Automatic homeward voyage method and system for failure of unmanned ship based on Beidou AI

Abstract

The invention discloses a Beidou AI-based unmanned ship fault automatic homeward voyage method and system, and relates to the technical field of fault automatic homeward voyage. According to the Beidou AI-based unmanned ship fault automatic return voyage method, the current three-dimensional position coordinates of a set unmanned ship, the target three-dimensional position coordinates of a preset dock wharf, water surface point cloud data and underwater point cloud data are acquired in real time, and a plurality of candidate paths are generated by combining a path planning algorithm; according to the method, the homeward voyage adaptive characteristic value of the first track point of each candidate path is analyzed in combination with an underwater point cloud environment recognition model, and homeward voyage iteration processing is carried out through the homeward voyage adaptive characteristic values until the unmanned ship reaches the preset dock wharf; therefore, the unmanned ship can autonomously select the homeward voyage path with safe obstacle avoidance and reasonable energy consumption, and the homeward voyage safety of the unmanned ship in a controllable fault state under a complex water area condition is further realized.

Description

Technical Field

[0001] This invention relates to the field of automatic return-to-base technology, specifically a method and system for automatic return-to-base of an unmanned vessel based on Beidou AI. Background Technology

[0002] With the development of intelligent shipping technology, unmanned vessels are gradually being applied to scenarios such as hydrological monitoring, environmental inspection, and emergency operations. In actual operation, unmanned vessels usually rely on battery power and achieve autonomous navigation through navigation systems and sensors. However, when unmanned vessels experience low battery or partial functional failure, most existing technologies lack a sound automatic return mechanism and usually rely on manual remote control or preset fixed return commands. Once the communication link is unstable or the unmanned vessel is in an operating area far from the dock, manual intervention is difficult to take effect in a timely manner, which can easily lead to the unmanned vessel drifting out of control due to power depletion or worsening of the malfunction, or even causing equipment loss, posing a safety hazard.

[0003] The limitations of existing technologies include at least the following issues: Existing technologies cannot guarantee that unmanned vessels can return safely and reliably in a malfunctioning state. Specifically, existing technologies lack the ability to fuse surface point clouds and underwater point clouds, making it difficult to accurately identify and avoid obstacles in the water during the return process. In the complex and ever-changing underwater environment, the return path is prone to conflict with obstacles, resulting in significant safety risks. Furthermore, the impact of the environment reflected by underwater point clouds on energy consumption is not fully considered during the return path assessment, making it difficult to effectively distinguish the energy consumption risks of different candidate paths. This can easily lead to inappropriate return path selection, making it difficult for unmanned vessels to successfully reach the pre-set dock in a malfunctioning state. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for automatic return to port of unmanned vessels based on Beidou AI malfunctions. This solves the problem that existing technologies lack point cloud fusion and path evaluation for return to port, making it difficult for unmanned vessels to return safely in malfunction situations.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for automatic return to port of a Beidou AI-based unmanned vessel in the event of a malfunction, comprising the following steps: real-time acquisition of the current three-dimensional position coordinates of the designated unmanned vessel and the target three-dimensional position coordinates of a preset dock, as well as surface point cloud data and underwater point cloud data; analysis of several candidate paths for the designated unmanned vessel based on a path planning algorithm, combined with the surface point cloud data and underwater point cloud data of the designated unmanned vessel, including the trajectory three-dimensional coordinates of several candidate trajectory points; analysis of the return-to-port adaptation feature value of the first trajectory point of each candidate path based on a pre-trained underwater point cloud environment recognition model, combined with the current three-dimensional position coordinates of the designated unmanned vessel, underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path; and iterative return-to-port processing of the designated unmanned vessel based on the return-to-port adaptation feature value of the first trajectory point of each candidate path, until the designated unmanned vessel arrives at the preset dock.

[0006] Furthermore, the specific steps for analyzing and setting several candidate paths for the unmanned vessel are as follows: input the surface point cloud data and underwater point cloud data of the unmanned vessel into a pre-trained obstacle recognition model to identify several obstacle regions of the unmanned vessel; based on each obstacle region of the unmanned vessel, analyze the target endpoint of the unmanned vessel; read the current three-dimensional position coordinates of the unmanned vessel, and based on the path planning algorithm, combined with each obstacle region and the target endpoint, analyze several candidate paths for the unmanned vessel.

[0007] Furthermore, the obstacle recognition model includes an input layer, an obstacle extraction layer, and an obstacle recognition output layer. The specific steps for identifying several obstacle regions of the designated unmanned vessel are as follows: In the input layer of the obstacle recognition model, the surface point cloud data and underwater point cloud data of the designated unmanned vessel are received and preprocessed; in the obstacle extraction layer of the obstacle recognition model, based on the preprocessed surface point cloud data and underwater point cloud data of the designated unmanned vessel, the surface obstacle point cloud set and underwater obstacle point cloud set of the designated unmanned vessel are extracted; in the obstacle recognition output layer of the obstacle recognition model, based on the surface obstacle point cloud set and underwater obstacle point cloud set of the designated unmanned vessel, several obstacle regions of the designated unmanned vessel are identified.

[0008] Furthermore, the specific steps for analyzing and setting the target endpoint of the unmanned vessel are as follows: Based on setting each obstacle area of ​​the unmanned vessel, analyze and set several candidate safe areas for the unmanned vessel; read the current three-dimensional position coordinates of the unmanned vessel and combine them with each candidate safe area to analyze and set the target endpoint of the unmanned vessel.

[0009] Furthermore, the underwater point cloud environment recognition model includes a trajectory classification subnetwork, a feature extraction subnetwork, and an output subnetwork. The trajectory classification subnetwork includes an input preprocessing layer and a classification and recognition output layer. The feature extraction subnetwork includes an input extraction layer and an extraction output layer.

[0010] Furthermore, the specific steps for analyzing the return-to-home adaptation feature value of the first trajectory point of each candidate path for the unmanned vessel are as follows: Input the current three-dimensional position coordinates of the unmanned vessel, underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path into the pre-trained underwater point cloud environment recognition model, and analyze the energy consumption risk feature set of the first trajectory point of the corresponding candidate path, including point cloud flow resistance feature value, echo energy attenuation feature value, underwater turbulence feature value, and point cloud distribution heterogeneity feature value; obtain the trajectory adaptation feature value of the first trajectory point of each candidate path for the unmanned vessel, and combine it with the energy consumption risk feature set to analyze the return-to-home adaptation feature value of the first trajectory point of the corresponding candidate path.

[0011] Furthermore, the specific steps for analyzing the energy consumption risk feature set of the first trajectory point of each candidate path of the unmanned vessel are as follows: In the trajectory classification sub-network of the underwater point cloud environment recognition model, the current three-dimensional position coordinates of the unmanned vessel, underwater point cloud data, and trajectory three-dimensional coordinates of the first trajectory point of each candidate path are received, and the underwater point cloud set of the first trajectory point of the corresponding candidate path is extracted; In the feature extraction sub-network of the underwater point cloud environment recognition model, based on the underwater point cloud set of the first trajectory point of each candidate path of the unmanned vessel, the point cloud energy consumption feature vector of the first trajectory point of the corresponding candidate path is extracted; In the output sub-network of the underwater point cloud environment recognition model, based on the point cloud energy consumption feature vector of the first trajectory point of each candidate path of the unmanned vessel...

