Automatic berthing dynamic path planning method and system based on deep learning

Abstract

This application provides a deep learning-based method and system for automatic berthing dynamic path planning, relating to the field of deep learning technology. First, it acquires a set of real-time navigation perception data of the target vessel during berthing, including shoreline echo point cloud data, instantaneous flow field vector distribution data of the surrounding waters, and real-time six-degree-of-freedom motion parameters of the vessel's attitude. Next, it analyzes the shoreline echo point cloud data to generate shoreline contour information and berth spatial location identifiers, and constructs a dynamic feasible domain constraint boundary. Then, it calls a berth tendency path generation network to perform multi-constraint path planning, generating an initial berth tendency path. The initial berth tendency path and related data are then input into a path dynamic fine-tuning network for optimization, generating an automatic berthing dynamic path planning result. Finally, based on this automatic berthing dynamic path planning result, it generates a set of berthing control commands to drive the vessel's propulsion system and steering gear system to execute berthing movements, improving the safety and accuracy of vessel berthing.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and more specifically, to a method and system for automatic docking dynamic path planning based on deep learning. Background Technology

[0002] Traditional ship berthing methods rely primarily on crew experience and manual operation; however, these methods have numerous limitations. Firstly, human judgment is susceptible to environmental factors and individual circumstances, making it difficult to comprehensively and accurately perceive the complex environmental information surrounding the ship, such as the precise geometry of the quay, instantaneous changes in the surrounding water flow, and the ship's own dynamic attitude. This can easily lead to safety accidents such as collisions and groundings during berthing. Secondly, manual operation struggles to quickly and accurately adjust the berthing path and attitude based on environmental changes and the ship's condition, resulting in low berthing efficiency. The limitations of traditional berthing methods are particularly pronounced when facing complex and changing port environments and adverse weather conditions. With the continuous development of the shipping industry, higher demands are being placed on the automation and intelligence of ship berthing operations. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method and system for automatic docking dynamic path planning based on deep learning.

[0004] According to a first aspect of this application, a deep learning-based method for automatic docking dynamic path planning is provided, the method comprising: Acquire a set of real-time navigation perception data of the target vessel during berthing operations. The set of real-time navigation perception data includes the shoreline echo point cloud data collected by the shipborne sensing equipment, the instantaneous flow field vector distribution data of the waters surrounding the target vessel, and the real-time six-degree-of-freedom motion parameters of the target vessel's hull attitude. The shoreline echo point cloud data is processed by shoreline geometry analysis to generate continuous shoreline contour information and spatial location reference identifier of the target berthing berth. At the same time, the dynamic feasible domain constraint boundary of the target vessel in the berthing water is constructed based on the instantaneous flow field vector distribution data and the real-time six-degree-of-freedom motion parameters. The pre-built berth approach path generation network is invoked to perform multi-constraint path planning processing on the continuous contour morphology information of the wharf shoreline, the spatial location reference identifier and the dynamic feasible domain constraint boundary, to generate an initial berth approach path for the target vessel to approach the target berth from its current position. The initial berth approach path includes a continuously distributed sequence of path nodes and a sequence of target vessel reference attitude angles at each path node. The initial berth orientation path, the continuous contour morphology information of the wharf shoreline, the instantaneous flow field vector distribution data, and the real-time six-degree-of-freedom motion parameters are input into a pre-constructed path dynamic fine-tuning network for path smoothing and attitude optimization, generating automatic berthing dynamic path planning results. The automatic berthing dynamic path planning results include a corrected path node sequence and a path node corrected attitude angle sequence. A set of berthing control commands is generated based on the automatic berthing dynamic path planning results. The set of berthing control commands is used to drive the propulsion system and steering gear system of the target ship to perform berthing movements along the automatic berthing dynamic path planning results.

[0005] According to a second aspect of this application, a deep learning-based automatic docking dynamic path planning system is provided. The deep learning-based automatic docking dynamic path planning system includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the deep learning-based automatic docking dynamic path planning system implements the aforementioned deep learning-based automatic docking dynamic path planning method.

[0006] Based on any of the above aspects, the technical effect of this application is as follows: By acquiring real-time navigation perception data of the target vessel during berthing operations, a comprehensive and accurate understanding of key information such as the wharf shoreline, the flow field of the surrounding waters, and the vessel's own attitude is obtained. The echo point cloud data of the wharf shoreline is analyzed and processed to generate accurate continuous contour information of the wharf shoreline and spatial location reference markers for the target berth. Simultaneously, a dynamic feasible domain constraint boundary is constructed by combining instantaneous flow field vector distribution data and real-time six-degree-of-freedom motion parameters, fully considering various constraints during the vessel berthing process. A pre-constructed berth tendency path generation network is invoked for multi-constraint path planning, generating an initial berth tendency path containing a continuously distributed path node sequence and a reference attitude angle sequence of the target vessel at each node. The initial path is then smoothed and its attitude optimized through a path dynamic fine-tuning network, further improving the path quality and the stability of vessel berthing. The final generated berthing control command set can precisely drive the target vessel's propulsion system and steering gear system to execute berthing movements, effectively improving the safety, accuracy, and efficiency of vessel berthing operations. Attached Figure Description

