes with good waterproof performance and high fault tolerance level to perform tasks, ensuring the smooth completion of key tasks. In the collaborative operation process, a dynamic partitioning collaboration strategy is adopted to allocate robot nodes to sub-areas suitable for their performance characteristics, and adjust the robot's motion trajectory in real time to avoid obstacles. , and dynamically adjusts its position according to changes in water flow, thereby improving the formation and obstacle avoidance capabilities of the robot cluster. Each robot node collects data in its assigned sub-area and transmits it to the leader node in real time. The leader node initially integrates and processes the received data, and in the event of a failure, the backup node takes over its role to ensure the continuity of data processing and task scheduling. The system also effectively responds to minor and serious failures and ensures the stable operation of the entire networking system. Finally, through online fine-tuning of various models and dynamic environment modeling, the present invention achieves improvements in the efficiency, stability, and robustness of the underwater robot networking system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A structural block diagram of a fault-tolerant method for improving the robustness of underwater robot networking in an embodiment of the present invention;

[0016] Figure 2 The present invention is a flowchart of a method for improving the robustness of underwater robot networking according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0018] The preferred implementation methods of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these implementation methods are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0019] It should be noted that when an element is considered to be "connected" to another element, it can be directly connected to the other element, or connected to the other element through an intermediate element. In addition, the "connection" in the following embodiments should be understood as "electrical connection", "communication connection", etc. if there is transmission of electrical signals or data between the connected objects.

[0020] When used herein, the singular forms "a", "an", and "said / the" may also include plural forms, unless the context clearly indicates otherwise. It should also be understood that the terms "include / comprise" or "have" and the like specify the presence of stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not exclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. At the same time, the term "and / or" used in this specification includes any and all combinations of the relevant listed items.

[0021] See also Figure 1-2 As shown, a fault-tolerant method for improving the robustness of underwater robot networking specifically includes:

[0022] In S101, high-resolution sonar and underwater cameras are used to collect target feature information, including type, size and location, and underwater environmental parameters such as water temperature, water flow and visibility are obtained. Based on the energy state, waterproof performance, data cache and computing load of heterogeneous underwater robots, combined with the depth of underwater operations and regional terrain characteristics, an improved ant colony algorithm is used to generate the optimal networking solution for the role division and spatial position distribution of robot nodes. A fuzzy comprehensive evaluation method is used to comprehensively consider the performance parameters of each robot node to determine the fault tolerance level of each node at different underwater operation depths.

[0023] Based on the target feature information collected by high-resolution sonar and underwater cameras, including target type, size and location, as well as underwater environmental parameters such as water temperature, current and visibility, a multi-dimensional data model of underwater targets and environments is established;

[0024] The fuzzy C-means clustering algorithm is used to classify the energy state, waterproof performance, data cache and computing load parameters of heterogeneous underwater robots to obtain the performance feature vectors of different types of robots. The relevant information in the underwater target and environment data model is integrated into the clustering results as the basis for the division of robot node roles.

[0025] Through the improved ant colony algorithm, the depth of underwater operations, regional terrain characteristics and the performance characteristics of each robot node are comprehensively considered, and a robot network coverage model based on graph theory is constructed. The quality of the communication link between nodes and the coverage area are used as optimization goals. The spatial position distribution of the robot nodes is searched and optimized to maximize the coverage of the robot network and meet the operation requirements under different water depths and terrain conditions.

[0026] The fuzzy comprehensive evaluation method is used to evaluate the fault tolerance of each robot node. The evaluation indicators include the remaining energy of the node, the quality of the communication link, and the computing load. Each indicator is quantified by scoring, and the weight is set according to the importance of the indicator. Finally, the weighted sum is used to obtain the comprehensive evaluation result of the node fault tolerance, and the fault tolerance level of each node at different underwater operating depths is obtained.

[0027] Dynamically adjust the role division of robot nodes according to the evaluation results, and assign nodes with stronger fault tolerance to more critical or higher-risk tasks. During the robot networking process, the node power supply voltage and current are regularly measured to estimate the node energy consumption rate, thus achieving real-time monitoring of the node energy status.

[0028] When the energy of a node drops to a preset threshold, some tasks of other nodes are reallocated through task migration and load balancing algorithms to optimize the energy balance between nodes, thereby extending the working time of the entire robot network.

[0029] For different types of underwater operation tasks, including seabed topography mapping, underwater archaeology, and pipeline laying, a task tree generation method based on semantic analysis is used to divide the task into multiple subtasks. Then, the market bidding mechanism and contract network protocol of multi-robot task allocation are combined to dynamically allocate and schedule tasks according to the capability characteristics of robot nodes, thereby improving the operation efficiency and reliability of the entire robot network.

[0030] During the task execution process, the target information obtained by the robot nodes is fed back to the network optimization model, and the spatial distribution of the nodes is dynamically adjusted to achieve adaptive optimization of the network performance.

[0031] Specifically, in S101, in the process of underwater target and environment data modeling, the octree space partitioning data structure is used to recursively partition the three-dimensional space, and the high-resolution sonar point cloud data and visual image features are mapped to the corresponding spatial grids, and the underwater space is divided into 256 grids, and the resolution of each grid is 1 cubic meter;

[0032] In each grid, the target feature information of the target is extracted, including shape, size, and texture, and the distribution of underwater environmental parameters such as water temperature, water flow velocity, and visibility is statistically analyzed to form a multi-dimensional data representation. For the clustering of robot performance parameters, the fuzzy C-means algorithm is used, and the number of clusters is set to 3-5, the number of iterations is 50, and the convergence threshold is 0.01;

[0033] In the clustering process, the similarity between nodes is calculated according to the energy state, waterproof performance level, and data cache capacity parameters of heterogeneous robots, and the fuzzy membership matrix is ​​generated. The cluster center is solved by optimizing the objective function to obtain the performance characteristic vectors of different types of robots.

[0034] In the optimization of robot network coverage, the maximum coverage problem model based on graph theory is adopted to divide the underwater space into 100 areas, each with an area of ​​100 square meters. The robot nodes are the nodes of the graph, the communication links between the nodes are the edges of the graph, and the signal attenuation value is the weight of the edge. An undirected weighted graph model is constructed. Then, a greedy algorithm is used to solve the maximum coverage subgraph. Each time, the node that covers the most new areas is selected to join the subgraph until the total coverage area reaches more than 90%;

[0035] In the fault tolerance evaluation, the hierarchical analysis method is used to determine the weight of each indicator. The weight of the remaining energy is 0.4, the weight of the communication link quality is 0.3, and the weight of the computing load is 0.3. For each indicator, it is divided into 5 levels and assigned 1-5 points respectively. Then, the weighted average method is used to calculate the comprehensive score of the node. Nodes with scores exceeding 4 points undertake key tasks, and nodes with scores below 2 points need to reduce the task load;

[0036] In the process of task decomposition and allocation, the task ontology description method based on semantic network is adopted to divide the task into multiple atomic tasks, and the hierarchical relationship and logical relationship between tasks are expressed by ontology language OWL. The submarine pipeline laying task is divided into seabed cleaning, pipeline transportation, pipeline docking, backfilling and burial atomic tasks, and the order relationship and resource demand relationship between tasks are defined;

[0037] In multi-robot task allocation, a market bidding mechanism is adopted. Each robot node estimates the cost of completing the task based on its own capabilities and status, and submits a bid price. Then, task allocation is carried out through the contract network protocol, and the task is allocated to the node combination with the lowest bid price and that can meet the task requirements.

[0038] In S102, when a robot node detects a new target, it transmits the target feature information to the leader node through the underwater acoustic communication link. The leader node evaluates the value of the target and the urgency of the task based on historical data, knowledge base and underwater environmental parameters. If the target value is high and the task is urgent, the leader node will prioritize the nodes with good waterproof performance and high fault tolerance level to perform the task, and broadcast the target information and task requirements to the neighboring nodes to trigger the collaborative operation mechanism.

[0039] When a robot node detects a new target through high-resolution sonar and underwater camera, it extracts the target’s type, size, and location target feature information, and combines the node’s own position and posture data to generate the target’s three-dimensional spatial coordinates and motion trajectory vector. The target information is encapsulated as a message in JSON format and reliably transmitted to the leader node through the underwater acoustic communication link. At the same time, an ACK confirmation mechanism is set to ensure the complete delivery of the information.