[0012] Furthermore, the specific steps for extracting the underwater point cloud set of the first trajectory point of each candidate path of the designated unmanned vessel are as follows: In the input preprocessing layer of the trajectory classification subnetwork, the current three-dimensional position coordinates of the designated unmanned vessel, the underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path are preprocessed; In the classification and recognition output layer of the trajectory classification subnetwork, based on the preprocessed current three-dimensional position coordinates of the designated unmanned vessel, the underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path, the underwater point cloud set of the first trajectory point of the corresponding candidate path is output.

[0013] Furthermore, the specific steps for extracting the point cloud energy consumption feature vector of the first trajectory point of each candidate path of the unmanned vessel are as follows: In the input extraction layer of the feature extraction subnetwork, the underwater point cloud set of the first trajectory point of each candidate path of the unmanned vessel is received and voxel normalization is performed; in the extraction output layer of the feature extraction subnetwork, based on the underwater point cloud set of the first trajectory point of each candidate path of the unmanned vessel after voxel normalization, the point cloud energy consumption feature vector of the first trajectory point of each candidate path of the unmanned vessel is output.

[0014] A system for automatic return to shore of a Beidou AI-based unmanned vessel in the event of a malfunction includes: a data acquisition module for real-time acquisition of the current 3D position coordinates of the unmanned vessel, the target 3D position coordinates of a preset dock, and surface and underwater point cloud data; a candidate path generation module for analyzing several candidate paths for the unmanned vessel based on a path planning algorithm and in conjunction with the surface and underwater point cloud data of the unmanned vessel, including the 3D coordinates of several candidate trajectory points; a return-to- shore adaptation analysis module for analyzing the return-to- shore adaptation feature value of the first trajectory point of each candidate path based on a pre-trained underwater point cloud environment recognition model and in conjunction with the current 3D position coordinates of the unmanned vessel, underwater point cloud data, and the 3D coordinates of the first trajectory point of each candidate path; and a return-to- shore iterative control module for performing iterative return-to- shore processing on the unmanned vessel based on the return-to- shore adaptation feature value of the first trajectory point of each candidate path until the unmanned vessel reaches the target 3D position coordinates of the preset dock.

[0015] The present invention has the following beneficial effects:

[0016] (1) The method of automatic return to shore for unmanned vessels based on Beidou AI faults is to construct a fusion analysis of surface point cloud and underwater point cloud, so as to simultaneously acquire spatial information of the surface and underwater environment during the return process. It also uses an obstacle recognition model to automatically divide the obstacle areas on the surface and underwater. Combined with the path planning algorithm, it can completely avoid obstacle areas when generating candidate paths, thereby ensuring that the unmanned vessel still has the ability to navigate safely in the state of low battery or controllable fault. At the same time, the return adaptation feature value is introduced in the path selection process, so that the unmanned vessel can autonomously select a return path that is safe in terms of obstacle avoidance and reasonable energy consumption, thereby realizing the return safety of unmanned vessels in the state of controllable fault in complex water conditions.

[0017] (2) The method of automatic return to shore for Beidou AI-based unmanned vessels in case of failure is to construct an underwater point cloud environment recognition model, thereby extracting and outputting point cloud energy consumption feature vectors during the candidate path analysis process, and forming an energy consumption risk feature set based on the output sub-network to characterize the energy consumption differences of candidate paths in the underwater environment. This enables the return path evaluation to achieve quantitative judgment of the energy consumption level of different candidate paths, and prioritizes the return route with lower energy consumption and better path stability among the candidate paths. This allows the unmanned vessel to avoid high energy consumption in the low power or failure state, thereby reducing the energy consumption risk during the return process and significantly improving the success rate of the return.

[0018] (3) The method of automatic return to shore for unmanned vessels based on Beidou AI is to acquire surface point cloud data and underwater point cloud data and input them into an obstacle recognition model for fusion recognition. The model can accurately divide the surface obstacle area and the underwater obstacle area and merge or expand them in the spatial domain to form a complete three-dimensional obstacle area expression. This ensures that the planning of the return path is not limited to the safety judgment of the surface environment, but can also cover the obstacle recognition of the underwater space. This effectively avoids potential risks caused by missing information and enables the unmanned vessel to achieve an overall safety assessment of the navigation waters during the return process.

[0019] (4) The system for automatic return to shore of the Beidou AI-based unmanned vessel in case of failure, through modular collaborative analysis, enables the unmanned vessel to maintain stable return capability even in the case of low battery or controllable failure. Through real-time perception of the data acquisition module, the system can continuously acquire the hull position, target dock position and point cloud environment information in the case of failure, providing high-precision input for subsequent path planning and analysis. The candidate path generation module and the return to shore adaptation analysis module are functionally connected to ensure that the return path is not only spatially feasible, but also dynamically optimized through energy consumption and adaptability indicators. The return to shore iteration control module executes the navigation command of the optimal path in an iterative manner, thereby significantly improving the real-time performance of the system operation and effectively reducing the risks caused by processing delay or information coupling during the return process.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of a method for automatic return to port of a Beidou AI-based unmanned vessel in case of malfunction, according to the present invention.

[0022] Figure 2 This is a flowchart illustrating the specific steps involved in analyzing and setting several candidate paths for an unmanned vessel in an automatic return-to-base method based on Beidou AI malfunction, as described in this invention.

[0023] Figure 3 This is a schematic diagram of the data for setting the first trajectory point in the candidate path sequence of an unmanned vessel in the method for automatic return to port of a Beidou AI-based unmanned vessel in the present invention.

[0024] Figure 4 This is a system block diagram of an unmanned vessel based on Beidou AI that automatically returns to port in case of malfunction. Detailed Implementation

[0025] Please see Figure 1 This invention provides a technical solution: a method for automatic return to shore based on Beidou AI unmanned vessel faults, comprising the following steps: when the unmanned vessel is in a low battery state or a controllable fault (i.e., the battery capacity decreases but can still supply the power required for return, the thruster thrust is abnormal but can still maintain low-speed navigation, some sensors fail but the overall environmental perception is usable, the communication link is unstable but can still maintain the return control command), the current three-dimensional position coordinates (in a geographic coordinate system, such as the WGS-84 coordinate system) of the set unmanned vessel (and the unmanned vessel has an IP68 protection level) and the target three-dimensional position coordinates of the preset dock, as well as (using the line connecting the current three-dimensional position coordinates and the target three-dimensional position coordinates of the preset dock as the direction of travel, and setting a semi-circular perception range with a radius of 60 meters in front of this direction) water surface point cloud data (which can be obtained through lidar) and underwater point cloud data (which can be obtained through ultrasonic multibeam echo sounder) are acquired in real time.