[0007] Figure 1 A flowchart illustrating the deep learning-based automatic docking dynamic path planning method provided in an embodiment of this application is shown. Figure 2 This illustration shows a schematic diagram of the component structure of the deep learning-based automatic docking dynamic path planning system provided in an embodiment of this application. Detailed Implementation

[0008] Figure 1 This paper illustrates a flowchart of the deep learning-based automatic docking dynamic path planning method and system provided in an embodiment of this application. The detailed steps include: Step S110: Obtain a set of real-time navigation perception data of the target vessel during the berthing operation. The set of real-time navigation perception data includes the shoreline echo point cloud data collected by the shipborne sensing equipment, the instantaneous flow field vector distribution data of the waters surrounding the target vessel, and the real-time six-degree-of-freedom motion parameters of the target vessel's hull attitude.

[0009] The shipborne lidar deployed on the target vessel outputs shoreline echo point cloud data at a fixed frequency. Each laser echo point in the shoreline echo point cloud data contains three-dimensional spatial coordinates and echo intensity values ​​in the ship's coordinate system. The shipborne acoustic Doppler current profiler emits acoustic pulses into the waters surrounding the target vessel and receives Doppler frequency-shifted echoes. After converting the frequency shift to flow velocity, it calculates the velocity vectors at each layer. Then, through spatial interpolation, it generates instantaneous flow field vector distribution data covering a rectangular water grid around the target vessel. Each grid node in the instantaneous flow field vector distribution data stores the east component Ue and the north component Un of the flow velocity. The shipborne inertial navigation system outputs real-time six-degree-of-freedom motion parameters, including sway velocity u, pitch velocity v, heave velocity w, roll angular velocity p, pitch angular velocity q, and bow angular velocity r. These three types of data are aligned to a common Coordinated Universal Time (UTC) timestamp t via a ship network time synchronization protocol.

[0010] Step S120: Perform shoreline geometric morphology analysis on the echo point cloud data of the wharf shoreline to generate continuous contour morphology information of the wharf shoreline and spatial location reference identifier of the target berthing berth. At the same time, construct the dynamic feasible domain constraint boundary of the target vessel in the berthing water area based on the instantaneous flow field vector distribution data and the real-time six-degree-of-freedom motion parameters.

[0011] Step S121: Perform voxel downsampling on the echo point cloud data of the wharf shoreline to generate a sparse shoreline point cloud set. The sparse shoreline point cloud set retains the spatial point location information representing the shoreline structural skeleton in the echo point cloud data of the wharf shoreline.

[0012] The 3D space containing the shoreline echo point cloud data is divided into cubic voxel grids of fixed size, with each cubic voxel having a side length of Lv. For each cubic voxel grid, all laser echo points falling within that grid are replaced by the laser echo point closest to the geometric center of that grid; the remaining laser echo points are discarded. If a cubic voxel grid contains no laser echo points, no representative point is generated for that grid. All retained laser echo points form a sparse shoreline point cloud set, where each point retains its 3D spatial coordinates and echo intensity value.

[0013] Step S122: Perform local point neighborhood normal vector consistency analysis on the sparse shoreline point cloud set to generate a set of local geometric morphology parameters for the degree of normal vector deflection and the degree of neighborhood surface fluctuation for each voxel unit.

[0014] For each retained point in the sparse shoreline point cloud set, a spherical space with radius Rs centered at that point is used as the neighborhood search range. All neighboring points within this spherical space are extracted, and the total number of neighboring points is denoted as Ns. A 3×Ns matrix is ​​constructed for the 3D spatial coordinates of these Ns neighboring points. Principal component analysis is performed on this matrix to obtain three eigenvalues ​​and three corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is the local surface normal vector n of that point, where n is a unit vector of length 3. The local surface normal vector of each point in the neighborhood is calculated in the same way. The standard deviation of the angle between the local surface normal vector n of the point and the local surface normal vectors of each point in the neighborhood is calculated, and this standard deviation is used as the degree of normal vector deflection of the point. The root mean square value of the vertical distance from each point in the neighborhood to the local fitted plane defined by the local surface normal vector n of the point is calculated, and this root mean square value is used as the degree of neighborhood surface undulation of the point. The normal vector deflection value reflects the consistency of the local surface normal vector at that point, while the neighborhood surface undulation value reflects the roughness of the local neighborhood surface at that point.