[0040] During the communication process, the AES encryption algorithm is used to encrypt and protect sensitive information. After receiving the target information, the leader node extracts key feature parameters and performs similarity matching with the historical target records in the local MongoDB database to identify the category and attributes of the target. At the same time, it accesses the ontology-based knowledge base and uses the SPARQL query language to retrieve the value assessment rules and disposal strategies for this type of target, taking into account the target characteristics, historical data analysis results, knowledge base, and current underwater environmental parameters.

[0041] The method of combining fuzzy reasoning and hierarchical analysis is used to evaluate the target value and task urgency. First, three levels are set for the target value and urgency, namely low, medium and high, and the triangular membership function of each level is defined. Then, the influencing factors are compared pairwise, the judgment matrix is ​​constructed and the weight vector is calculated. Finally, the membership and weight are multiplied and summed to obtain the comprehensive score of the target value and urgency.

[0042] If both the target value score and the task urgency score exceed the threshold, the emergency task handling mechanism is triggered, and the leader node builds a multi-constrained task allocation model based on the performance parameters of the heterogeneous robot cluster, including waterproof level, battery capacity, computing power, and fault tolerance;

[0043] Among them, the objective function is the task execution efficiency of the robot node, and the constraints include the node's energy budget, communication bandwidth limit and task completion time window. The improved whale optimization algorithm is used to solve the problem, and the optimal task allocation solution that meets the constraints is obtained. The task instructions are then sent to the corresponding robot nodes. After receiving the task instructions, the robot nodes assigned the tasks extract the target information and task requirements broadcast by the leader node, and combine their own status and capabilities. The autonomous path planning algorithm based on the Markov decision process is used to generate underwater navigation paths, speed planning and target capture strategy control instructions.

[0044] At the same time, it broadcasts collaborative work requests to neighboring nodes in the form of ROS-compatible messages, and synchronizes its own status information, target information, and control instructions to the collaborative nodes through the TCP / IP protocol. After receiving the collaborative work request, the collaborative node extracts the target information and task requirements, combined with its own status, waterproof performance, and fault tolerance;

[0045] A distributed task negotiation protocol based on Contract Net is used to negotiate task division and resource scheduling with the initiating node. The negotiation process sends Call for Proposal, Proposal, Accept / Reject messages and exchanges node resource status and task execution progress information to reach a consensus on the collaboration plan. Based on the division of labor results, each node generates and executes local control instructions, and keeps synchronization with the initiating node in time and space.

[0046] During the collaborative operation, each robot node tracks the target's motion trajectory and feature changes in real time through acoustic positioning, inertial navigation, and visual odometer multi-source sensor fusion, and regularly exchanges tracking results and status information through the ROSTf service. When the similarity of the target feature is lower than the preset minimum threshold, or the node's own energy level is lower than the red line, the task re-planning mechanism is activated to re-evaluate the target value and task urgency, adjust the node task division and resource scheduling, and generate updated control instructions to adapt to the dynamic changes of the target and environment. After each robot node completes its respective tasks, it packages the target capture status, node trajectory data, and sensor data and sends them to the leader node. The leader node performs statistical analysis on the data reported by each node, calculates the key performance indicators of time efficiency, energy efficiency, and target capture success rate of collaborative task completion, and compares them with the preset task completion standards to obtain quantitative task quality evaluation results. The evaluation results are fed back to each participating node together with the node's rewards and penalties to update the node's own strategy model and collaboration mechanism, and continuously improve the collaborative operation performance of the heterogeneous robot cluster. Finally, the leader node summarizes the task execution report and data log to the surface base station, which is archived and analyzed by the task management system.

[0047] Specifically, as described in S102, when the robot node detects a new target through high-resolution sonar and underwater cameras, it first extracts characteristic information of the target's type, size, and position, and identifies the target as a torpedo through an image segmentation algorithm, which is 5 meters long and 5 meters in diameter and is located 50 meters in front of the node. The node determines that it is located at 35 degrees north latitude, 135 degrees east longitude, 50 meters underwater, and has a pitch angle of 5 degrees and a yaw angle of 30 degrees through a SLAM algorithm combined with IMU and DVL sensor data.

[0048] According to the orientation of the target relative to the node and the spatial position of the node itself, the absolute three-dimensional coordinates of the target are calculated to be (100, 80, -40), and the Kalman filter algorithm is used to estimate the target's movement speed to be 2m / s and the heading angle to be 60 degrees. The node encapsulates the target information in JSON format, encrypts it with AES-256, and transmits it to the leader node through the underwater acoustic communication link;

[0049] After receiving the target information, the leader node extracts key features and performs similarity matching with the historical target records in the MongoDB database. It identifies the target as a training torpedo through the support vector machine (SVM) algorithm, accesses the ontology-based knowledge base, retrieves the value assessment rules and disposal strategies for this type of target, and uses a combination of fuzzy reasoning and hierarchical analysis (AHP) based on the target features, historical data analysis, knowledge base, and current environmental parameters such as water temperature (18 degrees Celsius) and turbidity (5 degrees Celsius) to obtain a target value score of 8 and a task urgency score of 9, both of which exceed the preset threshold of 7, triggering the emergency task disposal mechanism.

[0050] The leader node builds a multi-constrained task allocation model based on the performance parameters of the heterogeneous robot cluster, including waterproof level IPX8, battery capacity 5000mAh, and CPU main frequency 5GHz, and uses the improved whale optimization algorithm WOA to solve and obtain the optimal task allocation solution;

[0051] The target tracking task is sent to the underwater glider node, and the salvage task is sent to the ROV node. After receiving the command, the task node extracts the target information and task requirements, and uses the autonomous path planning algorithm based on the Markov decision process, combined with A* search and artificial potential field method, to generate underwater navigation path, speed planning and target capture strategy control instructions, and at the same time broadcasts collaborative operation requests to neighboring AUV nodes in the form of ROS messages;

[0052] The AUV node negotiates the division of tasks with the task node through the Contract Net protocol, exchanges resource status and progress information, reaches a collaboration plan, and controls the robotic arm and suction cup actuator to assist in target capture. Each node uses Kalman filtering to fuse acoustic positioning, inertial navigation, and visual odometer multi-source sensor data to track target motion and feature changes in real time, and exchanges tracking results through ROS tf services;

[0053] When the target similarity is lower than 6 or the node power is lower than 30%, task re-planning is started, and node division of labor and resource scheduling are adjusted. After the task is completed, each node packages the target capture status, trajectory data, and sensor data and sends them to the leader node. The leader node performs statistical analysis and finds that the average task completion time is 15 minutes, the energy consumption is 1500 joules, and the target capture success rate is 95%, which meets the preset task standards.

[0054] The evaluation results are fed back to each node to update the node's Q-Learning strategy model and collaboration mechanism parameters. Finally, the leader node summarizes the task execution report and data log to the surface base station, which is archived and used for secondary analysis by the task management system.

[0055] In S103, during the collaborative operation, a dynamic partitioning collaborative strategy based on Thiessen polygons is adopted, and the performance parameters of each robot node and the characteristics of the underwater environment are comprehensively considered to divide the area into several sub-areas. Robots with high waterproof levels and large thrust of thrusters are allocated to areas with high water flow speed, and robots with strong obstacle avoidance capabilities are allocated to areas with dense obstacles. The formation and obstacle avoidance behavior of the robot cluster are determined by the artificial potential field method, and the path planning algorithm is used to adjust the robot motion trajectory in real time to avoid obstacles, and the position is dynamically adjusted according to the change of water flow.

[0056] Obtain the position information of each robot node and underwater environmental parameters, including water depth, water temperature and water flow velocity, and use the sequential Gaussian simulation method to perform three-dimensional spatial dynamic modeling of the underwater environmental parameters;

[0057] The underwater operation area is divided by the Thiessen polygon algorithm, taking into account the three-dimensional characteristics of the underwater environment, using the octree spatial data structure, and taking the robot node as the occurrence point of the Thiessen polygon;

[0058] The shape and size of the Thiessen polygon are determined based on the performance parameters of each node. For areas with high water flow velocity, robot nodes with high waterproof levels and large thrust thrust are given priority to perform tasks.

