Self-adaptive regulation and control system for tension of multiple anchor cables in height limiting section of cutter suction dredger

Abstract

The invention discloses a cutter suction dredger height limit section multi-anchor cable tension self-adaptive regulation and control system, which relates to the technical field of intelligent ocean engineering equipment, and comprises an acquisition module for acquiring ship environment situation data, carrying out timestamp alignment, noise filtering and coordinate system unification, generating a multi-modal environment data packet, a model construction module, a control module and a control module, a three-dimensional dynamic situation field model is constructed by utilizing a potential field modeling engine based on a multi-modal environment data packet, a potential field distribution matrix is generated through a finite difference method, meanwhile, a partial derivative of the potential field distribution matrix is calculated by utilizing central difference, a three-dimensional potential field gradient tensor is generated, and a path planning module is used for planning a path on the basis of the three-dimensional potential field gradient tensor. Candidate path nodes are iteratively generated by adopting an RRT algorithm, smoothing processing is carried out through a cubic spline interpolation algorithm, an initial navigation path is generated, and the tension of the anchor cable can be dynamically adjusted according to actual environment changes by constructing a three-dimensional dynamic situation field model, so that the stability and the safety in the operation process are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent marine engineering equipment, and in particular to a multi-anchor cable tension adaptive control system in a height-limited section of a cutter suction vessel. Background Art

[0002] With the continuous deepening of marine development and port construction, dredger suction dredgers, as an important dredging tool, play an important role in seabed excavation, channel widening and maintenance. In recent years, in order to improve the operating efficiency and safety of dredger suction dredgers in complex environments, a variety of innovative solutions have emerged in related technical fields.

[0003] For example, the real-time environmental monitoring method achieved through sensor fusion can significantly improve the ship's perception of the surrounding environment, and the use of machine learning algorithms to optimize the path planning process can achieve a more efficient and safer navigation route design. In traditional methods, the adjustment of anchor cable tension during operation of dredger suction vessels usually relies on manual adjustment by experienced operators, which is not only inefficient but also difficult to ensure the accuracy and safety of the operation. Although the development of automated control technology has enabled some automated methods to be applied to assist in the regulation of anchor cable tension, they often lack flexibility and cannot adapt to dynamically changing working environments, especially when encountering complex working conditions such as height-restricted sections. In addition, the existing automated control mechanism also has obvious limitations in processing multi-source heterogeneous data integration and achieving accurate path planning. Specifically, the current automated control mechanism lacks the ability to integrate multiple types of data and generate the optimal navigation path based on it. In addition, since most of these methods are based on preset rules or static models and lack the ability to respond to dynamic environmental changes in real time, manual intervention is often required in actual operations, which increases the difficulty and risk of operation.

[0004] There are two main deficiencies in the existing technology. On the one hand, the traditional automatic control mechanism is obviously not adaptable when facing complex marine engineering environments, especially when it comes to operations in height-restricted sections. It cannot automatically adjust the anchor cable tension according to actual conditions, thus affecting the safety and efficiency of the operation. On the other hand, the existing technology lacks the ability to integrate multimodal environmental data, which limits the accuracy of the subsequent generation of the optimal navigation path. Summary of the invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the present invention provides a multi-anchor cable tension adaptive control system in a height-limited section of a dredger suction vessel to solve the problems of insufficient adaptive adjustment capability of anchor cable tension and insufficient fusion of multi-source data.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: The present invention provides a multi-anchor cable tension adaptive control system for a height-limited section of a cutter suction vessel, which comprises a collection module for collecting ship environmental situation data, performing timestamp alignment, noise filtering and coordinate system one, and generating a multi-modal environmental data packet; The model building module uses the potential field modeling engine to build a three-dimensional dynamic potential field model based on the multimodal environment data package, and generates the potential field distribution matrix through the finite difference method. At the same time, the partial derivatives of the potential field distribution matrix are calculated using the central difference to generate the three-dimensional potential field gradient tensor; The path planning module uses the RRT algorithm to iteratively generate candidate path nodes based on the three-dimensional potential field gradient tensor, and then smoothes them through the cubic spline interpolation algorithm to generate the initial navigation path; The instruction generation module inputs the initial navigation path and real-time dredger operation data into the dual-stream neural network, extracts spatial topological features using the spatial stream branch, extracts dynamic correlation features using the temporal stream branch, fuses the spatial topological features and dynamic correlation features, and inputs them into the strategy network to output the adjustment instruction set of the anchor cable tension; In the anchor cable tension control module, the potential field modeling engine parses the anchor cable tension adjustment instruction set, generates a new potential field distribution matrix and three-dimensional potential field gradient tensor, and inputs them into the two-stream neural network. The two-stream neural network performs parameter updates and path replanning to generate the final navigation path.

[0008] As a preferred solution of the multi-anchor cable tension adaptive control system in the height-limited section of the dredger suction vessel described in the present invention, the ship environmental situation data includes three-dimensional point cloud data of the height-limited section, sonar data of underwater obstacles and dredger suction vessel operation data.

[0009] As a preferred solution of the multi-anchor cable tension adaptive control system for the height-limited section of the cutter suction vessel of the present invention, wherein: the generation of the multi-modal environmental data packet specifically includes the following steps: The ship environment situation data is time-stamped and aligned through the PTP protocol. The three-dimensional point cloud data in the height-restricted section is filtered for outliers through the rectangular information granulation algorithm. The sonar data of underwater obstacles is denoised using the wavelet threshold. The operation data of the cutter suction vessel is clustered using the DBSCAN algorithm to remove outliers. The ICP algorithm is used to unify the preprocessed ship environmental situation data into the geographic coordinate system, and a hierarchical fusion strategy is adopted to integrate them to generate a multimodal environmental data package.

[0010] As a preferred solution of the multi-anchor cable tension adaptive control system in the height-limited section of the cutter suction vessel of the present invention, the potential field distribution matrix is ​​generated by the finite difference method, which specifically includes the following steps: Based on the multimodal environment data package, a standardized 3D voxel map is constructed using the point cloud voxelization downsampling method, and a spatial index is established through the KD-Tree. At the same time, a 3D dynamic potential field model is constructed through a dynamic weight allocation strategy combined with a hyperbolic secant function. Based on the three-dimensional dynamic potential field model, the spatial gradient field is calculated by the finite difference method to obtain the gravitational gradient and repulsive field strength. At the same time, based on the gravitational gradient and repulsive field strength, the potential field distribution matrix is ​​generated by the multi-physical field coupling superposition algorithm.