[0015] Step S123: Based on the degree of neighborhood surface fluctuation in the set of local geometric morphology parameters, perform shoreline feature boundary point extraction processing on the sparse shoreline point cloud set to generate a shoreline feature boundary point set. Each shoreline feature boundary point in the shoreline feature boundary point set carries spatial coordinate information and local normal vector pointing information.

[0016] Define a lower bound threshold Tlow and an upper bound threshold Thigh for volatility. Iterate through each retained point in the sparse shoreline point cloud set. If the volatility value of a point's neighborhood surface is greater than Tlow and less than Thigh, then that point is extracted as a shoreline feature boundary point. Points with volatility lower than Tlow are considered points within the flat shoreline surface, while points with volatility higher than Thigh are considered noise or vegetation patches. Each shoreline feature boundary point stores its 3D spatial coordinates and the corresponding local surface normal vector n. All shoreline feature boundary points constitute a shoreline feature boundary point set.

[0017] Step S124: Input the set of shoreline feature boundary points into the gated recurrent unit of the pre-constructed shoreline contour growth network to perform point-by-point state update processing along the spatial distribution of the set of shoreline feature boundary points, and generate the hidden state vector of each shoreline feature boundary point.

[0018] The shoreline contour growth network is a pre-trained neural network model. All shoreline feature boundary points in the shoreline feature boundary point set are sorted in ascending order according to their spatial coordinates projected onto the shoreline extension direction, forming a point sequence. For the t-th point in the point sequence, its 3D spatial coordinates and local surface normal vector n are concatenated to form an input feature vector xt with dimension 6. The calculation process of the gated recurrent unit is as follows: update gate zt = σ(Wz×xt + Uz×ht-1 + bz), reset gate rt = σ(Wr×xt + Ur×ht-1 + br), candidate hidden state... Currently hidden Where Wz, Uz, Wr, Ur, Wh, and Uh are trainable weight matrices, bz, br, and bh are trainable bias vectors, σ is the Sigmoid function, tanh is the hyperbolic tangent function, and ⊙ represents element-wise multiplication. The gated recurrent unit processes the point sequence one by one, outputting a hidden state vector ht for each shoreline feature boundary point.

[0019] Step S125: The hidden state vector of each shoreline feature boundary point and the hidden state vector of its neighboring shoreline feature boundary points in its local neighborhood are processed by the neighbor attention convergence layer of the shoreline contour growth network to generate an enhanced hidden state vector containing neighborhood context association information.

[0020] For the i-th shoreline feature boundary point in the point sequence, take the three-dimensional spatial coordinates of this point as the center, and select all other shoreline feature boundary points within a preset neighborhood radius Ra to form a neighbor point set N(i). Use the hidden state hi of this point as the query vector, and the hidden state hj of neighbor point j as the key vector and value vector. Calculate the scaled dot product attention score between the query vector and each key vector: sij=(Wq×hi)^T×(Wk×hj) / sqrt(d), where Wq and Wk are the trainable query weight matrix and key weight matrix, and d is the dimension of the hidden state vector. Normalize all sij using the softmax function to obtain the attention coefficients: aij=exp(sij) / sum(exp(sik)), k iterates through N(i). Enhance the hidden state vector h'i=hi+sum(aij×(Wv×hj)), j iterates through N(i), where Wv is the trainable value weight matrix. Attention convergence enables each boundary point to integrate the contextual morphological information of its surrounding neighboring boundary points.

[0021] Step S126: The enhanced hidden state vector of each shoreline feature boundary point is classified by the boundary point classification output layer of the shoreline contour growth network. Each shoreline feature boundary point is marked as a shoreline continuous contour constituent point or a shoreline break discontinuity point, generating a set of shoreline continuous contour constituent points.

[0022] The boundary point classification output layer consists of a fully connected layer and a sigmoid activation function. The fully connected layer maps the enhanced hidden state vector h'i to a scalar logit value: logit_i = Wc × h'i + bc, where Wc is the trainable classification weight matrix and bc is the trainable classification bias. The sigmoid function maps this logit value to a probability value in the interval (0, 1): pi = 1 / (1 + exp(-logit_i)). This probability value represents the probability that the i-th shoreline feature boundary point belongs to the shoreline continuous contour constituent point. A classification threshold Thr is set; if pi > Thr, the point is marked as a shoreline continuous contour constituent point; otherwise, it is marked as a shoreline break discontinuity point. All shoreline feature boundary points marked as shoreline continuous contour constituent points constitute the shoreline continuous contour constituent point set.

[0023] Step S127: Connect the continuous shoreline contour points in the set of continuous shoreline contour points in an orderly manner according to their spatial proximity to generate continuous shoreline contour morphology information of the wharf shoreline. The continuous shoreline contour morphology information of the wharf shoreline includes the sequence of shoreline contour points that constitute the wharf shoreline and the curvature of the connecting line segments between adjacent shoreline contour points.