[0059] According to the division results of Thiessen polygons, the high water flow velocity area is allocated to the above high-performance nodes, and the fuzzy control algorithm is used to adjust the speed and direction of the propeller in real time, so that the robot can maintain a stable motion posture and trajectory tracking accuracy in a high water flow environment;

[0060] In areas with dense obstacles, robot nodes equipped with forward-looking sonar, side-scan sonar and high-definition cameras are selected. The Kalman filter algorithm is used to fuse the data of different sensors to build a three-dimensional point cloud model of the obstacle. The artificial potential field method is used to plan the obstacle avoidance path. The obstacle is regarded as a repulsive field and the target position is regarded as a gravitational field. The direction of the resultant force of the robot's motion is solved by the gradient descent method. At the same time, the improved RRT* algorithm is used to plan the obstacle avoidance path in three-dimensional space, generate a series of waypoint coordinate sequences, and perform smoothing to improve the executableness of the path.

[0061] During the obstacle avoidance process, the robot node adjusts the heading and speed in real time through the adaptive PID control algorithm, and dynamically adjusts the PID parameters according to the curvature of the obstacle avoidance path and the distribution of obstacles to minimize the deviation between the actual motion trajectory and the planned path;

[0062] After the obstacle avoidance task is completed, the robot cluster enters the next operating area. In the dynamic water flow environment, it obtains the velocity profile data of each layer collected by the ADCP equipment in real time, builds a three-dimensional water flow field model through the spatiotemporal interpolation algorithm, and uses the ensemble Kalman filter algorithm to make real-time predictions of the water flow field.

[0063] According to the prediction results, the graph-based multi-robot formation control algorithm is used to adjust the spatial position of the nodes to balance the force of the entire formation and improve the anti-flow ability. During the formation control process, the spatial position of each node is optimized to minimize the communication delay and energy consumption between nodes.

[0064] When performing specific underwater operation tasks, including underwater archaeology and submarine pipeline inspection, first select robot nodes to form a formation according to task requirements, and determine the hierarchical structure and number of nodes of the formation according to the complexity of the task. Use a distributed task allocation algorithm based on market mechanisms, comprehensively consider the capabilities, locations and load conditions of each node, and allocate tasks to the optimal node combination through competitive bidding;

[0065] For complex collaborative tasks, a multi-robot collaborative framework based on the Max-sum algorithm is introduced to obtain the global optimal task decomposition and node scheduling solution through iterative communication and calculation between nodes;

[0066] In terms of communication and positioning of underwater robot clusters, D-OFDM technology is used to adaptively adjust the cyclic prefix and subcarrier spacing of OFDM signals according to the multipath delay and Doppler frequency shift characteristics of the underwater acoustic channel to improve the anti-interference capability of communication.

[0067] In terms of positioning, the robot node is positioned with high precision by combining multiple positioning technologies such as LBL, USBL and inertial navigation and performing multi-source information fusion through the loosely coupled extended Kalman filter algorithm.

[0068] A partition quality evaluation method based on the Voronoi diagram is used to quantitatively evaluate the effect of dynamic partitioning. The evaluation indicators include the compactness, coverage completeness and load balance of the partition. If the evaluation result is lower than the preset threshold, the adaptive partition adjustment mechanism is triggered. The location of the occurrence point of the Thiessen polygon is iteratively optimized, and the capacity parameters of the node are dynamically adjusted according to the task execution status.

[0069] Specifically, in S103, the underwater robot cluster obtains node performance parameters, node positions and underwater environment parameter information through acoustic communication, with a sampling frequency of 10 Hz, and uses a sequential Gaussian simulation algorithm to perform three-dimensional space interpolation on 1000 sampling points. A single iteration takes less than 1 second to generate a three-dimensional environment grid model with a resolution of 1 meter. The Thiessen polygon algorithm is used to divide the operation area, and the octree data structure is used to store the division results. The retrieval time complexity is O (logn);

[0070] For high water flow velocity areas, the fuzzy C-means clustering algorithm is used to select the optimal node according to the thrust of the robot node and the waterproof level matching degree. The clustering process converges after 20 iterations. The control frequency is 200Hz. The TS fuzzy controller is used to achieve trajectory tracking, and the control error is reduced to within 5cm.

[0071] In areas with dense obstacles, the obstacle detection accuracy is over 90% based on the fusion of forward-looking sonar, side-scan sonar and stereo camera data by the Kalman filter algorithm. The obstacle avoidance path planning is carried out by combining the artificial potential field method and the RRT algorithm. The gravity coefficient and repulsion coefficient of the artificial potential field method are 0.8 and 1.2 respectively, and the number of sampling points of the RRT algorithm is 1000. The shortest obstacle avoidance path length is shortened by 20%. The parameters of the adaptive PID controller are adjusted online according to fuzzy rules with an adjustment period of 1s, and the trajectory tracking error is reduced by 50%.

[0072] In a dynamic water flow environment, the ADCP device collects 128 layers of water flow profile data at a frequency of 5Hz, uses Kriging interpolation to build a three-dimensional water flow field model, and uses the ensemble Kalman filter algorithm for real-time prediction, with a prediction accuracy of 80%. The formation control uses a graph-based multi-agent consensus algorithm to optimize the spatial position of the 4×4 grid formation, reducing the average communication delay by 30% and the formation energy consumption by 20%.

[0073] In the underwater archaeological mission, six heterogeneous robot nodes formed a two-layer formation, and adopted a market-based task allocation algorithm to allocate tasks to the optimal node combination through three competitive biddings, which shortened the task completion time by 25%. The Max-sum algorithm was used for task decomposition in complex collaborative operations, and each node conducted five rounds of iterative communication with neighboring nodes to obtain the global optimal scheduling solution.

[0074] The underwater robot cluster uses D-OFDM technology for communication. According to the channel estimation results, the interval of 512 subcarriers and the length of 128 cyclic prefixes are adaptively adjusted, and the communication success rate is improved by 15%. In terms of positioning, the loosely coupled extended Kalman filter algorithm is used to fuse 10 sensor data including LBL, USBL, DVL and IMU. The positioning accuracy is better than 1m. When evaluating the dynamic partition quality, the geometric features of the Voronoi diagram are used to calculate the partition compactness and coverage completeness, and the Gini coefficient is used to quantify the load balancing degree. The occurrence point of the Thiessen polygon is adaptively adjusted. After 50 iterative optimizations, the partition quality is improved.

[0075] In S104, each robot node collects data in its assigned sub-area and transmits it to the leader node in real time through the underwater acoustic communication link. The leader node initially integrates and processes the received data. If the leader node fails, the backup node with excellent waterproof performance and the highest fault tolerance level takes over its role. The leader node uses a distributed joint Kalman filter algorithm and an iterative closest point algorithm to generate a fine three-dimensional model of the target, dynamically adjust the model parameters, and record the impact of water flow and temperature changes on the environment.

[0076] Each robot node determines the sub-area to which it belongs based on the results of the Thiessen polygon division, and autonomously adjusts the sampling frequency, range, and sensitivity parameters of the acoustic, optical, and chemical sensors according to the environmental characteristics of the area (water depth, terrain, and light) to optimize the data collection strategy;

[0077] The node uses a spatiotemporal registration algorithm, based on the method of cross-correlation or mutual information, to unify the heterogeneous data collected by different sensors into the same spatiotemporal reference, and uses Kalman filtering and particle filtering multi-sensor data fusion technology to extract multi-dimensional feature information of the target and environment;

[0078] The nodes transmit the fused data to the leader node through the software-defined underwater acoustic network (SDUAN). SDUAN adopts a distributed architecture. The nodes adaptively adjust the physical layer and link layer parameters according to the channel status, data priority and energy budget factors, and dynamically configure the transmission power, modulation mode and coding strategy to ensure the reliability and real-time performance of data transmission. At the same time, the network layer adopts the radial selective forwarding (VBF) routing protocol based on geographical location. According to the relative position of the relay node and the leader node, the remaining energy and the communication load index, the best relay path is selected to extend the network life cycle.