[0011] As a preferred solution of the multi-anchor cable tension adaptive control system in the height-limited section of the cutter suction vessel of the present invention, wherein: the generating of the three-dimensional potential field gradient tensor specifically comprises the following steps: The first-order partial derivative of the potential field distribution matrix is ​​calculated by using the central difference in the finite difference method, and the initial gradient vector field is obtained by using the vector synthesis algorithm based on the first-order partial derivative; The initial gradient vector field is reconstructed by the ADMM coupled reconstruction method, the second-order partial derivatives are cross-calculated by the central difference method, and the weight distribution ratio of the second-order partial derivatives is dynamically adjusted using the hyperbolic secant function. At the same time, regularization is performed by the alternating direction multiplier method to generate a three-dimensional potential field gradient tensor.

[0012] As a preferred solution of the multi-anchor cable tension adaptive control system in the height-limited section of the cutter suction vessel of the present invention, the specific steps of generating the initial navigation path are as follows: The RRT algorithm is used to perform node expansion on the three-dimensional potential field gradient tensor to obtain the dynamic direction vector, and the dynamic direction vector is converted into the path expansion direction through the elliptical sampling domain conversion method. At the same time, the node expansion step size is obtained by using the adaptive step size strategy. Generate candidate path nodes through a hierarchical search strategy according to the path expansion direction and node expansion step; The candidate path nodes are processed by the cubic spline interpolation algorithm to generate a cubic spline smooth trajectory; Based on the cubic spline smooth trajectory, the ADMM algorithm is used for dynamic regularization optimization to generate the initial navigation path.

[0013] As a preferred solution of the multi-anchor cable tension adaptive control system in the height-limited section of the cutter suction vessel of the present invention, wherein: the spatial flow branch is used to extract the spatial topological features, and the specific steps are as follows: The initial navigation path is input into the spatial stream branch of the two-stream neural network. The first level of the spatial stream branch performs time-space joint convolution and spatial dimension reduction to capture spatial structural features. The second level performs local iterative optimization and cross-layer feature fusion to obtain refined local features. The third layer dynamically assigns weights of different scales to spatial structural features and refined local features through the channel attention mechanism, and fuses them through splicing channels to generate spatial topological features.

[0014] As a preferred solution of the multi-anchor cable tension adaptive control system in the height-limited section of the cutter suction vessel of the present invention, wherein: the dynamic correlation features are extracted by using the time stream branch, and the specific steps are as follows: The real-time cutter suction vessel operation data is input into the time stream branch of the two-stream neural network, where the forward TCN of the bidirectional long short-term memory network captures the long-range trend through dilated convolution, and the backward TCN models the reverse degradation pattern through causal convolution; The long-range trends and reverse degradation patterns are bidirectionally spliced ​​through the temporal attention mechanism to generate dynamic correlation features.

[0015] As a preferred solution of the multi-anchor cable tension adaptive control system in the height-limited section of the cutter suction vessel of the present invention, the specific steps of the output anchor cable tension adjustment instruction set are as follows: The spatial topological features and dynamic temporal features are weightedly fused and input into the DDPG strategy network. The Actor-Critic collaborative optimization framework is used to perform policy gradient update to generate the anchor cable tension adjustment parameter vector. Based on the anchor cable tension adjustment parameter vector, the finite element algorithm is used to generate the anchor cable tension adjustment instruction set.

[0016] As a preferred solution of the multi-anchor cable tension adaptive control system for the height-limited section of the cutter suction vessel of the present invention, the specific steps of generating the final navigation path are as follows: Based on the adjustment instruction set of anchor cable tension, a new potential field distribution matrix is ​​generated through the finite element tension field coupling algorithm, and the implicit Euler method is used for nonlinear calculation to generate a new three-dimensional potential field gradient tensor; The new potential field distribution matrix and the three-dimensional potential field gradient tensor are input into the two-stream neural network, and the weight parameters are updated using the elastic weight solidification regularization method to generate an optimized two-stream neural network. Based on the optimized two-stream neural network, the initial navigation path deviation is corrected through the sliding mode control law to generate the final navigation path.

[0017] The beneficial effects of the present invention are as follows: a three-dimensional dynamic potential field model is generated by using a potential field modeling engine, and based on this, an accurate three-dimensional potential field gradient tensor is generated, ensuring that the anchor cable tension can be dynamically adjusted according to actual environmental changes, thereby improving stability and safety during the operation. Secondly, through efficient timestamp alignment, noise filtering, and coordinate system preprocessing steps, various types of environmental data (such as three-dimensional point cloud data, underwater obstacle sonar data, and dredger operation data) are seamlessly integrated to generate high-quality multimodal environmental data packets. These high-precision data provide a solid foundation for subsequent path planning, making the generated navigation path more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0019] Figure 1 Schematic diagram of the multi-anchor cable tension adaptive control system in the height-restricted section of a cutter suction dredger.

[0020] Figure 2 Flowchart for generating an initial navigation path based on the three-dimensional potential field gradient tensor.

[0021] Figure 3 This is a flowchart of the working principle of the path planning module.

[0022] Figure 4 Schematic diagram of the multi-anchor cable tension adaptive control system architecture. DETAILED DESCRIPTION

[0023] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0024] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figure 1~Figure 4 This embodiment provides a multi-anchor cable tension adaptive control system for a height-limited section of a cutter suction vessel, comprising the following steps: The acquisition module collects the ship's environmental situation data, performs timestamp alignment, noise filtering and coordinate system one, and generates a multi-modal environmental data packet.