[0024] For the set of points constituting the continuous shoreline contour, starting from one end of the shoreline, each time, within the neighborhood of the currently connected point, search for the spatially closest unconnected point constituting the continuous shoreline contour as the next connection point and add it to the end of the ordered contour point sequence. Repeat this greedy connection process until all points constituting the continuous shoreline contour are connected. For every three adjacent contour points in the ordered contour point sequence, let the spatial coordinates of the three points be Pa, Pb, and Pc, calculate the vectors v1 = Pb - Pa and v2 = Pc - Pb. The curvature κ of the spatial arc determined by the three points is κ = 2 × |v1 × v2| / (|v1| × |v2| × |v1 + v2|), where × represents the cross product of the vectors, and |·| represents the vector magnitude. The ordered contour point sequence and the curvature values ​​between each of the three adjacent points together constitute the continuous contour morphological information of the wharf shoreline.

[0025] Step S128: Obtain the pre-calibrated static spatial coordinates of the target berth, perform spatial registration processing between the static spatial coordinates of the berth and the shoreline contour point sequence in the continuous contour morphology information of the wharf shoreline, and generate a spatial position reference identifier for the target berth. The spatial position reference identifier for the target berth includes the coordinates of the berth center point and the direction vector of the berth leading edge line segment.

[0026] The static spatial coordinates of the target berth are provided in advance by the port authority and stored as the latitude and longitude coordinates of the two endpoints of the berth in the geographic coordinate system. These coordinates are then transformed to the three-dimensional spatial coordinates Pberth1 and Pberth2 in the ship's coordinate system. Pberth1 and Pberth2 are projected onto the plane containing the continuous contour point sequence of the wharf shoreline. The shoreline contour point with the closest spatial distance to Pberth1 and Pberth2 is searched within the shoreline contour point sequence as the matching point. The coordinates of the berth center point are Pcenter = (Pberth1 + Pberth2) / 2. The direction vector of the berth leading edge segment is vberth = (Pberth2 - Pberth1) / |Pberth2 - Pberth1|. Pcenter and vberth constitute the spatial location reference identifier for the target berth.

[0027] Step S129: Construct the flow field velocity vector field of the water area surrounding the target ship based on the instantaneous flow field vector distribution data; determine the instantaneous motion velocity vector of the target ship based on the target ship's sway velocity, pitch velocity, and bow roll angular velocity in the real-time six-degree-of-freedom motion parameters; perform vector superposition processing on the instantaneous motion velocity vector of the target ship and the flow field velocity vector at the corresponding spatial position in the flow field velocity vector field to generate the comprehensive disturbance velocity field of the target ship in the berthing water area.

[0028] The instantaneous flow field vector distribution data is a gridded velocity vector field, with each grid node storing the east component Ue and the north component Un. Assume the instantaneous velocity vector of the target ship is composed of the sway velocity u and the pitch velocity v, forming a planar velocity vector Vs = (u, v), and the target ship's bow roll rate is r. For each flow field grid node within a radius Rzone centered on the target ship, its flow field velocity vector is Vc = (Ue, Un). The comprehensive disturbance velocity field Vtotal at this grid node is Vtotal = Vs + Vc, with a vector direction of arctan(Vtotal_north / Vtotal_east) and a vector magnitude of |Vtotal| = sqrt(Vtotal_east² + Vtotal_north²). The comprehensive disturbance velocity field reflects the summative effect of the flow field and ship motion on the velocity distribution in the waters surrounding the target ship.

[0029] Step S1210: Based on the integrated disturbance velocity field and combined with the preset safety boundary rules, generate the dynamic feasible domain constraint boundary of the target vessel in the berthing water. The dynamic feasible domain constraint boundary includes the potential field boundary surface around the target vessel that is prohibited from being crossed.

[0030] The preset safety boundary rules include: the minimum safe distance d1 between the target vessel and the wharf shoreline, the minimum safe distance d2 between the target vessel and shallow water areas or obstacles, and a prohibited zone for areas where the velocity vector magnitude in the combined disturbance velocity field exceeds the safe velocity threshold Vsafe. Centered on the target vessel, these three safety constraints are mapped to a potential field boundary surface in three-dimensional space around the target vessel that is prohibited from crossing. The potential field boundary surface consists of a series of triangular facets, each storing its vertex spatial coordinates and normal vector orientation, collectively forming a dynamic feasible region constraint boundary. As the combined disturbance velocity field changes over time, the potential field boundary surface is also dynamically updated.

[0031] Step S130: Call the pre-built berth approach path generation network to perform multi-constraint path planning processing on the continuous contour morphology information of the wharf shoreline, the spatial position reference identifier and the dynamic feasible domain constraint boundary to generate an initial berth approach path for the target vessel to approach the target berth from its current position. The initial berth approach path includes a continuously distributed sequence of path nodes and a sequence of reference attitude angles of the target vessel at each path node.