[0079] After the leader node obtains the data reported by each node through the cognitive relay link, it adopts a consensus-based multi-agent reinforcement learning framework to autonomously learn and optimize the division of labor and cooperation strategy of multiple robot nodes. This framework combines the ideas of game theory and evolutionary computation, and regards each node as an intelligent agent that makes autonomous decisions. It continuously updates its own strategy through local observation and reward feedback. The global strategy is formed through strategy evaluation, strategy sharing and game equilibrium convergence among intelligent agents.

[0080] The state space of reinforcement learning includes the position, speed, and energy state information of each node, and the action space includes target allocation, path planning, and obstacle avoidance decision control instructions. To improve fault tolerance, the leader node adopts an active replication backup mechanism, which implements distributed consistency based on the Paxos protocol. The leader node periodically synchronizes state snapshots to the backup node and ensures the atomicity of the backup through a two-phase commit protocol. The backup node monitors the heartbeat signal of the leader node. Once it times out for multiple consecutive times, it triggers a view switch. The backup node elects a new leader through the Raft protocol and broadcasts identity change notifications to the cluster, rebuilding the TOPO relationship to ensure the continuity of data processing and task scheduling.

[0081] In terms of environmental change modeling, the leader node uses the Gaussian process regression (GPR) model to map the spatiotemporal discrete sampling data (temperature, salinity, light intensity) reported by the node to the continuous environmental parameter distribution. GPR describes the spatial correlation between different sampling points through the kernel function and uses Bayesian inference to estimate the posterior distribution of each location. The model training process uses the stochastic gradient descent optimization algorithm to minimize the negative log-likelihood loss function.

[0082] On this basis, the leader node further adopts the Bayesian optimization algorithm to adaptively adjust the sampling area and frequency of the node through mean-variance estimation and PI acquisition function to minimize energy consumption. If the leader node fails, the backup node with excellent waterproof performance and the highest fault tolerance level will take over its role and continue to perform the above tasks to ensure the continuity of data processing and task scheduling;

[0083] In terms of target 3D reconstruction, the leader node adopts the distributed joint multi-view stereo vision (DI-MVS) algorithm to globally stitch the local point cloud streams sent back by the nodes. The algorithm first uses the visual odometry to estimate the motion trajectory of the node, and then uses the 3D-NDT registration algorithm to achieve coarse alignment of local point clouds from different perspectives. The leader node initiates multiple rounds of pose graph optimization, uses the g2o framework to build an error function, and performs least squares solution on the node pose and target shape parameters to obtain a globally consistent dense 3D model. The algorithm incrementally updates the semantic label and texture mapping of the target according to changes in environmental parameters to improve the accuracy and granularity of perception. At the same time, the backup node records the changes in water flow and temperature in real time, and captures the impact of environmental changes on the target's fine 3D model through the Gaussian process regression model. An incremental learning strategy is adopted to use the newly collected data of the node to perform online fine-tuning on various models, and a long short-term memory network time series model is used to capture the dynamic changes of the environment. Active learning methods are used to select high-confidence samples for retraining.

[0084] Specifically, in S104, the robot node determines the sub-area to which it belongs based on the Thiessen polygon division result. A certain node finds that it is located in an area with a water depth of 30 meters and a light intensity of 500 lux by calculating the Voronoi diagram with adjacent nodes. Therefore, the sampling frequency of the sonar is adjusted to 10 Hz, the exposure time of the CCD camera is set to 50 ms, and the range of the pH sensor is set to 0-14.

[0085] The node uses a cross-correlation algorithm to perform spatiotemporal registration of different sensor data, and performs sliding window calculations in units of 1 second. The cross-correlation coefficient threshold is set to 8. The registered data is fused through a multimodal Kalman filter. The state vector includes six dimensions: target position, velocity, and acceleration. The observation vector includes 10 dimensions: image features, acoustic Doppler shift, and chemical signal intensity.

[0086] The EM algorithm is used for parameter estimation, and the confidence threshold is set to 95%. The fused data is transmitted to the leader node through SDUAN. The network adopts OFDM modulation, with 1024 subcarriers, 100us symbol period, and 10mW transmission power. The VBF routing protocol calculates the best relay path based on the distance between the relay node and the leader node, the remaining energy, and the data cache factor. The energy weight coefficient is 5, the distance weight coefficient is 3, and the load weight coefficient is 2.

[0087] The leader node adopts a consensus-based multi-agent reinforcement learning framework, treating each node as an independent Actor-Critic agent. The state space includes 10 dimensions: node position, speed, and energy. The action space includes 5 dimensions: target allocation, path planning, and obstacle avoidance decision. The reward function comprehensively considers target coverage, energy consumption, and safety distance factors. The learning rate is 0.1 and the discount factor is 9.

[0088] The softmax policy gradient algorithm is used for policy optimization, with 1000 iterations. To improve fault tolerance, the leader node synchronizes the state snapshot to the three backup nodes every 10 seconds. The two-phase commit protocol is used, with a timeout of 5 seconds. The backup nodes elect new leaders through the Raft protocol, with a heartbeat timeout of 3 seconds and a random timeout of 150-300ms.

[0089] In terms of environmental modeling, the leader node uses the Gaussian process regression model, the kernel function uses the square exponential kernel, the hyperparameters are obtained through maximum likelihood estimation, the number of training samples is 1000, the batch size is 100, the learning rate is 0.05, the Bayesian optimization algorithm uses the PI acquisition function, the confidence interval is 95%, the initial random sampling points are 10, and the maximum number of iterations is 50;

[0090] In terms of target three-dimensional reconstruction, the leader node adopts a distributed joint multi-view stereo vision algorithm and uses visual odometry to estimate the node's motion trajectory. The 3D-NDT registration algorithm is used to achieve coarse alignment of local point clouds from different perspectives. The g2o framework is used to construct an error function, and the least squares solution of the node pose and target shape parameters is performed to obtain a globally consistent dense three-dimensional model. The algorithm incrementally updates the target's semantic label and texture mapping according to changes in environmental parameters.

[0091] In S105, during the operation, if a robot node fails, the fault type and severity are determined by a multi-classification fault diagnosis model. For minor faults, the task priority of the node is reduced, and it is assigned to continue to perform low-priority tasks. For serious faults, the node sends a help signal to the neighboring nodes, requesting the neighboring nodes to take over the task, and at the same time starts the emergency response program, broadcasts the fault information through the underwater acoustic communication link, updates the network topology, and uses a dynamic path reconstruction algorithm to re-plan the location and path of the task node;

[0092] Based on the robot status data collected by multi-source heterogeneous sensors, voltage, current, temperature, and vibration characteristic parameters are extracted, and a multi-classification fault diagnosis model is constructed using a convolutional neural network. Through data enhancement and transfer learning methods, training is performed on a small-scale labeled data set, and adaptive sampling and active learning strategies are used to continuously expand the fault sample pool and improve the generalization performance of the model, realizing the diagnosis of typical fault types such as mechanical faults, electrical faults, and communication faults.

[0093] According to the fault diagnosis results and the fault grade classification standard in the knowledge base, the fuzzy comprehensive evaluation method is used to evaluate the severity of the fault and obtain the membership vector of the fault grade;

[0094] For minor faults whose severity is lower than the minimum threshold, the node's computing and communication resources are reduced, and the priority of its task queue is dynamically adjusted through the Quality of Service (QoS) scheduling mechanism to postpone the execution of tasks with lower real-time requirements;

[0095] For serious faults with a severity membership higher than the highest threshold, the alarm log is written to the fault information queue in a timely manner, and a help signal is sent to the relevant nodes through the message subscription and publishing mechanism. The SOS message is encapsulated in the XML-based Fault Information Description Language (FIDL), and the content includes the fault type ID, fault occurrence time, fault node ID, node current location and energy status;

[0096] The security of communication is ensured through message encryption and digital signature. After receiving the SOS message, the neighboring node evaluates whether to take over the task based on the preset fault handling strategy, its own capabilities and status. If it decides to take over, it migrates the task and synchronizes the status with the faulty node, and feeds back a confirmation message.

[0097] After the faulty node starts the emergency response procedure, it uses the software-defined underwater acoustic network (SDUAN) technology to broadcast the fault information, issue a new routing table and channel allocation strategy through the controller, dynamically adjust the network topology, isolate the faulty node, and restore the data transmission path. At the same time, the Value Network algorithm in reinforcement learning is used to optimize the node's energy consumption, task completion rate and survival time. The optimal emergency strategy is learned through the policy gradient method to guide the node's motion control and task replanning.