[0026] Based on path planning algorithms and combined with surface and underwater point cloud data of the unmanned surface vessel (USV), several candidate paths for the USV (from its current 3D position coordinates to the target endpoint within a pre-defined semi-circular perception range in its forward direction) are analyzed, including the 3D coordinates of the trajectories of several candidate trajectory points. Based on a pre-trained underwater point cloud environment recognition model and combined with the USV's current 3D position coordinates, underwater point cloud data, and the 3D coordinates of the trajectory of the first trajectory point of each candidate path, the return-to-course adaptation feature value of the first trajectory point of the corresponding candidate path is analyzed (this represents the overall adaptability of the candidate path containing a certain trajectory point during the USV's return process; a larger return-to-course adaptation feature value indicates that the candidate path is suitable for priority selection under conditions of low energy consumption, reasonable trajectory, and stable environment).

[0027] Based on the return-to-home adaptation feature value of the first trajectory point of each candidate path, the set unmanned vessel undergoes return-to-home iterative processing (i.e., selecting the first trajectory point of the candidate path corresponding to the maximum return-to-home adaptation feature value as the first target trajectory point, and driving the set unmanned vessel to the first target trajectory point; upon reaching the first target trajectory point, repeating the above steps of candidate path generation and return-to-home adaptation feature value analysis to obtain the return-to-home adaptation feature value of the second trajectory point of each candidate path; subsequently, selecting the second trajectory point of the candidate path corresponding to the maximum return-to-home adaptation feature value as the second target trajectory point, and driving the unmanned vessel to continue sailing to the second target trajectory point; the above process is iteratively executed until the unmanned vessel sails to the target endpoint within the current semi-circular perception range; when the unmanned vessel reaches the target endpoint position, this position is taken as the new current position, a new semi-circular perception range is reset, and the steps of candidate path generation, return-to-home adaptation feature value calculation, and target trajectory point selection are repeated again), until the set unmanned vessel arrives at the target three-dimensional position coordinates of the preset dock.

[0028] Specifically, such as Figure 2 As shown, the specific steps for analyzing several candidate paths for the unmanned surface vessel (USV) are as follows: The surface point cloud data and underwater point cloud data of the USV are input into a pre-trained obstacle recognition model to identify several obstacle regions within the USV's semi-circular perception range; based on each obstacle region, the target endpoint of the USV within its semi-circular perception range is analyzed; the current 3D position coordinates of the USV are read, and based on the path planning algorithm, combined with each obstacle region and the target endpoint (within the semi-circular perception range), several candidate paths for the USV (to the target endpoint within the semi-circular perception range) are analyzed, specifically as follows:

[0029] Within a semi-circular perception range between the current 3D position coordinates of the unmanned vessel and the target endpoint, the system is meshed or voxelized at a preset resolution to obtain a 3D search space composed of several candidate trajectory points. Each obstacle voxel point in each obstacle region is mapped into the 3D search space. Candidate trajectory points overlapping with obstacle regions are marked as impassable points. A path planning algorithm (such as the graph-based A* algorithm) is invoked to perform path search within the 3D search space. During the search process, the current 3D position coordinates of the unmanned vessel are used as the starting node, and the target endpoint is used as the target node. Candidate trajectory points adjacent to the current node are gradually expanded, and the comprehensive cost of each expanded trajectory point is calculated. The comprehensive cost consists of path length cost, obstacle distance cost, and turning radius cost (i.e., the above costs are weighted). The path length cost represents the cumulative length of the path, the obstacle distance cost represents the inverse relationship between the minimum distance between the trajectory point on the path and the nearest obstacle, and the turning radius cost represents the smoothness of the path turning angle. When an expanded candidate trajectory point is marked as an impassable point, the path branch is directly discarded and no further expansion is performed.

[0030] Repeat the above expansion process until the starting node and the target node are connected to generate multiple candidate paths to avoid the obstacle area. Smooth the generated candidate paths, such as by spline interpolation, and improve the continuity and executability of the paths by eliminating redundant points. Finally, several candidate paths are obtained for the unmanned ship (to the target endpoint within the semi-circular perception range), and each candidate path includes several candidate trajectory points and corresponding trajectory three-dimensional coordinates.

[0031] The specific steps for analyzing and setting the target endpoint of the unmanned vessel (within the semi-circular perception range) are as follows: Based on setting each obstacle area of ​​the unmanned vessel, analyze and set several candidate safe areas for the unmanned vessel (within the boundary neighborhood strip area of ​​the semi-circular perception range). Specifically, the boundary neighborhood strip area is located at the edge of the semi-circular perception range in the direction of travel of the unmanned vessel, covering the water area within a preset distance range (e.g., a strip area of ​​5-10 meters) from the current three-dimensional position coordinates of the unmanned vessel, and within a preset angle range to the left and right of the direction of travel (e.g., an angle range of 15° to 30° to the left and right). Each obstacle region of the unmanned vessel is mapped to a boundary neighborhood strip region (i.e., if any obstacle voxel point in the obstacle region is located within the strip region, the obstacle voxel point falling into it is marked as an impassable voxel point, and cluster analysis is performed to obtain several non-communication regions. The remaining water surface voxel points and their corresponding water surface voxel values ​​and water surface 3D coordinates in the boundary neighborhood strip region are uniformly marked as the passable voxel values ​​and passable 3D coordinates of passable voxel points, and all passable voxel points are clustered) to obtain several candidate safe regions (including the passable voxel values ​​and passable 3D coordinates of several passable voxel points).

[0032] The current 3D position coordinates of the designated unmanned vessel are read, and the target endpoint of the designated unmanned vessel is analyzed in combination with each candidate safe zone. Specifically, based on the 3D coordinates of each passage voxel point in each candidate safe zone, the mean is processed to obtain the center 3D coordinates of each candidate safe zone. Based on the current 3D position coordinates of the designated unmanned vessel and the center 3D coordinates of each candidate safe zone, the Euclidean distance from each candidate safe zone to the designated unmanned vessel is extracted. Several non-communication zones in the boundary neighborhood strip area are read, and the center 3D coordinates of each non-communication zone are extracted. For each candidate safe zone, the distances from it to several non-communication zones in the boundary neighborhood strip area are extracted and accumulated to obtain the safe distance of each candidate safe zone. The Euclidean distance and safe distance from each candidate zone to the designated unmanned vessel are weighted, and the Euclidean distance is inverted during the weighting process, i.e., 1 / (1+the Euclidean distance), to obtain the safety reliability value of each candidate safe zone. The candidate safe zone with the largest safety reliability value is selected as the safe zone, and the center 3D coordinates of the safe zone are taken as the target endpoint.