[0032] Step S131: The target vessel's current position coordinates, the target vessel's current heading angle, the berth center point coordinates and berth leading edge line segment direction vector in the spatial position reference identifier, the potential field boundary surface geometric parameters of the dynamic feasible domain constraint boundary, and the shoreline contour point sequence in the continuous contour morphology information of the wharf shoreline are uniformly encoded into a multi-channel state tensor with the same spatial reference.

[0033] The target vessel's current position coordinates Pnow are a three-dimensional vector. The target vessel's current heading angle ψnow is a scalar. The berth center point coordinates Pcenter and the berth leading edge line segment direction vector vberth are obtained from step S128. The potential field boundary surface geometric parameters are the vertex coordinate sequence and normal vector sequence of each triangular facet constituting the potential field boundary surface. The shoreline contour point sequence is the ordered shoreline contour points generated in step S127. The above multidimensional heterogeneous data is rasterized at a uniform spatial resolution onto a two-dimensional or three-dimensional raster plane centered on the target vessel and covering the predetermined berthing area. Each raster cell stores the multi-channel feature values ​​of that location, including: the nearest distance to the shoreline, the nearest distance to the potential field boundary surface, the relative azimuth of the berth center point, and the angle between the current heading and the berth direction. After normalization, each channel constitutes a multi-channel state tensor Mstate.

[0034] Step S132: Input the multi-channel state tensor into the multi-layer spatial graph convolutional layer of the berth tendency path generation network for spatial feature extraction processing to generate a spatial fusion feature map containing the ship's own state and water environment constraints.

[0035] The multi-layer spatial graph convolutional layer is composed of stacked multi-layer graph convolutional network layers. Each grid cell in the multi-channel state tensor Mstate is regarded as a graph node, and adjacency edges are established between adjacent grid cells to construct a spatial adjacency graph. For each graph node, each graph convolutional network layer aggregates the feature vectors of its neighboring nodes, linearly transforms them with the trainable weight matrix Wgcn, adds them to its own features, and then performs a non-linear transformation with the ReLU activation function to output the updated node feature vector. After processing by the multi-layer graph convolutional network layers, the feature vectors of each graph node are fused with the neighborhood environmental context information to form a spatial fusion feature map Fspatial. The feature vector at each spatial location encodes the spatial relationship between the ship's own state and the constraints of the water environment at that location.

[0036] Step S133: The spatial fusion feature map is processed by performing temporal dependency modeling through the bidirectional long short-term memory network layer of the berth tendency path generation network to generate a time series feature vector along the berthing direction. The time series feature vector along the berthing direction contains the temporal context information of the current state for the selection of future path nodes.

[0037] The spatial fusion feature map Fspatial is serialized and unfolded along the berthing direction axis from the current position of the target vessel to the center point of the berth, resulting in a sequence of feature vectors arranged along the berthing direction. This sequence of feature vectors is then input into a forward long short-term memory (LSTM) network layer and a backward LSM network layer in forward and reverse order, respectively. The forward LSM network layer processes the sequence from front to back, and the backward LSM network layer processes the sequence from back to front. At each position in the sequence, the forward hidden state and the backward hidden state are concatenated to obtain a time-series feature vector Htemporal along the berthing direction that contains bidirectional contextual information.

[0038] Step S134: Input the time series feature vector into the attention decoding layer of the berth tendency path generation network, and perform global feature interaction processing on the time series feature vector through the multi-head self-attention mechanism of the attention decoding layer to generate the attention weight distribution of each path node to be generated.

[0039] The attention decoding layer employs a Transformer decoder architecture. The time-series feature vector Htemporal is used as the input sequence, and the input sequence is mapped to the query matrix Q, key matrix K, and value matrix V respectively through trainable query weight matrix Wq, key weight matrix Wk, and value weight matrix Wv. For each attention head, the attention score matrix S = Q × K^T / sqrt(dk) is calculated, where dk is the dimension of the key vector. Each row of the attention score matrix is ​​normalized using the softmax function to obtain the attention weight matrix A = softmax(S). The value matrix V is weighted and summed using the attention weight matrix to obtain the output Ohead = A × V for each attention head. The outputs of multiple attention heads are concatenated along the feature dimension and linearly transformed using the output weight matrix Wo to obtain the attention weight distribution for each path node to be generated.

[0040] Step S135: The attention weight distribution is used to perform weighted summation on the corresponding spatial positions of the spatial fusion feature map to generate a spatial coordinate prediction vector for each path node to be generated. The spatial coordinate prediction vector contains the predicted spatial coordinates of the path node.