[0098] When reconstructing the network topology, the objective function is to maximize the network throughput, and the node connectivity requirements are used as constraints. An integer programming model is constructed to solve the optimal deployment plan. A distributed solution framework is used to decompose the global problem into local sub-problems of multiple nodes. Global convergence is achieved through iterative optimization. After the optimal deployment plan is obtained, each node adjusts its position according to the plan, and adopts a formation control algorithm based on artificial potential fields to form a stable formation while avoiding obstacles and collisions. The leading node comprehensively considers the fault diagnosis results, task importance, and resource constraints, and constructs a Markov decision process (MDP) model to replan the task. The state space includes the robot position, speed, energy, and fault state, and the action space includes task allocation, path planning, and obstacle avoidance decisions. The Gaussian process reinforcement learning (GPEL) algorithm is used to solve the optimal decision sequence, generate a task replanning plan after the fault, and send it to each node for execution.

[0099] During the task execution process, the node status is monitored in real time, and the status data is mapped to the spatiotemporal database through the consistent hashing algorithm to build a global view. If it is found that the node movement trajectory deviates from the predetermined path or the energy efficiency is abnormal, the local path correction and energy consumption optimization program are triggered in time. When correcting the path, the asynchronous advantage Actor-Critic (A3C) algorithm in deep reinforcement learning is used through interactive learning between the policy network and the value network.

[0100] The optimal trajectory is planned under the premise of meeting the task constraints. When optimizing energy consumption, the Sequence to Sequence (Seq2Seq) model is applied, combined with the attention mechanism to predict the energy consumption distribution of the node. The optimal working mode of the node is solved through the multi-objective evolutionary algorithm to balance energy consumption and task completion rate. Each node stores the execution log in the distributed ledger through blockchain technology to ensure that the data cannot be tampered with.

[0101] The leading node uses the federated learning framework to regularly trigger each node to conduct local training, protects model parameters through differential privacy and homomorphic encryption technology, aggregates learning results using multi-party secure computing protocols, and iteratively updates the global fault prevention and emergency response model. At the same time, the causal inference graph neural network (CIGNN) is used to mine causal dependencies from fault diagnosis, decision-making and disposal data, and the active learning strategy is combined to optimize the training samples, improve the comprehensive coverage of the fault knowledge graph, and form an explainable and transferable policy knowledge base to guide policy search and optimization in unknown scenarios.

[0102] Specifically, in S105, according to the robot state data collected by the multi-source heterogeneous sensors, voltage, current, temperature, and vibration characteristic parameters are extracted. For voltage data, the sampling frequency is set to 100 Hz, and each data contains 100 sampling points, and its mean, variance, and peak statistical features are extracted;

[0103] For vibration data, the sampling frequency is set to 1kHz, and each data contains 1000 sampling points. Its frequency domain features, including spectrum entropy and frequency band energy ratio, are extracted. A multi-classification fault diagnosis model is constructed using a convolutional neural network. The input is the extracted feature vector, and the output is the probability distribution of the fault type. The network structure contains 3 convolutional layers and 2 fully connected layers. The convolution kernel sizes are 3x3, 4x4, and 5x5, respectively. The number of convolution kernels is 32, 64, and 128, respectively. The ReLU activation function and maximum pooling operation are used. The number of neurons in the fully connected layer is 256 and 10, respectively. The ReLU and Softmax activation functions are used. The training set is expanded through data enhancement, including rotation and translation operations, and the ImageNet pre-trained model is used for transfer learning to improve the model performance.

[0104] During the training process, the batch size is dynamically adjusted using an adaptive sampling strategy. When the rate of decrease of the loss function value of the training set is less than 0.1%, active learning is triggered. The 10% samples with the largest information entropy are selected from the unlabeled sample pool for manual labeling and added to the training set for continued training until the accuracy of the model on the validation set reaches more than 95%;

[0105] According to the fault diagnosis results and the fault level classification standard in the knowledge base, the fuzzy comprehensive evaluation method is used to evaluate the severity of the fault. First, a fuzzy set containing three levels of slight, medium and severe is established, and the membership function of each level is set. For temperature abnormality faults, the membership function of slight faults is defined as: when x < 60, μ (x) = 1; when 60 ≤ x ≤ 70, μ (x) = (70-x) / 10; when x > 70, μ (x) = 0;

[0106] In the formula, x is the temperature value, and the membership function of medium and severe faults is defined. Then, the comprehensive membership is calculated using the weighted average method. The weight is determined according to the statistical results of the data, and the membership vector of the fault level is obtained.

[0107] For minor faults with a severity level lower than 3, the node's computing and communication resources are reduced, the CPU frequency is reduced by 20%, the communication power is reduced by 1dB, and the priority of its task queue is dynamically adjusted through the Quality of Service (QoS) scheduling mechanism, the priority of the data preprocessing task is reduced by 10%, and the execution of tasks with lower real-time requirements is postponed;

[0108] For serious faults with a severity membership higher than 7, the alarm log is written to the fault information queue in a timely manner, and a distress signal is sent to the relevant nodes through the message subscription and publishing mechanism. The SOS message is encapsulated in the XML-based Fault Information Description Language (FIDL) and encrypted using the AES-256 encryption algorithm, and the digital signature is generated using the SHA-256 algorithm.

[0109] In S106, during the multi-robot collaborative operation, each node estimates and models the probability distribution of the target by combining Bayesian reasoning and evidence theory, and uses transfer learning and incremental learning methods to update and optimize the target recognition model in real time. The robot in the deep water area adjusts the output power of the propeller, and the robot in the shallow water area improves the propeller efficiency and data transmission frequency.

[0110] Each robot node collects target data through vision, sonar, temperature, salinity and depth multimodal sensors, extracts color, texture and shape features, and uses the Bayesian network to build a joint probability distribution model of the target. The network structure adopts a directed acyclic graph, where nodes represent the attribute variables of the target and edges represent the conditional dependency between attributes. The expectation maximization algorithm is used to optimize the model parameters online and the network structure is adaptively adjusted. The reasoning time complexity is reduced to the logarithmic level, and the nodes use the Markov chain Monte Carlo sampling algorithm to approximate the posterior probability distribution of the new target to improve sampling efficiency.

[0111] In distributed target perception, nodes use the Dempster-Shafer theory to fuse heterogeneous evidence. The evidence format is represented by a mass distribution function and supports interval probability description. Before fusion, the quality of evidence is evaluated using a projection tracking algorithm to remove evidence sources with low quality. Then, a new evidence similarity metric is preset based on the manifold structure, and the Murphy and Dempster combination rules are adaptively selected for fusion. The Gini uncertainty index is used to achieve conflict detection and credibility assessment of fusion results. Nodes use deep transfer learning and incremental learning methods to achieve rapid modeling of target recognition models. MobileNet-SSD is selected as the backbone network for target detection and pre-trained on ImageNet. Then, the knowledge of the teacher network is transferred to the student network of the node through knowledge distillation. Finally, it is fine-tuned on a small-scale labeled sample. The ESDM edge sample selection strategy is used in the fine-tuning process. The samples with the greatest value to the model are selected based on the sample gradient norm and density estimation to construct increments.

[0112] When the target state drifts over time, an active learning method is used to generate a pool of samples to be labeled using an offline strategy. The difficult samples that can improve the model the most are selected online using the criterion of maximizing the target probability entropy. The model is then incrementally updated using knowledge distillation and the FRCL forgetting control strategy to improve overall detection accuracy.

[0113] At different underwater operating depths, the node needs to comprehensively consider multiple performance requirements such as detection distance, energy consumption, and motion efficiency. Based on the data of MEMS depth sensors and hyperbolic sonar, the sensor fusion is realized by using Kalman filtering, and the uncertainty of depth measurement is estimated to be less than 1%.

[0114] On this basis, a multi-objective optimization model is established. The objective function is the weighted detection distance and the inverse of energy consumption. The constraint condition is the feasible domain of the controller. A multi-strategy co-evolutionary algorithm is used to search for the optimal control sequence, reducing the solution time to less than 1s.