[0027] The specific operations include the following: First, the ship's environmental situation data is collected, including the three-dimensional point cloud data of the height-restricted section, the sonar data of underwater obstacles, and the operation data of the dredger. The laser radar scanning equipment is installed on the dredger to scan the surrounding environment in real time, and collects the three-dimensional point cloud data of the height-restricted section by emitting high-frequency laser beams and receiving reflected echo signals; The multi-beam sonar unit is installed on the bottom of the dredger to collect sonar data of underwater obstacles by emitting multiple sound beams underwater at the same time and receiving echo signals returned from underwater objects; The operation data of the dredger includes the engine speed, fuel consumption, wheel speed, ship speed and position data. The engine speed and fuel consumption are collected through the on-board diagnostic unit (OBD), the wheel speed and ship speed are collected through the wheel speed sensor, and the ship position is collected through the GPS device. After the acquisition is completed, the PTP protocol is used to align the timestamps of the ship's environmental situation data. In the specific operation, the PTP hardware clock synchronization unit is deployed in the multi-sensor unit, and the master-slave time reference architecture is adopted. Based on the PTP protocol, the multi-sensor units are synchronized by using message interaction and round-trip delay measurement. After the synchronization is completed, the linear interpolation algorithm and the cubic spline interpolation algorithm are used to achieve cross-frequency domain timestamp alignment. At the same time, the clock offset compensation mechanism and the abnormal clock jump detection algorithm are integrated to perform jump detection and dynamic time window sliding verification to ensure the timestamp synchronization of the ship's environmental situation data; Next, the ship environment situation data is subjected to noise filtering. In the specific operation, for the 3D point cloud data of the height-limited section, the rectangular information granulation algorithm is used to divide the 3D point cloud data into multiple small rectangular areas, and the statistical characteristics of the points in each area are calculated. Based on the statistical characteristics, the outliers are identified by the Grubbs criterion and eliminated by the information granule similarity analysis, thereby improving the data quality. For the sonar data of underwater obstacles, the wavelet threshold denoising method is used to analyze the signal characteristics of the sonar data at different scales through wavelet decomposition, and the denoising threshold is set through the adaptive threshold adjustment mechanism. The high-frequency noise exceeding the denoising threshold is removed through the soft threshold function to retain useful information. For the operation data of the cutter suction vessel, the DBSCAN clustering algorithm is used to generate the main data clusters through density accessibility analysis, and the abnormal points far away from the main data clusters are effectively identified through the quantitative analysis of abnormality, and removed through the dynamic sliding window filtering mechanism to ensure the accuracy and reliability of the data; After the noise filtering is completed, the pre-processed ship environment situation data is unified into the geographic coordinate system for subsequent fusion and analysis. In the specific operation, based on the ICP algorithm, the FPFH feature descriptor is used to construct the point cloud local feature space, and the Tukey weight function is used to remove the erroneous corresponding point set caused by occlusion or noise in the point cloud local feature space to find the best correspondence between the two corresponding point sets. According to the best correspondence, the position of one corresponding point set is adjusted through SVD decomposition to make it coincide with the other corresponding point set. After the adjustment is completed, the ship environment situation data is unified into the WGS-84 geographic coordinate system through the seven-parameter coordinate transformation; Next, the processed ship environment situation data is fused to generate a multimodal environmental data packet. In the specific operation, a hierarchical fusion strategy is adopted for integration: the spatial index of the three-dimensional point cloud data in the height-restricted section is constructed by an octree, and the compressed point cloud data is obtained by aggregating the features within the point cloud voxels; the sonar data of underwater obstacles are enhanced by the wavelet packet reconstruction method, and the echo delay and intensity characteristics are extracted by time domain peak detection and energy integral calculation; the Kalman filter is applied to the ship operation data for state estimation to obtain the optimal estimate of the ship's state; the compressed point cloud data, echo delay and intensity characteristics, and the optimal estimate of the ship's state are encapsulated in the JSON-LD format for multi-source data encapsulation, and the Zstd compression algorithm is used to generate a standardized multimodal environmental data packet, which is then verified by the SHA-256 hash to ensure data integrity.

[0028] The model building module, based on the multimodal environment data package, uses the potential field modeling engine to build a three-dimensional dynamic potential field model, and generates the potential field distribution matrix through the finite difference method. At the same time, it uses the central difference to calculate the partial derivatives of the potential field distribution matrix to generate a three-dimensional potential field gradient tensor.