[0041] For the k-th path node to be generated, the attention weight distribution of the node generated in step S134 is used as a weighting coefficient, and weighted and summed with the feature vectors of each spatial location in the spatial fusion feature map Fspatial to obtain the aggregated feature vector of the node. The aggregated feature vector is input into the feedforward fully connected network, which outputs the spatial coordinate prediction vector of the path node, containing the predicted three-dimensional spatial coordinates of the path node.

[0042] Step S136: Construct a path node sequence based on the spatial coordinate prediction vector, and analyze and process the spatial coordinate difference vector between adjacent path nodes in the path node sequence through the attitude angle prediction branch of the berth tendency path generation network to generate a target ship reference attitude angle sequence at each path node. The target ship reference attitude angle sequence includes the ship's predicted heading angle at each path node.

[0043] Arrange the spatial coordinate prediction vectors of all path nodes generated in step S135 in the order of generation to form a path node sequence. For each adjacent two path nodes in the path node sequence, the attitude angle prediction branch calculates the spatial coordinate difference vector obtained by subtracting the coordinates of the previous node from the coordinates of the subsequent node. The horizontal component of this spatial coordinate difference vector is the predicted heading angle of the target vessel at that path node, measured clockwise from true north. Arrange the predicted heading angles at all path nodes in node order to obtain the reference attitude angle sequence of the target vessel.

[0044] Step S137: Perform point-by-point distance calculation on the potential field boundary surface of the path node sequence, the target ship reference attitude angle sequence and the dynamic feasible domain constraint boundary, and reproject the path nodes that are less than the preset safety distance margin from the potential field boundary surface to the safe area outside the potential field boundary surface to generate the corrected path node sequence.

[0045] For each path node in the path node sequence, calculate the shortest distance from the node's spatial coordinates to the nearest triangular facet on the potential field boundary surface. Compare this shortest distance with a preset safety margin dmin. If the shortest distance is less than dmin, translate the node's spatial coordinates along the outward normal vector direction from the node to the nearest triangular facet to a position exactly equal to dmin from the potential field boundary surface. All adjusted path nodes form the corrected path node sequence in their original order.

[0046] Step S138: Perform path node density homogenization processing on the corrected path node sequence, insert new path nodes between path node pairs where the spacing between adjacent path nodes exceeds the preset node spacing upper limit, and generate a path node sequence with uniform density.

[0047] In the corrected path node sequence, for every two adjacent path nodes, calculate the Euclidean distance Dpair between them. If Dpair is greater than the preset node spacing upper limit dmax, then insert ceil(Dpair / dmax)−1 new path nodes at equal intervals along the line connecting the two points. The spatial coordinates of each new path node are calculated using linear interpolation. After inserting the new nodes, recalculate the predicted heading angle at each node. The resulting sequence has a uniform density of path nodes.

[0048] Step S139: Combine the uniformly dense path node sequence and the corresponding target ship reference attitude angle sequence into an initial berth approach path, wherein the initial berth approach path includes a continuously distributed path node sequence and a target ship reference attitude angle sequence at each path node.

[0049] Step S140: Input the initial berth orientation path, the continuous contour morphology information of the wharf shoreline, the instantaneous flow field vector distribution data, and the real-time six-degree-of-freedom motion parameters into the pre-constructed path dynamic fine-tuning network for path smoothing and attitude optimization, and generate an automatic berthing dynamic path planning result containing the corrected path node sequence and the path node corrected attitude angle sequence.

[0050] Step S141: Obtain the path node sequence of the initial berth approach path and the target ship reference attitude angle sequence at each path node, and construct a spatiotemporal sequence sample of the path nodes. The spatiotemporal sequence sample of the path nodes includes the spatial coordinates of each path node, the corresponding target ship reference attitude angle, and the path node number.

[0051] The data structure of the initial berth approach path is analyzed, and the path node sequence and the corresponding target vessel reference attitude angle sequence are extracted. Each path node in the path node sequence is assigned a number from 1 to N according to its arrangement. The spatial coordinates of each path node, the target vessel reference attitude angle, and the path node number are combined into a basic spatiotemporal data unit. Multiple basic spatiotemporal data units are arranged in ascending order of their numbers to form a spatiotemporal sequence sample.

[0052] Step S142: The flow field velocity vector in the instantaneous flow field vector distribution data is spatially interpolated according to the spatial coordinates of the path nodes to generate an instantaneous flow velocity vector at each path node. The instantaneous flow velocity vector includes the flow velocity direction and the flow velocity magnitude.

[0053] For the spatial coordinates of each path node, the four nearest grid nodes in the instantaneous flow field vector distribution data are found, and the instantaneous velocity vector at that spatial coordinate is calculated using bilinear interpolation. The interpolated instantaneous velocity vector contains an east component Ue_int and a north component Un_int. The velocity magnitude |Vcurrent|=sqrt(Ue_int²+Un_int²) and the velocity direction are calculated from these two components. The velocity direction is the angle between the velocity vector and true north.