[0115] Specifically, the propeller optimization variables are propeller diameter, pitch, and speed, which are searched in the discrete domain by the adaptive differential evolution algorithm. The motion control optimization variables are speed, heading, and depth, which are searched in the continuous domain by the MOEA / D algorithm. The two algorithms are iterated alternately and converge quickly to the Pareto frontier.

[0116] In deep water areas, we focus on low power consumption and long-distance detection, use MEMS depth sensors as the main depth measurement method, reduce sonar pulse frequency, reduce propeller speed, increase pitch, and improve overall performance;

[0117] In shallow waters, the focus is on rapid maneuvering and close-range detection, using multi-beam forward-looking sonar as the primary obstacle avoidance method, increasing the frequency, dynamically optimizing the propeller speed and pitch, and reducing trajectory tracking errors;

[0118] In the target search task, multiple heterogeneous nodes adopt a distributed collaborative perception and planning method based on reinforcement learning. The nodes decompose the task into search, classification, and tracking atomic operations according to the heterogeneous constraints of detection range and motion ability. The abstraction is a Markov decision process. The state is the node position and target probability map. The action is local path planning and allocation atomic operations. The reward is the number of targets found and the trajectory cost. The local strategy is learned through the A3C algorithm. The collaborative layer uses a graph convolutional neural network to encode the global target probability map. The graph attention network dynamically aggregates local strategies. The MADDPG algorithm is used to achieve multi-agent joint strategy optimization, and the overall reward is increased by 20%;

[0119] For heterogeneous node formations, a virtual leader-follower architecture is used to decompose motion constraints. The leader plans a virtual reference trajectory based on the global task and uses the sampling points on the trajectory as the intermediate targets of the follower nodes. The follower nodes optimize the speed and heading of the intermediate targets based on their own heterogeneous characteristics to ensure that nodes with different motion performances arrive synchronously under the virtual structure. The formation control rate uses a consistent Hamiltonian, and the two-way ring communication topology ensures the robustness of data interaction. The controller uses an improved sliding mode control law to reduce the convergence time.

[0120] In view of the low bandwidth and long delay characteristics of the underwater acoustic channel, software-defined underwater acoustic network technology is used between nodes to coordinately optimize the physical layer, link layer, and network layer. The physical layer adopts FH-MFSK modulation, and the frequency hopping rate and number of orthogonal tones are adaptive to the underwater acoustic channel status. The link layer uses the Sliding-ALOHA protocol to dynamically allocate time slots to ensure a low collision rate for data burst transmission. The network layer uses the Q-Learning method to learn routing strategies based on end-to-end delay and available bandwidth to improve average throughput.

[0121] A simulation platform for multi-robot autonomous collaborative system is implemented on the ROS robot operating system, and an experimental platform composed of heterogeneous surface or underwater robots is built.

[0122] Specifically, each robot node described in S106 collects target data through vision, sonar, temperature, salt and depth multimodal sensors, uses a high-resolution camera to obtain target images, extracts color histograms and SIFT features, and uses a support vector machine to classify textures, with an average accuracy of more than 95%; estimates the three-dimensional shape of the target through binocular stereo vision, with a point cloud density of more than 1,000 per square meter, and the node uses a Bayesian network to build a joint probability distribution model of the target. The network structure uses a directed acyclic graph, including 15 attribute nodes of color, texture, and shape. The dependency structure between attributes is learned through the maximum weight spanning tree algorithm, and the expectation maximization algorithm is used to optimize the model parameters online. The average number of iterations is less than 20 times. A pruning strategy is used during reasoning to remove nodes with edge probabilities lower than 0.1, and the time complexity is reduced to a logarithmic level;

[0123] For new targets, the node generates 20,000 samples using the Metropolis-Hastings sampling algorithm, and makes an approximate inference on the posterior probability distribution, with an average acceptance rate of more than 60%. In distributed target perception, the node first uses the projection tracking algorithm to calculate the quality of evidence, removes evidence sources with a confidence level lower than 6, and then uses the geodesic distance in the manifold space to measure the similarity of evidence. The fuzzy C-means clustering algorithm is used to divide the evidence into three categories, and the Murphy, Dempster and Yager combination rules are used for fusion. Finally, the Gini uncertainty index is used to evaluate the credibility of the fusion result, which is considered credible when the index is less than 3. The node uses MobileNet-SSD as the backbone network for target detection, first pre-trains on ImageNet, then uses KL divergence as the distillation loss function, transfers the knowledge of the teacher network to the student network, and finally fine-tunes on 500 small samples, with an overall mAP of more than 90%. The ESDM method is used for incremental learning. Based on the sample gradient norm and local density estimation, the 20% difficult samples that have the greatest improvement on the model are adaptively selected to construct increments, which improves the detection accuracy by 5 percentage points on average.

[0124] For targets with state drift, the node uses an active learning method to generate 1,000 samples to be labeled offline, selects 30 difficult samples online using the criterion of maximizing the target probability entropy, and incrementally updates the model using a forgetting control strategy, which improves the detection accuracy by 8 percentage points on average.

[0125] In S107, after the operation task is completed, all data are transmitted to the data processing center on the mother ship. The data processing center uses big data analysis and virtual reality technology to comprehensively process and visualize the collected data, generate a panoramic three-dimensional model and virtual reality scene of the area, and use the water flow and temperature data collected by the robot to create a dynamic environment model, and intuitively display the environment and conditions of underwater operations through virtual reality technology;

[0126] The data processing center on the mother ship receives massive amounts of heterogeneous data from the robot nodes, including multi-beam sonar data, side-scan sonar data, pressure sensor data, and temperature, salinity, and depth sensor data. The total amount of data reaches TB level. Hadoop Distributed File System (HDFS) and HBase column-based database are used to perform distributed storage and management of heterogeneous data, supporting efficient data retrieval and aggregate query.

[0127] Using the Flink stream computing framework, we modeled data dependencies through directed acyclic graphs (DAGs) and implemented real-time monitoring and response to underwater targets and environments based on event-driven real-time computing tasks. The end-to-end latency was controlled at the millisecond level. At the same time, we used the Spark memory computing framework to perform batch processing and analysis on historical data, extract the spatiotemporal evolution pattern of the underwater environment, and optimize the autonomous decision-making model of the underwater robot.

[0128] For sonar point cloud data, adaptive voxelization and random downsampling algorithms are first used to filter and reduce noise and compress the original point cloud, and an octree index is constructed to accelerate subsequent processing. Then, the regional growing criterion is constructed through normal vector consistency, curvature similarity, and feature entropy multi-scale features, and the merging threshold is adaptively adjusted to achieve robust point cloud segmentation.

[0129] On this basis, HOG and SHOT 3D shape descriptors are extracted, and the support vector machine (SVM) and random forest classifier are combined to semantically annotate the seabed features. The Markov random field (MRF) is used to smooth and optimize the classification results to eliminate noise interference. Finally, the Poisson reconstruction algorithm is used to reconstruct the gridded surface of the point cloud blocks to generate a high-precision, semantically rich 3D model.

[0130] Considering the complementary characteristics of sonar data and image data, a loosely coupled registration method of multi-view images and point clouds based on graph optimization is introduced. By minimizing the reprojection error and distance error, accurate registration of different modal data is achieved to generate a more complete and accurate panoramic 3D scene model.

[0131] For the water flow and temperature physical parameter data, the ensemble Kalman filter algorithm is used for multi-sensor fusion. By adaptively adjusting the noise covariance matrix, the state changes of the underwater environment are dynamically tracked. In terms of data interpolation, the Kriging interpolation algorithm is used to build a spatial correlation model of environmental parameters. According to the spatial distribution characteristics of the sampling points, the variation function is adaptively fitted, and anisotropy factors are introduced to improve the interpolation accuracy. At the same time, numerical simulation is combined with machine learning. By coupling the Reynolds average Navier-Stokes (RANS) equation and the long short-term memory network (LSTM), the refined physical modeling and data-driven parameter prediction of the underwater environment are realized, and the physical processes of water flow and temperature in a dynamic environment are more accurately described.