[0029] The specific operations include the following: Based on the multimodal environment data package, the point cloud voxelization downsampling method is used to construct a three-dimensional voxel map, and the spatial index is established through KD-Tree. In the specific operation, first, the high-density point cloud data is extracted from the multimodal environment data package through the JSON-LD parser, and the point cloud voxelization downsampling method is used to convert the high-density point cloud data into a uniformly distributed voxel grid through voxel grid division and centroid feature retention, thereby reducing the data volume and improving the efficiency of subsequent processing; Based on the uniformly distributed voxel grid, octree recursive segmentation is used to perform null value voxel pruning and multi-resolution LOD generation to build a 3D voxel map. Based on KD-Tree, median axis selection segmentation and BBF search optimization are used to establish a spatial index for the 3D voxel map, so as to quickly query and access data in a specific area. KD-Tree can quickly locate target points in large-scale data sets, greatly improving the speed of data retrieval. Next, based on the 3D voxel map, a 3D dynamic potential field model is constructed by using a dynamic weight allocation strategy combined with a hyperbolic secant function. In the specific operation, first, the 3D voxel map is subjected to multi-scale spatial convolution through the Euclidean distance transformation algorithm to obtain the obstacle distance; secondly, the 3D voxel map is subjected to mixed clustering through the Bayesian probability inference algorithm to obtain the channel probability distribution; finally, the 3D voxel map is subjected to dynamic tracking of the optical flow field through a feature extraction algorithm to extract the dynamic target clustering features; The hyperbolic secant function is used to quantify the obstacle distance through the adaptive attenuation coefficient to generate the obstacle repelling potential field; the gradient field mapping is used to normalize the channel probability distribution through the probability density function to generate the channel potential field; the time sliding window method is used to process the dynamic target clustering characteristics through exponential weighting to obtain the dynamic target influence domain; It should be noted that the adaptive attenuation coefficient is defined based on the obstacle distance and in combination with the hyperbolic secant function, and its value range is usually [0.01, 1.5]; Based on the obstacle rejection potential field, channel potential field and dynamic target influence domain, the initial environment feature set is constructed by splicing feature tensors, and the entropy weight method is used to perform inverse weighted calculation to obtain the static weight of the initial environment feature set. At the same time, the BPTT algorithm is used to obtain the dynamic weight of the initial environment feature set through time gradient back propagation. The improved Log-Sum-Exp function is used to fuse the static weights and dynamic weights into multiple potential fields to generate a continuous and differentiable three-dimensional dynamic potential field. Based on the Lyapunov exponent, the stability of the three-dimensional dynamic potential field is verified by calculating the divergence rate of the phase space trajectory. Based on the three-dimensional dynamic potential field, a real-time maintenance mechanism based on ROS nodes is established, and incremental KD-Tree is used to quickly respond to environmental changes. At the same time, OpenGL visual rendering is used for verification to form a three-dimensional dynamic potential field model that is both distance sensitive and environmentally adaptive. The method combining dynamic weight allocation strategy and hyperbolic secant function not only improves the adaptability of the three-dimensional dynamic potential field model, but also enhances its stability in complex environments. Based on the constructed three-dimensional dynamic potential field model, the potential field distribution matrix is ​​generated by the finite difference method. In the specific operation, based on the finite difference method, grid points are established by the equally spaced voxel division method, including the center point and 26 neighborhood points, and based on the central difference format, the three-directional synchronous gradient solution is performed on each grid point to obtain the potential field gradient; Tikhonov regularization is used to deal with boundary singularities in the potential field gradient, and Jacobi iteration method is used to separate and solve the Poisson equation of the potential field gradient to generate the gravitational gradient. At the same time, the potential field gradient is convoluted and filtered through Gaussian kernel density estimation to obtain the repulsive field strength. The multi-physics field coupling superposition algorithm is used to perform Log-Sum-Exp weighted fusion of the gravitational gradient and the repulsive field strength to generate a potential field distribution matrix. Based on the potential field distribution matrix, a three-dimensional potential field gradient tensor is generated. In the specific operation, the central difference in the finite difference method is used to perform synchronous calculations in three directions. After performing two calculations, weighted average is performed to generate an initial difference result. Based on the initial difference result, error correction is performed through Richardson extrapolation compensation to generate the first-order partial derivative of the potential field distribution matrix. The quaternion spherical linear interpolation method is used to correct the directional continuity of the first-order partial derivatives to generate directional gradient components. The directional gradient components are then subjected to noise suppression and outlier removal through Gauss-Seidel smoothing filtering. At the same time, the directional gradient components are superimposed through a vector synthesis algorithm to obtain an initial gradient vector field, which describes the potential field change trend at each position in the environment. In order to further improve the quality of the initial gradient vector field, the ADMM coupled reconstruction method is used to reconstruct it. In the specific operation, the ADMM coupled reconstruction method is used to remove the noise in the initial gradient vector field through alternating direction multipliers, and the mixed regularization based on the Huber function is used to remove the irregular parts in the initial gradient vector field, making it smoother and more accurate, so as to generate the reconstructed gradient vector field. Based on the reconstructed gradient vector field, the central difference method is continued to be used to cross-calculate the second-order partial derivatives. In the specific operation, the Hessian matrix elements of the reconstructed gradient vector field are calculated based on the central difference method to generate the original second-order differential result. Based on the original second-order differential result, the truncation error is eliminated through Richardson extrapolation compensation to generate the second-order partial derivative. The use of the central difference method not only improves the accuracy of gradient calculation, but also enhances the ability to capture complex terrain features. Next, the hyperbolic secant function is used to dynamically adjust the weight distribution ratio of the second-order partial derivative. In the specific operation, the hyperbolic secant function is used to re-weight the original second-order differential result to generate an adaptive second-order differential field. Based on the adaptive second-order differential field, the feature space projection is performed through principal component analysis to redistribute the weight ratio of the second-order partial derivative, thereby realizing dynamic weight adjustment based on terrain characteristics. Finally, the alternating direction multiplier method is used for regularization to generate the final three-dimensional potential field gradient tensor. In the specific operation, the enhanced ADMM algorithm is used to perform 50 iterations, and the residual noise is eliminated by the mixed norm constraint. At the same time, the Schur complement decomposition is used to accelerate the inversion operation of the second-order partial derivatives. The final three-dimensional potential field gradient tensor is generated. The three-dimensional potential field gradient tensor is processed by the elastic network regularization method to prevent overfitting and improve the generalization ability, ensuring that the generated three-dimensional potential field gradient tensor has high reliability and stability in practical applications.

[0030] The path planning module, based on the three-dimensional potential field gradient tensor, uses the RRT algorithm to iteratively generate candidate path nodes, and then smoothes them through the cubic spline interpolation algorithm to generate the initial navigation path.