[0054] Step S143: Perform feature concatenation processing on the instantaneous flow velocity vector and the real-time six-degree-of-freedom motion parameters to generate a hydrodynamic coupling feature vector at each path node.

[0055] The east component Ue_int and north component Un_int of the instantaneous flow velocity vector are concatenated with the sway velocity u, pitch velocity v, and yaw rate r of the real-time six-degree-of-freedom motion parameters along the feature dimension to form a hydrodynamic coupling feature vector of length 5.

[0056] Step S144: Input the hydrodynamic coupling feature vector and the target ship reference attitude angle of the path node into the attitude angle gating fine-tuning unit of the path dynamic fine-tuning network, calculate the fine-tuning offset of the target ship reference attitude angle through the gating mechanism of the attitude angle gating fine-tuning unit, and generate the preliminary corrected attitude angle of the path node.

[0057] The attitude angle gating fine-tuning unit is a small neural network composed of fully connected layers and a gating structure. The input is a concatenation of the hydrodynamic coupling feature vector and the target ship reference attitude angle of the path node. After mapping through a fully connected layer, a gating coefficient g is generated by the Sigmoid activation function, with g ranging from (0, 1). Simultaneously, another fully connected layer outputs a candidate attitude angle offset value Δψ_candidate. The fine-tuning offset Δψ = g × Δψ_candidate. The path node initially corrects the attitude angle ψ_adj = ψ_ref + Δψ, where ψ_ref is the target ship reference attitude angle.

[0058] Step S145: Perform coordinate transformation on the initial corrected attitude angle and spatial coordinates of the path node, and calculate the minimum distance interval between the outer contour of the ship and the shoreline contour at the path node by combining the curvature of the connecting line segment between adjacent shoreline contour points in the continuous contour morphology information of the wharf shoreline.

[0059] The outer contour of the target vessel is simplified into a rectangular bounding box. A local coordinate system is established at each path node, with the vessel's center at the path node spatial coordinates and the initial corrected attitude angle at the path node as the vessel's heading. Within this local coordinate system, the coordinates of the four vertices of the vessel's outer contour are constructed. Similarly, in the global coordinate system, adjacent shoreline contour points from the continuous shoreline contour morphology information are connected by curvature to form a shoreline contour curve. The shortest Euclidean distances from the four vertices of the vessel's outer contour to the shoreline contour curve are calculated, and the minimum of these four distances is taken as the minimum distance interval between the vessel's outer contour and the shoreline contour at each path node.

[0060] Step S146: Input the minimum distance interval into the lateral offset compensation layer of the path dynamic fine-tuning network, calculate the lateral offset correction amount of each path node in the direction perpendicular to the path tangent, and perform lateral translation processing on the spatial coordinates of the path node based on the lateral offset correction amount to generate the preliminary correction result of the spatial coordinates of the path node.

[0061] The lateral offset compensation layer is a proportional controller. The minimum distance interval is set to d_min, and the preset shoreline safety distance is d_safe. Lateral offset correction amount. The lateral translation direction is a unit vector pointing away from the shoreline along a direction perpendicular to the path tangent. The spatial coordinates of the path node are translated by Δd along this lateral translation direction to generate a preliminary correction of the path node's spatial coordinates.

[0062] Step S147: Input the preliminary correction result of the path node spatial coordinates and the preliminary correction attitude angle of the path node into the path curvature linkage optimization layer of the path dynamic fine-tuning network, extract the local curvature change features of the path node sequence through convolution operation, and perform curvature smoothing processing on the preliminary correction result of the path node spatial coordinates according to the local curvature change features to generate the secondary correction result of the path node spatial coordinates.

[0063] The path curvature linkage optimization layer contains a one-dimensional convolutional neural network. The initial correction results of the path node spatial coordinates are arranged into a coordinate sequence by node number. A one-dimensional convolution operation is performed on this sequence, and the convolution kernel slides along the sequence to extract local coordinate change features between adjacent path nodes. Curvature estimation is performed on the extracted features. At positions where the curvature exceeds a preset curvature threshold, the spatial coordinates of the path nodes are locally smoothed to ensure that the path curvature meets the maximum curvature constraint, generating a secondary correction result for the path node spatial coordinates.

[0064] Step S148: The global path shape encoder of the path dynamic fine-tuning network performs global shape encoding processing on the secondary correction result of the spatial coordinates of the path nodes to generate a global path shape latent variable, which contains the overall morphological features of the initial berth tendency path.

[0065] The global path shape encoder is a point encoder network. It takes the spatial coordinates of all path nodes from the secondary correction result as an unordered set of points as input, embeds and encodes the coordinates of each point using a multilayer perceptron, and then aggregates the embeddings of all points using a global max-pooling layer to generate a fixed-dimensional global path shape latent variable. This global path shape latent variable encodes the global shape features of the entire path.