[0033] In this implementation scheme, by inputting both surface and underwater point clouds into the obstacle recognition model, obstacle areas on and under the water can be identified and labeled simultaneously and projected into a three-dimensional search space. This effectively avoids conflicts between the return path and potential obstacles and improves navigation safety. Secondly, by constructing a boundary neighborhood strip region at the edge of the perception range and extracting candidate safe regions within this region, the potential safety hazards of directly using the boundary intersection as the target point are avoided, making the selection of the target endpoint more robust. Finally, a safety reliability value is introduced into the candidate safe region, comprehensively considering the distance from the current position of the unmanned vessel to the candidate region and the safety of the candidate region relative to the obstacle region. This ensures both the efficiency of the return path and the stability during navigation. Ultimately, the candidate safe region with the highest safety reliability value is selected as the target endpoint, making the return path adaptive and significantly improving the return success rate of the unmanned vessel in low battery or controllable fault conditions.

[0034] Specifically, the surface point cloud data includes surface voxel values ​​and three-dimensional coordinates of several surface voxel points. The underwater point cloud data includes underwater voxel values ​​(i.e., echo intensity values) and three-dimensional coordinates of each underwater voxel point. It should be noted that both the surface and underwater three-dimensional coordinates are in the ship's coordinate system, i.e., the bow direction of the unmanned vessel is the positive X-axis, the port side direction is the positive Y-axis, and the vertically upward direction is the positive Z-axis. The obstacle recognition model includes an input layer, an obstacle extraction layer, and an obstacle recognition output layer. The specific steps for recognizing several obstacle areas of the designated unmanned vessel are as follows: In the input layer of the obstacle recognition model, the surface point cloud data and underwater point cloud data of the designated unmanned vessel are received. Cloud data is processed and preprocessed, specifically as follows: the pose information of the unmanned vessel is acquired, which can be obtained through attitude sensors such as heading angle, roll angle, and pitch angle, as well as the current three-dimensional position coordinates of the unmanned vessel in the geographic coordinate system obtained by Beidou navigation; based on the pose information, a coordinate transformation matrix between the ship coordinate system and the geographic coordinate system (WGS-84 coordinate system) is established, and a combination of translation and rotation transformation is used to transform the surface point cloud data and underwater point cloud data from the ship coordinate system to a unified coordinate expression in the WGS-84 coordinate system. After the above preprocessing, the point cloud data in the WGS-84 coordinate system is consistent with the current three-dimensional position coordinates of the unmanned vessel and the target three-dimensional position coordinates of the preset dock.

[0035] In the obstacle extraction layer of the obstacle recognition model, based on the preprocessed surface point cloud data and underwater point cloud data of the designated unmanned vessel, the surface obstacle point cloud set and underwater obstacle point cloud set of the designated unmanned vessel are extracted. Specifically, the preprocessed surface point cloud data and underwater point cloud data of the designated unmanned vessel are processed by an existing 3D point cloud recognition network (such as PointNet, PointNet++, etc.). (That is, the preprocessed surface point cloud data and underwater point cloud data are input into the network. The network automatically extracts and encodes the spatial distribution features, geometric features and local neighborhood features of the point cloud, and distinguishes obstacle voxels from non-obstacle voxels based on classification or segmentation mechanisms. The category information of obstacle voxels includes, but is not limited to, floating objects, ships, bridge piers, shoals, reefs, etc.) to obtain the surface obstacle point cloud set (including several surface obstacle voxels) and the underwater obstacle point cloud set (including several underwater obstacle voxels).

[0036] In the obstacle recognition output layer of the obstacle recognition model, based on the set surface obstacle point clusters and underwater obstacle point clusters of the unmanned vessel, several obstacle regions of the set unmanned vessel are identified. Specifically, the surface obstacle point clusters are clustered, that is, based on the spatial proximity of obstacle voxels (i.e., the relative distance between different obstacle voxels; for example, if the difference between the three-dimensional coordinates of two obstacle voxels on the water surface meets a preset proximity threshold, then these two voxels are considered to be spatially adjacent) and category information (including but not limited to floating objects, ships, bridge piers, etc.), mutually adjacent and consistent obstacle voxels are grouped into the same cluster, thus dividing them into several surface obstacle regions. Similarly, several underwater obstacle regions can be obtained. For each underwater obstacle region, it is determined whether they are in the same spatial range. Does the corresponding surface obstacle region exist within the space domain? If it does, the underwater obstacle region is merged with the corresponding surface obstacle region, and their outer boundary is used as a complete obstacle region. If it does not exist, the space domain where the underwater obstacle region is located, together with the corresponding surface region above it, is marked as an obstacle region to ensure that the vertical space domain is determined to be an impassable area. Similarly, for surface obstacle regions not covered by any underwater obstacle regions, they are also treated as obstacle regions. Based on the above fusion and expansion determination results, several obstacle regions are output (including the obstacle voxel values ​​of several obstacle voxel points and the obstacle 3D coordinates). Each obstacle region corresponds to the surface and underwater obstacle regions within the same space domain after clustering and merging or spatial domain expansion processing.

[0037] The pre-training steps for the obstacle recognition model are as follows:

[0038] A labeled dataset was obtained, consisting of surface and underwater point cloud data collected by an unmanned vessel in different aquatic environments. Data acquisition was achieved using lidar and multibeam echo sounders. The labeled dataset was annotated by waterway monitoring experts based on the actual on-site environment, forming an obstacle dataset with ground truth labels. The labeled categories include, but are not limited to, floating objects, ships, bridge piers, shoals, and reefs. Each sample in the labeled dataset includes: the 3D coordinates and voxel value of each voxel in the point cloud; the corresponding obstacle category label (or non-obstacle point label); and the category information of the surface or underwater point cloud. All samples have complete annotations. The labeled dataset was divided proportionally into training, validation, and test sets, for example, 80% for training, 10% for validation, and 10% for testing, to ensure that generalization performance can be evaluated during training.

[0039] The preprocessed point cloud data is input into a 3D point cloud recognition network (such as PointNet, PointNet++, VoxelNet, PointPillars, etc.). During the training process, the model automatically extracts the spatial distribution features, geometric features, and local neighborhood features of the point cloud. For water surface point cloud samples, the network learns the features of obstacles such as floating objects, ships, and bridge piers. For underwater point cloud samples, the network learns the features of obstacles such as shoals, reefs, and submerged reefs. The network distinguishes obstacle voxels from non-obstacle voxels through point-by-point classification and cluster-level segmentation mechanisms, and predicts the category label of obstacle voxels.

[0040] During training, the cross-entropy loss function or weighted cross-entropy loss function is used to minimize the difference between the predicted class and the ground truth label. At the same time, the IoU (Intersection over Union) evaluation metric is introduced to optimize the boundary recognition accuracy of obstacle areas. The network parameters are adjusted by optimization algorithms (such as Adam optimizer and SGD optimizer), and hyperparameters such as learning rate and batch size are set.

[0041] After each training cycle, the model performance is evaluated using a validation set, and the hyperparameters are adjusted based on the validation results to avoid overfitting. After training is complete, the model's generalization ability is evaluated using a test set to ensure that the model can accurately identify surface and underwater obstacles in different water areas and under different environmental conditions.