[0031] The specific steps include: The RRT algorithm is used to perform node expansion on the three-dimensional potential field gradient tensor to obtain the dynamic direction vector. In the specific operation, first, the RRT algorithm is used to randomly select target points in the three-dimensional potential field gradient tensor through Gaussian distribution sampling, and the target points are progressively expanded using adaptive step size control to generate an initial path tree. Based on the initial path tree, the dynamic direction vector is obtained by the vector field integral method; Next, the dynamic direction vector is converted into the path extension direction through the ellipse sampling domain conversion method. In the specific operation, based on the dynamic direction vector, the equidistant spherical projection is performed through the unit sphere mapping to construct the initial unit circle sampling domain, and the distribution characteristics of the dynamic direction vector are extracted through principal component analysis (PCA). According to the distribution characteristics, the axial variability is calculated through eigenvalue decomposition to obtain the major and minor axis ratio of the ellipse. The specific mathematical formula is as follows: ; in, represents the ratio of the major and minor axes of the ellipse, represents the maximum distribution eigenvalue extracted by principal component analysis, represents the second largest distribution eigenvalue extracted by principal component analysis; It should be noted that axial variability refers to the degree of discreteness of the distribution of the dynamic direction vector in three-dimensional space. Its core function is to guide the morphological construction of the elliptical sampling domain by quantifying the spatial distribution characteristics of the dynamic direction vector. According to the ratio of the major and minor axes of the ellipse, the unit circle sampling domain is converted into the ellipse sampling domain by the affine transformation method. Then, quasi-Monte Carlo sampling is used to perform hierarchical importance sampling in the ellipse sampling domain to generate a set of candidate directions. According to the set of candidate directions, the third-order control points are optimized by the Bézier curve interpolation method to generate continuous path segments. Based on the continuous path segments, the path extension direction is generated by Frenet frame projection. The application of the ellipse sampling domain conversion method makes the path extension direction more reasonable and flexible. At the same time, an adaptive step strategy is used to obtain the node expansion step. In the specific operation, the potential field strength is normalized by the hyperbolic tangent function based on the dynamic direction vector to generate a benchmark step. Based on the benchmark step, dynamic attenuation is performed through the local curvature feedback mechanism to generate the node expansion step. The PID controller is used to smooth the node expansion step in real time through incremental error adjustment to ensure that the expansion process meets the dynamic constraints and maintains the exploration effect. The application of the adaptive step strategy ensures that the path expansion is both efficient and safe. Then, according to the path expansion direction and node expansion step, the candidate path nodes are generated through the hierarchical search strategy. In the specific operation, the hierarchical search strategy uses the spatial multi-resolution decomposition algorithm to perform dynamic hierarchical division to build a three-level search architecture, including a coarse-grained search layer, a medium-grained optimization layer, and a fine-grained verification layer. Among them, the coarse-grained search layer uses the octree space partitioning algorithm to screen the feasibility of the path extension direction, and uses the potential field intensity threshold segmentation method to perform binary marking to obtain the high-risk area for obstacle collision. At the same time, the AABB collision detection algorithm is used to eliminate the high-risk area for obstacle collision and generate a primary feasible area set; based on the primary feasible area set, the Dijkstra algorithm is used to search for the minimum potential field integral path to generate the initial candidate node set; The medium-granularity optimization layer uses the Metropolis-Hastings sampling method guided by the potential field gradient to perform probability re-screening based on the initial candidate node set to generate an optimized node distribution. Based on the optimized node distribution, B-spline interpolation is used for local smoothing optimization to obtain curvature-constrained path segments. The fine-grained verification layer verifies node accessibility based on curvature-constrained path segments using a parallelized ray detection method through multi-resolution voxel penetration testing, and evaluates node dynamic feasibility through time-domain trajectory simulation in combination with a dynamic window method, and finally outputs a set of candidate path nodes. The hierarchical search strategy can quickly exclude obviously infeasible path options at a macro scale, and avoid potential obstacles and optimize path smoothness at a micro scale. After the candidate path nodes are generated, they are processed by the cubic spline interpolation algorithm to generate a cubic spline smooth trajectory. In the specific operation, firstly, based on the candidate path node set, the parameterized mapping relationship between the candidate path nodes is constructed by the regularized least squares method through least squares fitting, and the tridiagonal matrix solution method is used to obtain the spline coefficient matrix through Thomas calculation; then, the piecewise continuity constraint algorithm is used to make the first-order and second-order partial derivatives continuous by using the C² continuity boundary condition to achieve the optimization of the spline coefficient matrix; Based on the optimized spline coefficient matrix, the L-BFGS optimizer is used to adjust the position of the candidate path nodes. After the adjustment, the Fritsch-Carlson monotonicity preserving algorithm is used to generate a cubic spline smooth trajectory through piecewise cubic Hermite interpolation. At the same time, the adaptive node encryption method is used to eliminate RANSAC outliers to ensure the accuracy of the cubic spline smooth trajectory. Based on the cubic spline smooth trajectory, the ADMM algorithm is used for dynamic regularization optimization to generate the initial navigation path. In the specific operation, the cubic spline trajectory is first discretized into a dense path point set through equal arc length sampling, and the optimization objective function is constructed through the multi-objective weighted fusion method. Based on the optimization objective function, the alternating direction multiplier method is used to iteratively solve: Firstly, the curvature-optimized path segments are generated by accelerating the curvature calculation through Chebyshev polynomial approximation using the projected gradient method; Secondly, based on the curvature optimization path segment, the proximal gradient method is used to deal with non-smooth terms, and the Nesterov acceleration method is combined to improve the convergence speed of the potential field adapter to obtain the potential field compatible path; Finally, based on the potential field compatible path, the dynamic constraints are solved by quadratic programming to obtain a dynamically feasible solution. In each round of iteration of the alternating direction multiplier method, the condition number of the dynamically feasible solution is optimized by the Schur complementary decomposition method to generate a coordinated global solution. Based on the coordinated global solution, the initial navigation path is generated by the B-spline reparameterization method. The application of the ADMM algorithm can effectively solve large-scale optimization problems and ensure that the generated initial navigation path has high feasibility and safety.

[0032] The instruction generation module inputs the initial navigation path and real-time dredger operation data into the dual-stream neural network, uses the spatial stream branch to extract the spatial topological features, uses the time stream branch to extract the dynamic correlation features, fuses the spatial topological features and the dynamic correlation features, and inputs them into the strategy network to output the adjustment instruction set of the anchor cable tension.

[0033] The specific steps include: First, a two-stream neural network is constructed and trained. In the specific operation, a heterogeneous two-branch architecture is constructed: the spatial stream branch uses ResNet-50 as the backbone network, initialized by loading ImageNet pre-trained weights, and specifically processes static potential field features; the temporal stream branch uses a 3D ConvNet structure and uses a 10-frame optical flow sequence generated by the TV-L1 algorithm as input to capture dynamic potential field features; The training process adopts a three-stage alternating training strategy: in the first stage, the spatial stream branch is trained, the cross entropy loss function is used to classify multi-category potential field features, and the Adam optimizer is used to fine-tune on a multi-sample static potential field data set. During this period, the statistics of the batch normalization layer are frozen to keep the feature distribution stable. The statistics of the batch normalization layer refer to the mean and variance dynamically calculated by the sliding average method based on the input data of each small batch during the training process; In the second stage, the time stream branch is trained, triplet loss is used for temporal feature embedding learning, and the SGD optimizer with momentum is used to train on the dynamic potential field features of multiple samples. At the same time, three-level temporal pyramid pooling is used to enhance the temporal feature expression ability. In the third stage, the two-stream features are concatenated through a learnable adaptive fusion layer, and a graded learning rate strategy is used for end-to-end fine-tuning. The fusion feature classification results are output, and an early stopping mechanism is set to prevent overfitting of the two-stream neural network. After the two-stream neural network training is completed, the initial navigation path is input into the two-stream neural network, and the spatial topological features are extracted using the spatial stream branch. In the specific operation, the initial navigation path is converted into a three-dimensional terrain tensor through a multi-source data fusion method, and the three-dimensional terrain tensor is input into the spatial stream branch of the two-stream neural network through three-dimensional convolution; The spatial stream branch includes three levels of feature extraction, a cross-scale feature fusion layer, and a channel attention enhancement layer. The first level of feature extraction uses a three-dimensional convolution layer to perform spatial context-aware convolution to generate a primary feature map. Based on the primary feature map, multi-scale spatial downsampling is performed through a maximum pooling layer to capture spatial structural features. The temporal-spatial joint convolution performs temporal dimension convolution on the spatial structural features, so that the spatial stream branch can better understand the spatial distribution of the path and its changing trend over time. The spatial dimensionality reduction reduces the dimension of the spatial structural features through a 1×1×1 convolution, thereby improving computational efficiency while retaining key information. The second-level feature extraction layer uses a three-dimensional convolution layer to perform local feature refinement convolution to generate a mid-level feature map. Based on the mid-level feature map, cross-layer feature enhancement is performed through a residual connection layer to obtain refined local features. The local iterative optimization method optimizes the refined local features multiple times through gradient accumulation to improve the quality of the features. The cross-layer feature fusion uses a channel attention gating mechanism to perform adaptive weighted fusion of the refined local features to generate a richer feature representation. The third-level feature extraction layer uses the global pooling layer to perform feature compression aggregation to generate a global context descriptor; based on the global context descriptor, high-order semantic encoding is performed through the fully connected layer to obtain a compact feature representation; based on the compact feature representation, the spatial structure features and refined local features are dynamically assigned weights of different scales through the channel attention mechanism. In the specific operation, firstly, global average pooling is used for the spatial structure features to generate global channel statistics; 3×3 depth-separable convolution is used for the refined local features to generate local channel statistics; based on the global channel statistics and local channel statistics, two layers of fully connected layers are used to assign weights of different scales to the spatial structure features and refined local features; Based on different scale weights, spatial structural features and refined local features are fused through splicing channels and feature recalibration technology to generate spatial topological features. Among them, the channel attention mechanism can automatically adjust the weight of each channel according to the importance of the feature, thereby enhancing important features and suppressing irrelevant features; At the same time, the real-time operation data of the cutter suction vessel is input into the time stream branch to extract the dynamic correlation features. In the specific operation, the time series tensor is constructed based on the real-time operation data of the cutter suction vessel using the time-frequency synchronous coding method. Next, based on the time series tensor, a bidirectional long short-term memory network is constructed. In the specific operation, the time series tensor is used to construct the forward TCN (temporal convolutional network) and the backward TCN through dilated causal convolution. Each TCN layer contains 32 LSTM units; each LSTM unit contains a gating mechanism for peephole connection; the forward layer captures long-range trends through multi-layer dilated convolution, and the backward layer lags the correlation pattern through reverse time convolution; A three-stage optimization strategy is adopted in the training phase: in the pre-training phase, training samples are generated using the Wasserstein generative adversarial network based on the time series tensor; In the main training phase, the temporal focus loss function is used to dynamically weight the training samples for difficult samples, and the Adam optimizer is used to iterate the training samples for multiple rounds. The gradient clipping method is applied in each iteration to prevent gradient explosion. In the fine-tuning stage, the fully connected layer is used to optimize the hybrid attention mechanism and output the trained bidirectional long short-term memory network.