[0132] In the process of environmental modeling, we fully consider the FAIR data management principles, adopt the standard netCDF and HDF5 scientific data formats, standardize the annotation and semantic description of metadata, and facilitate the integration and sharing of multi-source heterogeneous data. In terms of virtual reality scene construction, in response to the rendering challenges of large-scale marine environments, we use a rendering pipeline based on real-time ray tracing. By introducing screen space reflection (SSR) and volume ray tracing (VXGI) acceleration algorithms, we can ensure an interactive frame rate of more than 30 frames per second while meeting the requirements of high-realistic rendering. At the same time, we also introduce a physically based ocean rendering model, taking into account the absorption, scattering, and refraction optical properties of seawater, and use high-resolution skyboxes and dynamic environment maps to simulate realistic sea and sky and sparkling effects. In order to improve the user interaction experience, we integrate multi-modal interaction methods such as facial expression recognition, gesture recognition, and voice recognition, and use force feedback devices to simulate the resistance and touch of underwater operations to enhance the sense of reality.

[0133] In addition, gaze tracking technology is used to dynamically adjust the rendering strategy and viewpoint according to the user's line of sight and focus area, achieving a smoother and smarter roaming experience;

[0134] In terms of autonomous decision-making of underwater robots, a method combining deep reinforcement learning and transfer learning is used to optimize the control strategy from an end-to-end perspective. The global map information is encoded through a graph convolutional neural network (GCN), and the local sensor data is modeled in time series using a gated recurrent unit (GRU). The Transformer structure is used to achieve global-local feature fusion and output the optimal target detection and path planning strategy.

[0135] During the training process, imitation learning is used to initialize the policy network, and a meta-learning algorithm is used to achieve cross-scenario and cross-task policy migration and rapid adaptation, thereby improving the robot's environmental perception and autonomous decision-making capabilities. At the same time, adversarial training and model uncertainty estimation techniques are introduced to actively explore noisy data and improve the generalization ability of the strategy.

[0136] Specifically, the data processing center on the mother ship described in S107 uses Hadoop and HBase to build a distributed data storage system, divides the unstructured data into 64MB data blocks, uses a 3-copy mechanism to achieve fault tolerance, and uses Zookeeper to implement distributed coordination services, supporting elastic expansion of thousands of nodes;

[0137] Flink processes underwater target detection tasks in real time, uses RocksDB as the StateBackend to save intermediate states, sets up three concurrent TaskManager nodes, each equipped with 8GB of memory and four CPU cores, and uses the EventTime mechanism to process out-of-order data, achieving millisecond-level end-to-end latency.

[0138] Spark batch processes the spatiotemporal prediction task of the marine environment, uses HDFS to store the training data set, uses the stochastic gradient descent algorithm for model optimization, sets a learning rate of 0.01 and a momentum factor of 0.9, uses F1-score as the evaluation indicator, and after 10 iterations, the model prediction accuracy reaches 95%. The point cloud data uses the octree and kdtree spatial index method, with a compression rate of 80%, and the retrieval response time is reduced to less than 50ms;

[0139] When segmenting point clouds, the weighted Euclidean distance of normal vector and curvature is used to measure the similarity of local features, the weight ratio is adaptively adjusted, and the seed point strategy of regional growing is used to merge supervoxels, and the segmentation accuracy reaches 92%;

[0140] During semantic annotation, the FPFH and VFH feature descriptors of each point were extracted, and a random forest classifier containing 1000 decision trees was constructed. The model parameters were optimized through grid search, and the classification results were Markov optimized using the graph cut algorithm. The F1 value reached 0.9. During Poisson reconstruction, the conjugate gradient descent method was used to solve the Poisson equation, and the maximum depth of the Oct-tree was set to 8. The Hausdorff distance between the generated mesh surface and the real model was less than 1 cm. During the registration of sonar and image data, SIFT feature point matching was used to initialize the rigid body transformation matrix, and then the pose parameters were optimized through ICP iteration. The average registration error was controlled within 5 mm.

[0141] When the underwater environment state is integrated, the volumetric Kalman filter algorithm is used to approximate the posterior probability distribution of temperature and salinity parameters, and the covariance matrix of system noise and observation noise is adaptively adjusted to achieve the optimal estimation of nonlinear and non-Gaussian distribution.

[0142] When interpolating the underwater environment, the fractal analysis method is used to fit the variation function model of the temperature field and flow field, and the optimal range and base value are obtained through cross-validation. The interpolation variance is less than 5%. When coupling physics and machine learning modeling, the k-ε turbulence model is used to solve the Reynolds averaged NS equation to generate a flow field data set, which is then used to train the long short-term memory neural network to learn the time correlation of the water flow and realize the flow velocity prediction within the next 24 hours, with an average error of less than 10%;

[0143] When rendering the ocean environment in virtual reality, 3D textures are used to represent volume scattering media, and realistic rendering of clouds and smoke is achieved through light stepping and pre-calculated phase functions, with a frame rate of 90fps. When force feedback is interactive, the force of the fluid on the manipulator is simulated by solving the particle spring system, and 5 degrees of freedom force feedback is provided in conjunction with the tactile device, with a peak value of up to 16N;

[0144] When multimodal perception is fused, FasterR-CNN is used to detect the target area in the sonar image, and a 512-dimensional feature vector is extracted through the backbone network. At the same time, the inertial navigation information is sent to the GRU to learn motion compensation, realizing end-to-end learning of target detection, tracking and positioning, with a target positioning accuracy of better than 1m.

[0145] When the underwater robot makes autonomous decisions, the DDPG algorithm is used offline to train the control strategy network. The state space includes the relative position of the target and the 18-dimensional vector of its own posture. The action space includes the speed and yaw angular velocity in the three directions of x / y / z. The reward function comprehensively considers the shortest path length and energy consumption factors, and uses data enhancement and strategy distillation methods to alleviate the overfitting problem. After the strategy network converges, the average path length is shortened by 25%;

[0146] When reconstructing underwater ruins in 3D, the moving least squares method is used to estimate the local implicit surface from the point cloud data, and then the isosurface is extracted by the Bloomenthal polygon algorithm to generate a high-precision mesh model. The Poisson reconstruction method is used to refine the mesh topology structure, eliminate holes and noise, and the reconstruction accuracy is better than 5mm.

[0147] When displaying archaeological data in digital form, the PBR physical rendering model is used to generate high-fidelity material maps of the shipwreck site. Volume rendering technology is used to simulate the underwater lighting environment, and particle special effects are introduced to render suspended objects and bubbles. The rendering frame rate is stabilized at above 60fps.

[0148] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0149] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present invention shall be included in the protection scope of the present invention.

Claims

1. A fault-tolerant method for improving the robustness of underwater robot networking, characterized in that: S101, using sonar and underwater cameras to collect target feature information, the target feature information includes type, size and location, and at the same time obtain underwater environmental parameters such as water temperature, water flow and visibility, and generate the optimal networking scheme of the role division and spatial position distribution of robot nodes through the improved ant colony algorithm according to the energy state, waterproof performance, data cache and computing load of the heterogeneous underwater robots, combined with the depth of underwater operation and regional terrain, and comprehensively consider the performance parameters of each robot node to determine the fault tolerance level of each node at different underwater operation depths; S102, when a robot node detects a new target, it transmits the target feature information to the leader node through the underwater acoustic communication link. The leader node evaluates the value of the target and the urgency of the task based on historical data, knowledge base and underwater environmental parameters, and triggers the collaborative operation mechanism; S103. In collaborative work, a dynamic partitioning collaborative strategy based on Thiessen polygons is adopted to divide the area into several sub-areas; Corresponding robots are assigned to high water flow speed areas and dense obstacle areas. The formation and obstacle avoidance behavior of the robot cluster are determined by the artificial potential field method. The path planning algorithm is used to adjust the robot's motion trajectory in real time, and the position is dynamically adjusted according to the changes in water flow. S104, each robot node collects data in its assigned sub-area and transmits it to the leader node in real time through the underwater acoustic communication link. The leader node initially integrates and processes the received data, uses the distributed joint Kalman filter algorithm and iterative closest point algorithm to generate a three-dimensional model of the target, dynamically adjusts the three-dimensional model parameters, and records the impact of water flow and temperature changes on the environment; S105. During the operation, if a robot node fails, the fault type and severity are determined through a multi-classification fault diagnosis model. For minor faults, the task priority of the node is reduced. For serious faults, the neighboring node is requested to take over the task. The fault information is broadcast through the underwater acoustic communication link to update the network topology. The dynamic path reconstruction algorithm is used to re-plan the location and path of the task node. S106. In multi-robot collaborative operations, each node estimates and models the probability distribution of the target through a combination of Bayesian reasoning and evidence theory, updates and optimizes the target recognition model in real time, and adjusts the robot parameters based on the data characteristics of different underwater operation depths. The robot parameters include the output power of the propeller, the propeller efficiency, and the data transmission frequency; S107. After the operation task is completed, all data are transmitted to the data processing center on the mother ship. The data processing center uses the water flow and temperature data collected by the robot to create a dynamic environmental model to display the environment and conditions of underwater operations.