[0042] In this implementation scheme, by uniformly inputting surface and underwater point clouds into the obstacle recognition model, obstacle information in both the surface and underwater environments can be simultaneously covered, achieving three-dimensional perception across the entire water area. Secondly, attitude sensor and BeiDou navigation pose information are introduced into the input layer, and the point cloud data is unified to the WGS-84 geographic coordinate system through coordinate transformation. This not only ensures the coordinate consistency of the data but also provides a reliable spatial reference for subsequent path planning and return control. Furthermore, the three-dimensional point cloud recognition network called in the obstacle extraction layer allows the model to automatically extract point cloud features, thereby achieving accurate differentiation between obstacle voxels and non-obstacle voxels, while also identifying obstacle category information. Finally, in the obstacle recognition output layer, a clustering and spatial domain merging strategy is used to fuse and identify surface obstacles and corresponding underwater obstacles, ensuring that any vertical spatial domain is completely identified as an obstacle area, thus significantly improving the completeness and safety of the recognition.

[0043] Specifically, the underwater point cloud environment recognition model includes a trajectory classification subnetwork, a feature extraction subnetwork, and an output subnetwork. The trajectory classification subnetwork includes an input preprocessing layer and a classification and recognition output layer. The feature extraction subnetwork includes an input extraction layer and an extraction output layer.

[0044] The specific steps for analyzing the return-to-home adaptation feature value of the first trajectory point of each candidate path for the unmanned vessel are as follows: Input the current three-dimensional position coordinates of the unmanned vessel, underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path into the pre-trained underwater point cloud environment recognition model, and analyze the energy consumption risk feature set of the first trajectory point of the corresponding candidate path, including point cloud flow resistance feature value, echo energy attenuation feature value, underwater turbulence feature value, and point cloud distribution heterogeneity feature value; obtain the trajectory adaptation feature value of the first trajectory point of each candidate path for the unmanned vessel, and combine it with the energy consumption risk feature set to analyze the return-to-home adaptation feature value of the first trajectory point of the corresponding candidate path.

[0045] The specific steps for obtaining the trajectory adaptation feature value of the first trajectory point of each candidate path for the designated unmanned vessel are as follows: Read the current three-dimensional position coordinates of the designated unmanned vessel and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path, and analyze the travel distance between the designated unmanned vessel and the first trajectory point of each candidate path based on the Euclidean distance formula; read the current three-dimensional position coordinates of the designated unmanned vessel and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path, and perform difference processing to obtain the travel direction vector between the designated unmanned vessel and the first trajectory point of each candidate path.

[0046] The current 3D position coordinates of the designated unmanned vessel and the target 3D position coordinates of the preset dock are read and interpolated to obtain the target direction vectors of the designated unmanned vessel and the preset dock. The vector angle between the vector vector vector and the navigation direction vector of the first trajectory point of each candidate path is calculated (e.g., using the vector dot product formula) to obtain the heading deviation value between the designated unmanned vessel and the first trajectory point of each candidate path. Then, the deviation value is standardized with the navigation distance value. The standardization result is then weighted and inverted, i.e., 1 / (1+the weighting), to obtain the trajectory adaptation feature value of the first trajectory point of each candidate path of the designated unmanned vessel.

[0047] The specific steps for analyzing the return-to-navigation adaptation characteristic value of the first trajectory point of each candidate path for the unmanned vessel are as follows: The point cloud flow resistance characteristic value, echo energy attenuation characteristic value, underwater turbulence characteristic value, and point cloud distribution heterogeneity characteristic value of the first trajectory point of each candidate path are weighted (the weighted result is then mapped to the 0-1 interval using a Sigmoid activation function). During the weighting process, the point cloud flow resistance characteristic value, echo energy attenuation characteristic value, and underwater turbulence characteristic value are all inverted, e.g., 1 / (1+point cloud flow resistance characteristic value), to obtain the energy consumption navigation characteristic value of the first trajectory point of each candidate path for the unmanned vessel. Combined with the trajectory adaptation characteristic value, the return-to-navigation adaptation characteristic value of the first trajectory point of each candidate path for the unmanned vessel is analyzed. The specific formula for calculating the return-to-navigation adaptation characteristic value of the first trajectory point of a candidate path for the unmanned vessel is as follows: Wherein, FhY is the return-to-home adaptation characteristic value of the first trajectory point of a candidate path of the unmanned vessel, NhS is the energy consumption seaworthiness characteristic value of the first trajectory point of a candidate path of the unmanned vessel, HjS is the track adaptation characteristic value of the first trajectory point of a candidate path of the unmanned vessel, α is the interaction adjustment coefficient stored in the database, β is the difference adjustment coefficient stored in the database, and η is the smoothing adjustment coefficient stored in the database. In this embodiment, the interaction adjustment coefficient α, the difference adjustment coefficient β, and the smoothing adjustment coefficient η stored in the database are respectively 1.500, 1.250, and 0.864.

[0048] The following is a specific implementation example for calculating the return-to-home adaptation characteristic value of the first trajectory point of a candidate path for a given unmanned vessel. Existing data includes: energy consumption airworthiness characteristic values ​​and trajectory adaptation characteristic values ​​of the first trajectory point of three candidate paths for the unmanned vessel, as detailed in Table 1 and... Figure 3 As shown:

[0049] Table 1. Example of data for the first trajectory point in the candidate path sequence of the unmanned vessel.

[0050] Energy consumption airworthiness characteristic values Track adaptation feature value Candidate Path 1 0.824 0.548 Candidate Path 2 0.716 0.739 Candidate Path 3 0.764 0.625

[0051] The interaction moderating coefficient α stored in the database is 1.500;

[0052] The difference adjustment coefficient β stored in the database is 1.250;

[0053] The smoothing adjustment coefficient η stored in the database is 0.864;

[0054] Substituting the data from Table 1 and the aforementioned coefficients into the specific formula for calculating the return-to-home adaptation characteristic value of the first trajectory point of a candidate path for a given unmanned vessel, we obtain:

[0055] The return-to-home adaptation feature value of the first trajectory point of the first candidate path of the unmanned ship is set as ln(1+1.500×√(0.824×0.548)) / (1+1.250×|0.824-0.548|)+0.864×min(0.824,0.548)≈0.991;

[0056] Set the return-to-home adaptation characteristic value of the first trajectory point of the second unmanned ship as ln(1+1.500×√(0.716×0.739)) / (1+1.250×|0.716-0.739|)+0.864×min(0.716,0.739)≈1.335;

[0057] The return-to-home adaptation feature value of the first trajectory point of the third candidate path of the unmanned ship is set as ln(1+1.500×√(0.764×0.625)) / (1+1.250×|0.764-0.625|)+0.864×min(0.764,0.625)≈1.146.