[0034] Furthermore, the trained bidirectional long short-term memory network is used to extract dynamic correlation features: the forward TCN includes a dilated convolution layer and a gated activation layer. The dilated convolution layer generates multi-scale temporal features through hole convolution. Based on the multi-scale temporal features, the long-range trend is captured through residual connection to reduce the risk of gradient vanishing. The backward TCN includes a causal convolution layer and a reverse gating layer. The causal convolution layer generates reverse time series features through time reverse convolution. Based on the reverse time series features, the degradation pattern detector models the reverse degradation pattern to identify potential ship performance degradation. Through the combination of the forward TCN and the backward TCN, not only can the long-term trend in the data be identified, but also its potential reverse change pattern can be analyzed, ensuring that the provided time series features are more comprehensive and accurate. Then, the long-term trend and reverse degradation pattern are input into the spatiotemporal fusion layer, and the temporal attention mechanism is used for feature fusion to generate dynamic correlation features. In the specific operation, the spatiotemporal fusion layer performs time alignment and bidirectional splicing of the long-term trend features and reverse degradation pattern features, and inputs them into the temporal attention mechanism for processing; the temporal attention mechanism calculates the interaction relationship between the query, key and value matrices through the multi-head attention algorithm to obtain the time-dependent features; the time-dependent features are residually connected through the dilated time convolution layer, and the dynamic correlation features are output; the whole process is optimized with a pre-layer normalization structure and learnable fusion weights, which significantly improves the feature similarity while ensuring the training stability. Among them, the application of the temporal attention mechanism highlights the data of important time periods, so that the generated dynamic correlation features are more representative; After the extraction of spatial topological features and dynamic association features is completed, the two are weighted and fused and then input into the DDPG strategy network to output the anchor cable tension adjustment instruction set. In the specific operation, the spatial topological features and dynamic association features are fused through the adaptive feature gating fusion mechanism to generate a 1152-dimensional joint feature vector, which is then input into the DDPG strategy network through the feature normalization method. Based on the DDPG policy network, the Actor-Critic collaborative optimization framework is called through the deep deterministic policy gradient algorithm. The Actor part in the Actor-Critic collaborative optimization framework uses a three-layer fully connected network to perform Gaussian policy parameterization to generate an initial action distribution. The Tanh activation function is used to scale the initial action distribution through dynamic action boundary mapping to generate an anchor cable tension adjustment strategy. The Critic part uses a dual four-layer fully connected network combined with a Q-value function to evaluate the quality of the tension adjustment strategy. For example, the Critic part calculates a Q value based on the action distribution generated by the Actor through the Bellman optimal equation. The Q value reflects the expected return of taking the action in the current state. If the Q value exceeds 0.8, it means that the tension adjustment strategy performs well in the current environment; conversely, if the Q value is lower than 0.5, it means that the tension adjustment strategy needs to be improved. Based on the anchor cable tension adjustment strategy after Critic evaluation, the Ornstein-Uhlenbeck process method is used to perform action exploration and strategy optimization to generate the anchor cable tension adjustment parameter vector. The ADDPG strategy network combines the advantages of the value function method and the policy gradient method, and can efficiently learn the optimal strategy in the continuous action space. Based on the anchor cable tension adjustment parameter vector, the anchor cable tension adjustment instruction set is generated through the finite element algorithm. In the specific operation, the nonlinear finite element solver (ANSYS APDL) is used to perform transient dynamic simulation on the anchor cable tension adjustment parameter vector to calculate the force distribution of each node; based on the force distribution, the anchor cable tension adjustment parameter vector is iteratively adjusted through the constrained optimization algorithm, and the stability of the anchor cable tension adjustment parameters is verified through Lyapunov exponent calculation; based on the iteratively adjusted anchor cable tension adjustment parameters, the anchor cable tension adjustment instruction is output through cubic spline interpolation and PID controller.

[0035] In the anchor cable tension control module, the potential field modeling engine parses the anchor cable tension adjustment instruction set, generates a new potential field distribution matrix and three-dimensional potential field gradient tensor, and inputs them into the two-stream neural network. The two-stream neural network performs parameter updates and path replanning to generate the final navigation path.