2. The fault-tolerant method for improving the robustness of underwater robot networking according to claim 1, characterized in that: In the process of underwater target and environment data modeling, the S101 adopts an octree space partitioning data structure to recursively partition the three-dimensional space, and maps high-resolution sonar point cloud data and visual image features into a spatial grid. In the process of task decomposition and allocation, a task ontology description method based on a semantic network is adopted to divide the task into multiple atomic tasks, and the hierarchical and logical relationships between tasks are expressed in the ontology language OWL, and task allocation is performed through a contract network protocol.

3. The fault-tolerant method for improving the robustness of underwater robot networking according to claim 1, characterized in that: In S102, when the robot node detects a new target through sonar and underwater camera, it first extracts characteristic information of the target's type, size, and position. The characteristic information includes identifying the target as a torpedo through image segmentation algorithm, which is preset to be 5 meters long and 5 meters in diameter and located 50 meters in front of the node. The node uses SLAM algorithm combined with IMU and DVL sensor data to preset itself to be located at 35 degrees north latitude, 135 degrees east longitude, 50 meters underwater, with a pitch angle of 5 degrees and a yaw angle of 30 degrees. According to the relative position of the target to the node and the spatial position of the node itself, the absolute three-dimensional coordinates of the target are calculated to be 100,8 0, -40, and calculates the target's movement speed as 2m / s and heading angle as 60 degrees through the Kalman filter algorithm. The node encapsulates the target information in JSON format, encrypts it with AES-256, and transmits it to the leader node through the underwater acoustic communication link. After receiving the target information, the leader node extracts key features and performs similarity matching with the historical target records in the MongoDB database. The support vector machine SVM algorithm is used to identify the target as a training torpedo, and the improved whale optimization algorithm WOA is used for solution. The target tracking task is sent to the underwater glider node, and the salvage task is sent to the ROV node.

4. The fault-tolerant method for improving the robustness of underwater robot networking according to claim 1, characterized in that: In S103, the underwater robot cluster obtains node performance parameters, node positions and underwater environment parameter information through acoustic communication, sets the sampling frequency to 10 Hz, uses the sequential Gaussian simulation algorithm to perform three-dimensional space interpolation on 1000 sampling points, and a single iteration takes less than 1 second to generate a three-dimensional environment grid model with a resolution of 1 meter. The Thiessen polygon algorithm is used to divide the operation area, and the octree data structure is used to store the division results. The retrieval time complexity is Ologn; In areas with dense obstacles, the artificial potential field method and the RRT algorithm are combined to plan obstacle avoidance paths. The attraction coefficient and repulsion coefficient of the artificial potential field method are 0.8 and 1.2 respectively. The number of sampling points of the RRT algorithm is 1000, the shortest obstacle avoidance path length is shortened by 20%, and the parameters of the adaptive PID controller are adjusted online according to fuzzy rules with an adjustment period of 1 s; In a dynamic water flow environment, the ADCP device collects 128 layers of water flow profile data at a frequency of 5 Hz, uses Kriging interpolation to construct a three-dimensional water flow field model, and uses the ensemble Kalman filter algorithm for real-time prediction. The formation control adopts a graph-based multi-agent consensus algorithm to optimize the spatial position of the 4×4 grid formation.

5. The fault-tolerant method for improving the robustness of underwater robot networking according to claim 1, characterized in that: In S104, the robot node determines the sub-area to which it belongs according to the result of the Thiessen polygon division, and transmits the fused data to the leader node through SDUAN. The network adopts OFDM modulation, the number of subcarriers is 1024, the symbol period is 100us, and the transmission power is 10mW. The VBF routing protocol calculates the relay path according to the distance between the relay node and the leader node, the remaining energy and the data cache factor, with an energy weight coefficient of 5, a distance weight coefficient of 3, and a load weight coefficient of 2; The leader node adopts a consensus-based multi-agent reinforcement learning framework, sets each node as an independent Actor-Critic agent, synchronizes the state snapshot to three backup nodes every 10 seconds, adopts a two-phase commit protocol, and the timeout is set to 5 seconds. The backup nodes elect new leaders through the Raft protocol, the heartbeat timeout is 3 seconds, and the random timeout is 150-300ms.

6. The fault-tolerant method for improving the robustness of underwater robot networking according to claim 1, characterized in that: In S105, a fuzzy comprehensive evaluation method is used to evaluate the severity of the fault. First, a fuzzy set of three levels, namely, slight, medium, and severe, is established, and a membership function of each level is set. The membership function includes a membership function for a slight fault defined as follows for an abnormal temperature fault: When x<60, μ(x)=1; When 60≤x≤70, μ(x)=(70-x) / 10; When x>70, μ(x)=0; In the formula, x is the temperature value, and the membership function of medium and severe faults is defined. Then, the membership is calculated using the weighted average method. The weight is determined according to the statistical results of the data, and the membership vector of the fault level is obtained. For minor faults with a severity level lower than 3, the CPU frequency is reduced by 20% and the communication power is reduced by 1dB. The priority of the task queue is dynamically adjusted through the QoS scheduling mechanism. The priority includes reducing the priority of the data preprocessing task by 10% and postponing the task execution. For serious faults with a severity level higher than 7, the alarm log is written to the fault information queue in time, and a rescue signal is sent to the node through the message subscription and publishing mechanism. The SOS message of the assistance signal is encapsulated by the XML-based fault information description language FIDL, encrypted by the AES-256 encryption algorithm, and a digital signature is generated by the SHA-256 algorithm.

7. The fault-tolerant method for improving the robustness of underwater robot networking according to claim 1, characterized in that: In the S106, each robot node collects target data through vision, sonar, temperature, salt and depth multimodal sensors, uses a camera to obtain a target image, extracts a color histogram and SIFT features, and uses a support vector machine to classify textures. The node uses a Bayesian network to build a joint probability distribution model of the target. The network structure uses a directed acyclic graph, including color, texture, and shape attribute nodes. The dependency structure between attributes is learned through a maximum weight spanning tree algorithm. The expectation maximization algorithm is used to optimize model parameters online. The average number of iterations is less than 20 times. A pruning strategy is used during reasoning to remove nodes with edge probabilities lower than 0.

1. For new targets, the node generates 20,000 samples using the Metropolis-Hastings sampling algorithm to approximate the posterior probability distribution; In distributed target perception, the node first uses the projection pursuit algorithm to calculate the quality of evidence to eliminate the evidence sources with a confidence level lower than 6. Then, the geodesic distance is preset in the manifold space to measure the similarity of evidence. The fuzzy C-means clustering algorithm is used to divide the evidence into three categories. The Murphy, Dempster and Yager combination rules are used for fusion respectively. Finally, the Gini uncertainty index is used to evaluate the credibility of the fusion result.

8. The fault-tolerant method for improving the robustness of underwater robot networking according to claim 1, characterized in that: The data processing center on the mother ship in S107 uses Hadoop and HBase to build a distributed data storage system, divides unstructured data into 64MB data blocks, adopts a 3-copy mechanism for fault tolerance, and uses Zookeeper distributed coordination service for elastic expansion; Flink processes underwater target detection tasks, uses RocksDB as StateBackend to save intermediate states, sets up three concurrent TaskManager nodes, each equipped with 8GB of memory and four CPU cores, and uses the EventTime mechanism to process out-of-order data; Spark batch processes the spatiotemporal prediction task of the marine environment, uses HDFS to store the training data set, adopts the stochastic gradient descent algorithm, sets the learning rate of 0.01 and the momentum factor of 0.9, and uses the F1-score as the evaluation indicator.

Images