[0058] In this implementation scheme, by introducing trajectory adaptation feature values ​​and energy consumption seaworthiness feature values, candidate paths can be comprehensively evaluated from both geometric path rationality and underwater energy consumption environment. This ensures that the selection of return paths not only depends on geometric optimality but also fully considers energy consumption and environmental adaptability, thereby improving the comprehensiveness of decision-making. Secondly, in the analysis of trajectory adaptation feature values, the standardization and weighting of navigation distance and heading deviation are combined, and inversion is introduced, so that paths with shorter distances and smaller deviations have a numerical advantage. This ensures that the return path is geometrically more in line with the requirements of fast and stable navigation. Finally, in the extraction of energy consumption seaworthiness feature values, energy consumption risk features are uniformly mapped to the 0-1 interval through the Sigmoid activation function. This not only enhances the comparability between different feature dimensions but also avoids the interference of extreme values ​​on the overall evaluation, thereby achieving accurate screening of return paths under complex water conditions and significantly improving the return success rate of unmanned vessels under controllable fault conditions.

[0059] Specifically, the steps for analyzing the energy consumption risk feature set of the first trajectory point of each candidate path of the unmanned vessel are as follows: In the trajectory classification sub-network of the underwater point cloud environment recognition model, the current three-dimensional position coordinates of the unmanned vessel, underwater point cloud data, and trajectory three-dimensional coordinates of the first trajectory point of each candidate path are received, and the underwater point cloud set of the first trajectory point of the corresponding candidate path is extracted; In the feature extraction sub-network of the underwater point cloud environment recognition model, based on the underwater point cloud set of the first trajectory point of each candidate path of the unmanned vessel, the point cloud energy consumption feature vector of the first trajectory point of the corresponding candidate path is extracted.

[0060] In the output subnetwork of the underwater point cloud environment recognition model, based on the point cloud energy consumption feature vector of the first trajectory point of each candidate path of the unmanned vessel, the energy consumption risk feature set of the first trajectory point of the corresponding candidate path is output. Specifically, the output subnetwork includes a feature normalization layer and a feature output layer. In the feature normalization layer, the point cloud flow resistance feature, echo energy attenuation feature, underwater turbulence feature, and point cloud distribution heterogeneity feature in the point cloud energy consumption feature vector are processed by the Sigmoid activation function, and each feature is mapped to the 0-1 interval. In the feature output layer, the point cloud flow resistance feature, echo energy attenuation feature, underwater turbulence feature, and point cloud distribution heterogeneity feature processed by the Sigmoid activation function are output as the point cloud flow resistance feature value, echo energy attenuation feature value, underwater turbulence feature value, and point cloud distribution heterogeneity feature value of the first trajectory point of the corresponding candidate path.

[0061] The pre-training steps of the underwater point cloud environment recognition model are as follows: Obtain the underwater annotation dataset, which consists of surface point cloud data and underwater point cloud data collected synchronously by the lidar, ultrasonic multibeam echo sounder and water flow sensor carried by the unmanned vessel under different navigation environments. The annotation is completed manually by hydrological and ship engineering experts based on the measured environmental conditions and energy consumption monitoring results. Each sample includes: the underwater point cloud set corresponding to the first trajectory point of the candidate path, the ground truth label of point cloud flow resistance, the ground truth label of echo energy attenuation, the ground truth label of underwater turbulence, and the ground truth label of point cloud heterogeneity. The underwater annotation dataset is divided into training set, validation set and test set.

[0062] During the training phase, taking the feature extraction subnetwork as an example, the underwater point cloud corresponding to the candidate path trajectory points is input into the feature extraction subnetwork, which is then processed by voxelization and normalization. The spatial distribution features, geometric structure features, and neighborhood correlation features are extracted layer by layer. The feature extraction subnetwork automatically captures the density changes, curvature fluctuations, and energy attenuation patterns of the point cloud in local regions through its internal multi-layer perception structure (including voxel convolutional layers and local aggregation layers), and outputs the corresponding point cloud energy consumption feature vector. The point cloud energy consumption feature vector includes point cloud flow resistance features, echo energy attenuation features, underwater turbulence features, and point cloud distribution heterogeneity features.

[0063] During training, the Adam optimizer is used to update parameters and adjust hyperparameters (such as learning rate, batch size, and number of sub-network layers). The model performance is evaluated and early stopping is controlled through the validation set to avoid overfitting. Finally, the generalization ability of the model is tested using the test set to ensure that the model can accurately extract features from unseen point cloud data.

[0064] Once the model converges and reaches optimal performance, the trained underwater point cloud environment recognition model is saved. Its model parameters can be called in the subsequent actual deployment stage to realize real-time energy consumption risk feature extraction and return-to-go adaptive analysis of candidate paths during the return process of the unmanned vessel.

[0065] The specific steps for extracting the underwater point cloud set of the first trajectory point of each candidate path of the set unmanned vessel are as follows: In the input preprocessing layer of the trajectory classification subnetwork, the current three-dimensional position coordinates of the set unmanned vessel, the underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path are preprocessed. Specifically, the underwater point cloud data is subjected to statistical filtering, that is, the average distance of the neighboring points is calculated within the neighborhood of each voxel. When the average distance between a voxel and its neighboring points exceeds a preset distance threshold, the point is determined to be an isolated point and deleted, thereby removing isolated outliers. The filtered underwater point cloud data is then subjected to interpolation smoothing to maintain the continuity of the point cloud in spatial distribution, thereby retaining the effective point cloud data related to the candidate trajectory path.

[0066] In the classification and recognition output layer of the trajectory classification subnetwork, based on the preprocessed current 3D position coordinates of the unmanned vessel, underwater point cloud data, and the trajectory 3D coordinates of the first trajectory point of each candidate path, the underwater point cloud set of the first trajectory point of the corresponding candidate path is output. Specifically, based on the current 3D position coordinates of the unmanned vessel and the trajectory 3D coordinates of the first trajectory point of each candidate path, a spatial connection between the current position of the unmanned vessel and the candidate trajectory point is determined, and the spatial connection is defined as the travel area axis of the candidate path. A neighborhood range with a predetermined width and depth is set on both sides of the travel area axis of the candidate path to form a travel area band extending from the current 3D position coordinates to the trajectory 3D coordinates of the first trajectory point. All underwater voxel points falling within the travel area band are selected from the preprocessed underwater point cloud data, which are the underwater point cloud set of the first trajectory point of each candidate path.

[0067] The specific steps for extracting the point cloud energy consumption feature vector of the first trajectory point of each candidate path of the set unmanned vessel are as follows: In the input extraction layer of the feature extraction sub-network, the underwater point cloud set of the first trajectory point of each candidate path of the set unmanned vessel is received and voxel normalization processing is performed (that is, the voxel value of each underwater voxel point is normalized to the 0-1 range).