[0036] The specific steps include: The potential field modeling engine parses the adjustment instruction set of the anchor cable tension to generate a new potential field distribution matrix and a three-dimensional potential field gradient tensor. In the specific operation, based on the adjustment instruction set of the anchor cable tension, the anchor cable tension is discretized into 50 Timoshenko beam units through a nonlinear finite element solver, and the finite element tension field coupling algorithm is used to perform bidirectional coupling of the structural field and the fluid field on the Timoshenko beam unit. At the same time, the Fixed-Stress splitting algorithm is used to ensure the coupling stability. After the coupling is completed, the displacement-tension coupling field data is extracted through the finite element algorithm, and the displacement-tension coupling field data is visualized and verified through the Paraview post-processing software. Based on the displacement-tension coupling field data after verification, the Helmholtz equation solver based on the FEniCS framework is used to perform potential energy redistribution calculations: first, Gmsh is used to generate an unstructured tetrahedral mesh through the Delaunay triangulation algorithm; secondly, the GMRES iterative solver of the PETSc linear algebra library is used to solve the Helmholtz equation on the unstructured tetrahedral mesh to generate a node potential energy distribution field; then, based on the node potential energy distribution field, the potential energy gradient is automatically extracted through a Python script; finally, based on the potential energy gradient, the potential field distribution matrix is ​​generated through a bicubic spline interpolation algorithm; Next, the implicit Euler method is used to generate a new three-dimensional potential field gradient tensor. In the specific operation, based on the adjustment instruction set of the anchor cable tension, the implicit Euler method is used to numerically solve the nonlinear potential field evolution equation: first, in the FEniCS framework, the discrete weak form equation is established through the variational form conversion method; based on the discrete weak form equation, the SNES nonlinear solver of PETSc is used to process the nonlinear terms and obtain the numerical solution; based on the numerical solution, the Krylov subspace iterative calculation is accelerated by the ILU (3) preprocessing technology, and the transient potential field distribution is output; Based on the instantaneous potential field distribution, a new three-dimensional potential field gradient tensor is calculated by the structured central difference method: first, a difference template is constructed by the grid step parameterization tool; then, the Scharr operator is used to calculate the isotropic gradient of the difference template to obtain preliminary gradient field data; secondly, based on the preliminary gradient field data, CUDA parallel computing is used to achieve grid point level parallel processing to generate a new three-dimensional potential field gradient tensor, which is stored in HDF5 format for subsequent analysis; After the new potential field distribution matrix and the three-dimensional potential field gradient tensor are input into the two-stream neural network, the weight parameters are updated using the elastic weight solidification regularization method to generate an optimized two-stream neural network. In the specific operation, the new potential field distribution matrix and the three-dimensional potential field gradient tensor are input into the two-stream neural network through the data preprocessing pipeline; The two-stream neural network uses the elastic weight solidification (EWC) regularization method to update the weight parameters. The specific process is as follows: first, the potential field distribution matrix and the three-dimensional potential field gradient tensor are input into the spatial stream branch and the temporal stream branch respectively; through the feature fusion layer of the two-stream neural network, they are fused through the cross-modal attention mechanism to obtain the joint feature vector; then, based on the real-time ship environment situation data, the loss is calculated through the cross entropy loss function and the triplet loss function to generate a comprehensive loss value; then, based on the comprehensive loss value, the weight update of the joint feature vector is constrained by the EWC regularization term method to construct a constrained optimization target; based on the constrained optimization target, the weight of the spatial stream branch is updated by layered gradient clipping, and the weight of the temporal stream branch is updated by temporal gradient normalization; Finally, the Adam optimizer is used for back-propagation update: the spatial stream branch uses layered weight solidification to freeze the differentiated parameters; the temporal stream branch implements timing-sensitive solidification to protect motion features; the entire weight parameter update process is implemented under the PyTorch framework, and automatic mixed precision (AMP) is used to accelerate training. The resulting optimized two-stream neural network can not only improve the spatial perception ability of complex environments, but also enhance the capture and response efficiency of dynamic temporal features. Based on the optimized two-stream neural network, the final navigation path is generated. In the specific operation, the initial navigation path is first input into the spatial stream branch to extract 128-dimensional spatial topological features, and the operation data of the dredger is input into the temporal stream branch to generate 64-dimensional dynamic correlation features. The two types of features are fused into a 192-dimensional joint feature vector through the gated attention mechanism; based on the joint feature vector, the improved RRT* algorithm is used to generate a candidate path set through dynamic constraint sampling; the MLP layer is used to verify the safety of the candidate path set through obstacle distance field analysis, and the Sigmoid output layer is used to perform nonlinear probability weighting on the candidate path set through cubic polynomial regression to verify the path safety; finally, the MPC framework is used for path tracking optimization, and the B-spline curve path and control instruction sequence are generated through the adaptive weight adjustment mechanism; based on the B-spline curve path and control instruction sequence, the path is tracked and adjusted through the modular control law to generate the final navigation path.

[0037] In summary, the present invention generates a three-dimensional dynamic potential field model through a potential field modeling engine, and generates an accurate three-dimensional potential field gradient tensor based on the model, thereby ensuring that the anchor cable tension can be dynamically adjusted according to actual environmental changes, thereby improving stability and safety during the operation. Secondly, through efficient timestamp alignment, noise filtering, and coordinate system preprocessing steps, various types of environmental data (such as three-dimensional point cloud data, underwater obstacle sonar data, and dredger operation data) are seamlessly integrated to generate high-quality multimodal environmental data packets. These high-precision data provide a solid foundation for subsequent path planning, making the generated navigation path more accurate and reliable.

[0038] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A multi-anchor cable tension adaptive control system for a height-limited section of a cutter suction vessel, characterized in that: include, The acquisition module collects the ship's environmental situation data, performs timestamp alignment, noise filtering and coordinate system one, and generates a multi-modal environmental data packet; The model building module uses the potential field modeling engine to build a three-dimensional dynamic potential field model based on the multimodal environment data package, and generates the potential field distribution matrix through the finite difference method. At the same time, the partial derivatives of the potential field distribution matrix are calculated using the central difference to generate the three-dimensional potential field gradient tensor; The path planning module uses the RRT algorithm to iteratively generate candidate path nodes based on the three-dimensional potential field gradient tensor, and then smoothes them through the cubic spline interpolation algorithm to generate the initial navigation path; The instruction generation module inputs the initial navigation path and real-time dredger operation data into the dual-stream neural network, extracts spatial topological features using the spatial stream branch, extracts dynamic correlation features using the temporal stream branch, fuses the spatial topological features and dynamic correlation features, and inputs them into the strategy network to output the adjustment instruction set of the anchor cable tension; In the anchor cable tension control module, the potential field modeling engine parses the anchor cable tension adjustment instruction set, generates a new potential field distribution matrix and three-dimensional potential field gradient tensor, and inputs them into the two-stream neural network. The two-stream neural network performs parameter updates and path replanning to generate the final navigation path.