[0068] In the extraction output layer of the feature extraction subnetwork, based on the underwater point cloud of the first trajectory point of each candidate path of the unmanned vessel after voxel normalization, the point cloud energy consumption feature vector of the first trajectory point of each candidate path of the unmanned vessel is output. Specifically, for the underwater point cloud corresponding to the first trajectory point of any candidate path, it is divided into voxel units of fixed size (such as a 3D grid of 0.5m×0.5m×0.5m), and the number of underwater voxel points in each voxel unit is counted to obtain the voxel point density of each voxel unit; the mean and variance of underwater voxel points in each voxel unit are counted and processed as a ratio, i.e., underwater voxel point variance / underwater voxel point mean, and the result is standardized with the voxel point density. Based on the standardized result, a weighted processing is performed and the mean is taken to extract the point cloud flow resistance feature (the larger the value, the denser and unevenly distributed the point cloud in the local area, the more significant the water resistance effect, and the more power the unmanned vessel needs to consume when passing through).

[0069] For the underwater point cluster corresponding to the first trajectory point of any candidate path, the area is divided into several equidistant segments (e.g., each segment is 1 meter apart) according to the direction of the line connecting the current position of the unmanned vessel to the first trajectory point of the candidate path. The average echo intensity value (i.e., the mean value of the voxels) of the underwater voxels in each segment is extracted. The average echo intensity difference (absolute value) between the center points of adjacent segments is analyzed and divided by the segment spacing (i.e., 1 meter) to obtain the echo intensity attenuation rate of that segment (i.e., the next adjacent segment). To reflect the differentiated impact of different segments on the overall energy consumption environment, the echo intensity attenuation rate of each segment (except the first segment) is multiplied by a weighting factor to obtain the weighted attenuation rate. The weighting factor is defined as the distance from the current position (i.e., the current three-dimensional position coordinates of the unmanned vessel) to the next segment. The ratio of the horizontal distance of a segment (the distance between the segment and the center point of the current three-dimensional position coordinates of the unmanned vessel on the same horizontal plane, where the center point is the mean of the underwater three-dimensional coordinates of all underwater voxels on the same horizontal plane) to the total distance from the current position to the first trajectory point of the candidate path (the Euclidean distance between the current three-dimensional position coordinates of the unmanned vessel and the trajectory three-dimensional coordinates of the first trajectory point of the candidate path) is used to reflect the relative contribution of the segment in the overall path. The weighted attenuation rate of all segments (except the first segment) is accumulated and normalized to obtain the echo energy attenuation characteristic of the first trajectory point of the candidate path. When this characteristic is large, it indicates that the echo intensity attenuates significantly with distance, which indicates high water turbidity, strong disturbance, and increased energy consumption when the vessel passes through.

[0070] For each underwater voxel in the underwater point cloud corresponding to the first trajectory point of any candidate path, a neighborhood is established centered on it (with an optional radius of 0.5m). All underwater voxels falling within the spherical area of ​​this neighborhood are extracted, and a plane is fitted using the least squares method. That is, based on the underwater 3D coordinates of each underwater voxel in the neighborhood, a reference plane is found such that the vertical distance from all neighborhood points to this plane is minimized overall. An optimal combination of plane parameters is calculated using the least squares method, which minimizes the sum of squares of the deviations of all neighborhood points from the plane. This results in a reference plane that best approximates the overall distribution trend of the neighborhood point set, which is the fitted plane. The deviation of each underwater voxel in the neighborhood from the fitted plane is calculated one by one, i.e., the vertical distance from the underwater voxel to the plane, and root mean square processing is performed to obtain the curvature of each underwater voxel. The mean curvature is then extracted. Curvature variance is calculated and weighted to extract underwater turbulence features. When the feature is large, it indicates that the geometric fluctuations of the regional point cloud are severe, representing a potential strong eddy or turbulent zone. Frequent corrections are required when the ship passes through, and the return energy consumption is significantly increased.

[0071] For each underwater voxel in the underwater point cloud corresponding to the first trajectory point of any candidate path, it is projected onto three orthogonal coordinate planes: the X-Y plane, the Y-Z plane, and the X-Z plane. Each coordinate plane is divided into several grid cells, and the number of underwater voxels in each grid cell is counted. The variance of the number of underwater voxels on each coordinate plane is also counted. The ratio of the variances of the number of underwater voxels between any two coordinate planes is weighted to extract the heterogeneous features of the point cloud distribution. The smaller this feature is, the more uniform the environment and the lower the energy consumption risk. The point cloud flow resistance feature, echo energy attenuation feature, underwater turbulence feature, and point cloud distribution heterogeneous feature are then concatenated into a point cloud energy consumption feature vector.

[0072] In this implementation scheme, in the trajectory classification subnetwork, a travel area zone is constructed based on the spatial connection between the current 3D position coordinates and candidate trajectory points, and relevant voxel points are accurately selected from underwater point cloud data to ensure the relevance and effectiveness of the input data and avoid interference from irrelevant noise on the analysis results. Secondly, in the feature extraction subnetwork, deep extraction from geometric features to energy consumption-related indicators is achieved. In the output subnetwork, all features are normalized by the Sigmoid activation function to eliminate the differences in the dimensions of different features, thus ensuring comparability and the robustness of the return-to-base adaptive calculation. Finally, this step integrates key features into a unified energy consumption risk feature set, which can effectively distinguish the energy consumption level and passage risk of different candidate paths in the return-to-base path determination, thereby enabling the unmanned vessel to have a more accurate path selection capability in the fault state, and thus significantly improving the safety of the return-to-base.

[0073] Please see Figure 4 This invention provides a technical solution: a system for automatic return to shore based on Beidou AI-powered unmanned vessels in case of malfunction, comprising: a data acquisition module for real-time acquisition of the current three-dimensional position coordinates of the designated unmanned vessel, the target three-dimensional position coordinates of a preset dock, and surface and underwater point cloud data; a candidate path generation module for analyzing several candidate paths of the designated unmanned vessel based on a path planning algorithm and in combination with the surface and underwater point cloud data of the designated unmanned vessel, including the trajectory three-dimensional coordinates of several candidate trajectory points; a return-to- shore adaptation analysis module for analyzing the return-to- shore adaptation feature value of the first trajectory point of each candidate path based on a pre-trained underwater point cloud environment recognition model and in combination with the current three-dimensional position coordinates of the designated unmanned vessel, underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path; and a return-to- shore iteration control module for performing return-to- shore iteration processing on the designated unmanned vessel based on the return-to- shore adaptation feature value of the first trajectory point of each candidate path until the designated unmanned vessel arrives at the target three-dimensional position coordinates of the preset dock.

[0074] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0075] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for automatic return to port of a Beidou AI-based unmanned vessel in the event of a malfunction, characterized in that, Includes the following steps: Real-time acquisition of the current three-dimensional position coordinates of the set unmanned vessel and the target three-dimensional position coordinates of the preset dock, as well as surface point cloud data and underwater point cloud data; Based on the path planning algorithm, and combined with the surface point cloud data and underwater point cloud data of the unmanned vessel, several candidate paths of the unmanned vessel are analyzed, including the three-dimensional coordinates of the trajectory points of several candidate trajectory points. Based on the pre-trained underwater point cloud environment recognition model, and combined with the current three-dimensional position coordinates of the unmanned vessel, underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path, the return-to-home adaptation feature value of the first trajectory point of the corresponding candidate path is analyzed. The unmanned vessel is subjected to return-to-home adaptation feature value based on the first trajectory point of each candidate path, and the return-to-home iterative process is carried out until the unmanned vessel arrives at the preset dock.