2. The self-adaptive control system for multiple anchor cable tension in the height-limited section of a cutter suction vessel according to claim 1, characterized in that: The ship environment situation data includes three-dimensional point cloud data of the height-restricted section, sonar data of underwater obstacles and operation data of the dredger.

3. The self-adaptive control system for multiple anchor cable tension in the height-limited section of a cutter suction vessel according to claim 1, characterized in that: The generating of the multimodal environment data packet specifically comprises the following steps: The ship environment situation data is time-stamped and aligned through the PTP protocol. The three-dimensional point cloud data in the height-restricted section is filtered for outliers through the rectangular information granulation algorithm. The sonar data of underwater obstacles is denoised using the wavelet threshold. The operation data of the cutter suction vessel is clustered using the DBSCAN algorithm to remove outliers. The ICP algorithm is used to unify the preprocessed ship environmental situation data into the geographic coordinate system, and a hierarchical fusion strategy is adopted to integrate them to generate a multimodal environmental data package.

4. The self-adaptive control system for multiple anchor cable tension in the height-limited section of a cutter suction vessel according to claim 1, characterized in that: The potential field distribution matrix is ​​generated by the finite difference method, which includes the following steps: Based on the multimodal environment data package, a 3D voxel map is constructed using the point cloud voxelization downsampling method, and a spatial index is established through the KD-Tree. At the same time, a 3D dynamic potential field model is constructed through a dynamic weight allocation strategy combined with a hyperbolic secant function. Based on the three-dimensional dynamic potential field model, the spatial gradient field is calculated by the finite difference method to obtain the gravitational gradient and repulsive field strength. At the same time, based on the gravitational gradient and repulsive field strength, the potential field distribution matrix is ​​generated by the multi-physical field coupling superposition algorithm.

5. The self-adaptive control system for multiple anchor cable tension in the height-limited section of a cutter suction vessel according to claim 1, characterized in that: The generating of the three-dimensional potential field gradient tensor specifically comprises the following steps: The first-order partial derivative of the potential field distribution matrix is ​​calculated by using the central difference in the finite difference method, and the initial gradient vector field is obtained by using the vector synthesis algorithm based on the first-order partial derivative; The initial gradient vector field is reconstructed by the ADMM coupled reconstruction method, the second-order partial derivatives are cross-calculated by the central difference method, and the weight distribution ratio of the second-order partial derivatives is dynamically adjusted using the hyperbolic secant function. At the same time, regularization is performed by the alternating direction multiplier method to generate a three-dimensional potential field gradient tensor.

6. The self-adaptive control system for multiple anchor cable tension in the height-limited section of a cutter suction vessel according to claim 5, characterized in that: The specific steps of generating the initial navigation path are as follows: The RRT algorithm is used to perform node expansion on the three-dimensional potential field gradient tensor to obtain the dynamic direction vector, and the dynamic direction vector is converted into the path expansion direction through the elliptical sampling domain conversion method. At the same time, the node expansion step size is obtained by using the adaptive step size strategy. Generate candidate path nodes through a hierarchical search strategy according to the path expansion direction and node expansion step; The candidate path nodes are processed by the cubic spline interpolation algorithm to generate a cubic spline smooth trajectory; Based on the cubic spline smooth trajectory, the ADMM algorithm is used for dynamic regularization optimization to generate the initial navigation path.

7. The self-adaptive control system for multiple anchor cable tension in the height-limited section of a cutter suction vessel according to claim 6, characterized in that: The specific steps of extracting spatial topological features by using spatial stream branches are as follows: The initial navigation path is input into the spatial stream branch of the two-stream neural network. The first level of the spatial stream branch performs time-space joint convolution and spatial dimension reduction to capture spatial structural features. The second level performs local iterative optimization and cross-layer feature fusion to obtain refined local features. The third layer dynamically assigns weights of different scales to spatial structural features and refined local features through the channel attention mechanism, and fuses them through splicing channels to generate spatial topological features.

8. The self-adaptive control system for multiple anchor cable tension in the height-limited section of a cutter suction vessel according to claim 7, characterized in that: The specific steps of extracting dynamic correlation features by using time stream branches are as follows: The real-time cutter suction vessel operation data is input into the time stream branch of the two-stream neural network, where the forward TCN of the bidirectional long short-term memory network captures the long-range trend through dilated convolution, and the backward TCN models the reverse degradation pattern through causal convolution; The long-range trends and reverse degradation patterns are bidirectionally spliced ​​through the temporal attention mechanism to generate dynamic correlation features.

9. The self-adaptive control system for multiple anchor cable tension in the height-limited section of a cutter suction vessel according to claim 1, characterized in that: The specific steps of the output anchor cable tension adjustment instruction set are as follows: The spatial topological features and dynamic temporal features are weightedly fused and input into the DDPG policy network. The Actor-Critic collaborative optimization framework is used to perform policy gradient updates to generate the anchor cable tension adjustment parameter vector. Based on the anchor cable tension adjustment parameter vector, an adjustment instruction set of the anchor cable tension is generated by a finite element algorithm.

10. The self-adaptive control system for multiple anchor cable tension in the height-limited section of a cutter suction vessel according to claim 1, characterized in that: The specific steps of generating the final navigation path are as follows: Based on the adjustment instruction set of anchor cable tension, a new potential field distribution matrix is ​​generated through the finite element tension field coupling algorithm, and the implicit Euler method is used for nonlinear calculation to generate a new three-dimensional potential field gradient tensor; The new potential field distribution matrix and the three-dimensional potential field gradient tensor are input into the two-stream neural network, and the weight parameters are updated using the elastic weight solidification regularization method to generate an optimized two-stream neural network. Based on the optimized two-stream neural network, the initial navigation path deviation is corrected through the sliding mode control law to generate the final navigation path.

Images