2. The method for automatic return to port of a Beidou AI-based unmanned vessel in case of malfunction, as described in claim 1, is characterized in that... The specific steps for analyzing and defining several candidate paths for the unmanned vessel are as follows: The surface point cloud data and underwater point cloud data of the unmanned vessel are input into a pre-trained obstacle recognition model to identify several obstacle areas of the unmanned vessel. Based on the defined obstacle zones of each unmanned vessel, the target endpoint of the unmanned vessel is analyzed and defined. The system reads the current three-dimensional position coordinates of the unmanned vessel, and analyzes several candidate paths for the unmanned vessel based on the path planning algorithm, combined with each obstacle area and the target endpoint.

3. The method for automatic return to port of a Beidou AI-based unmanned vessel in case of malfunction, as described in claim 2, is characterized in that... The obstacle recognition model includes an input layer, an obstacle extraction layer, and an obstacle recognition output layer. The specific steps for recognizing several obstacle regions of the unmanned vessel are as follows: In the input layer of the obstacle recognition model, surface point cloud data and underwater point cloud data of the unmanned vessel are received and preprocessed. In the obstacle extraction layer of the obstacle recognition model, based on the preprocessed surface point cloud data and underwater point cloud data of the designated unmanned vessel, the surface obstacle point cloud set and underwater obstacle point cloud set of the designated unmanned vessel are extracted. In the obstacle recognition output layer of the obstacle recognition model, based on the set surface obstacle point cluster and underwater obstacle point cluster of the unmanned vessel, several obstacle areas of the set unmanned vessel are identified.

4. The method for automatic return to port of a Beidou AI-based unmanned vessel in case of malfunction, as described in claim 2, is characterized in that... The specific steps for setting the target endpoint of the unmanned vessel are as follows: Based on the defined obstacle areas of each unmanned vessel, several candidate safe zones for the unmanned vessel are analyzed. Read the current three-dimensional position coordinates of the designated unmanned vessel and analyze the target endpoint of the designated unmanned vessel in combination with each candidate safe zone.

5. The method for automatic return to port of a Beidou AI-based unmanned vessel in case of malfunction, as described in claim 1, is characterized in that... The underwater point cloud environment recognition model includes a trajectory classification subnetwork, a feature extraction subnetwork, and an output subnetwork. The trajectory classification subnetwork includes an input preprocessing layer and a classification and recognition output layer. The feature extraction subnetwork includes an input extraction layer and an extraction output layer.

6. The method for automatic return to port of a Beidou AI-based unmanned vessel in case of malfunction, as described in claim 5, is characterized in that... The specific steps for analyzing and setting the return-to-home adaptation feature value of the first trajectory point for each candidate path of the unmanned vessel are as follows: The current three-dimensional position coordinates of the unmanned vessel, underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path are input into the pre-trained underwater point cloud environment recognition model. The energy consumption risk feature set of the first trajectory point of the corresponding candidate path is analyzed, including point cloud flow resistance feature value, echo energy attenuation feature value, underwater turbulence feature value, and point cloud distribution heterogeneity feature value. The trajectory adaptation feature value of the first trajectory point of each candidate path of the unmanned vessel is obtained, and the return adaptation feature value of the first trajectory point of the corresponding candidate path is analyzed in combination with the energy consumption risk feature set.

7. The method for automatic return to port of a Beidou AI-based unmanned vessel in case of malfunction, as described in claim 6, is characterized in that... The specific steps for analyzing and defining the energy consumption risk feature set of the first trajectory point for each candidate path of the unmanned vessel are as follows: In the trajectory classification subnetwork of the underwater point cloud environment recognition model, the current three-dimensional position coordinates of the unmanned vessel, the underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path are received, and the underwater point cloud set of the first trajectory point of the corresponding candidate path is extracted. In the feature extraction subnetwork of the underwater point cloud environment recognition model, the point cloud energy consumption feature vector of the first trajectory point of each candidate path of the unmanned vessel is extracted based on the underwater point cloud set of the first trajectory point of each candidate path. In the output subnetwork of the underwater point cloud environment recognition model, the point cloud energy consumption feature vector is based on the first trajectory point of each candidate path of the unmanned vessel.

8. The method for automatic return to port of a Beidou AI-based unmanned vessel in case of malfunction, as described in claim 7, is characterized in that... The specific steps for extracting the underwater point cloud of the first trajectory point for each candidate path of the designated unmanned vessel are as follows: In the input preprocessing layer of the trajectory classification subnetwork, the current three-dimensional position coordinates of the unmanned vessel, the underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path are preprocessed. In the classification and recognition output layer of the trajectory classification subnetwork, based on the preprocessed current three-dimensional position coordinates of the unmanned vessel, underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path, the underwater point cloud set of the first trajectory point of the corresponding candidate path is output.

9. The method for automatic return to port of a Beidou AI-based unmanned vessel in case of malfunction, as described in claim 7, is characterized in that... The specific steps for extracting the point cloud energy consumption feature vector of the first trajectory point of each candidate path of the designated unmanned vessel are as follows: In the input extraction layer of the feature extraction subnetwork, the underwater point cloud of the first trajectory point of each candidate path of the unmanned vessel is received and voxel normalization is performed. In the extraction output layer of the feature extraction subnetwork, the point cloud energy consumption feature vector of the first trajectory point of each candidate path of the unmanned vessel is output based on the underwater point cloud set of the first trajectory point of each candidate path of the unmanned vessel after voxel normalization.

10. A system for automatic return to port of a Beidou AI-based unmanned vessel in the event of a malfunction, employing the method for automatic return to port of a Beidou AI-based unmanned vessel in the event of a malfunction as described in any one of claims 1-9, characterized in that... include: The data acquisition module is used to acquire in real time the current three-dimensional position coordinates of the set unmanned vessel, the target three-dimensional position coordinates of the preset dock, as well as surface point cloud data and underwater point cloud data. The candidate path generation module is used to analyze several candidate paths for the unmanned vessel based on the path planning algorithm and combined with the surface point cloud data and underwater point cloud data of the unmanned vessel, including the three-dimensional coordinates of the trajectory of several candidate trajectory points. The return-to-home adaptation analysis module is used to analyze the return-to-home adaptation feature value of the first trajectory point of the corresponding candidate path based on the pre-trained underwater point cloud environment recognition model, combined with the current three-dimensional position coordinates of the unmanned vessel, underwater point cloud data, and the trajectory three-dimensional coordinates of the first trajectory point of each candidate path. The return-to-go iterative control module is used to perform return-to-go iterative processing on the set unmanned vessel based on the return-to-go adaptation feature value of the first trajectory point of each candidate path, until the set unmanned vessel arrives at the target three-dimensional position coordinates of the preset dock.

Images