[0066] Step S149: Perform shape consistency constraint processing on the global path shape latent variable and the overall shape features of the shoreline contour point sequence in the continuous contour morphology information of the wharf shoreline, and fine-tune the spatial coordinates of the path node spatial coordinates under global shape constraints to generate a path node correction sequence.

[0067] The shoreline contour point sequence is also processed using a global path shape encoder to obtain the overall shoreline shape feature vector. The cosine similarity between the global path shape latent variable and the overall shoreline shape feature vector is calculated. If the cosine similarity is lower than a preset similarity threshold, the secondary correction result of the path node spatial coordinates is fine-tuned through gradient backpropagation optimization to improve the cosine similarity between the adjusted path's global shape latent variable and the overall shoreline shape feature. The resulting fine-tuned path node correction sequence is then obtained.

[0068] Step S1410: Input the path node correction sequence and the path node preliminary correction attitude angle into the attitude angle fine-tuning layer of the path dynamic fine-tuning network. The attitude angle fine-tuning layer recalculates the path tangent direction based on the coordinate difference vector of adjacent path nodes in the path node correction sequence, and performs fine-tuning processing on the path node preliminary correction attitude angle according to the path tangent direction and the hydrodynamic coupling feature vector to generate the path node correction attitude angle sequence.

[0069] For each path node in the path node correction sequence, the coordinate difference vector between that node and the next node is taken as the path tangent direction. The horizontal component direction angle of this path tangent direction is taken as the attitude angle reference value. The drift angle, which is the result of the sway velocity u and pitch velocity v in the hydrodynamic coupling characteristic vector, is subtracted from this attitude angle reference value to obtain the fine-tuned path node corrected attitude angle. All path node corrected attitude angles are arranged according to the node number to form the path node corrected attitude angle sequence. The path node correction sequence and the path node corrected attitude angle sequence are combined to form the automatic berthing dynamic path planning result.

[0070] Step S150: Generate a set of berthing control commands based on the automatic berthing dynamic path planning results. The set of berthing control commands is used to drive the propulsion system and steering gear system of the target ship to perform berthing movements along the automatic berthing dynamic path planning results.

[0071] The path node correction sequence and path node attitude angle correction sequence from the automatic berthing dynamic path planning results are converted into control commands executable by the target vessel. For each path node, the travel distance from the current node to the next node and the desired heading angle are calculated. The travel distance is converted into thrust commands for the propulsion system, and the difference between the desired heading angle and the current heading angle is converted into rudder angle commands for the steering system. The thrust commands and rudder angle commands for all path nodes are arranged sequentially to form the berthing control command set. The berthing control command set is sent to the propulsion system and steering system for execution in real time via the ship's motion control bus.

[0072] Step S210: Obtain the historical navigation perception data set recorded during the historical berthing operation. The historical navigation perception data set includes historical wharf shoreline echo point cloud data, historical instantaneous flow field vector distribution data, and historical real-time six-degree-of-freedom motion parameters.

[0073] Multiple sets of historical berthing operation records were extracted from the ship berthing history database by berthing scenario type. Each set of historical berthing operation records contains complete navigation perception data stored in time sequence during that berthing process: historical quay echo point cloud data collected by shipborne lidar, with each laser echo point containing three-dimensional spatial coordinates and echo intensity values ​​in the ship's coordinate system; historical instantaneous flow field vector distribution data collected by shipborne acoustic Doppler current profiler, with each grid node storing the east and north components of the flow velocity; and historical real-time six-degree-of-freedom motion parameters collected by the shipborne inertial navigation system, including sway velocity, pitch velocity, heave velocity, roll angular velocity, pitch angular velocity, and bow angular velocity. The three types of data are aligned to a common Coordinated Universal Time (UTC) timestamp through a ship network time synchronization protocol.

[0074] Step S220: Construct a training sample data set based on the historical navigation perception data set, wherein the training sample data set includes historical berth trend path annotation data and historical path dynamic fine-tuning annotation data.

[0075] Step S221: Extract historical wharf shoreline echo point cloud data from the historical navigation perception data set, perform historical shoreline geometric morphology analysis processing on the historical wharf shoreline echo point cloud data, and generate continuous contour morphology information of historical wharf shoreline and spatial location reference identifier of historical berthing target berths.

[0076] The historical shoreline geometric morphology analysis and processing operation adopts the same processing method as steps S121 to S128. First, the historical wharf shoreline echo point cloud data is downsampled into a cubic voxel grid with fixed side lengths. The data point closest to the geometric center within each cubic voxel grid is retained to generate a sparse shoreline point cloud set. For each retained point in the sparse shoreline point cloud set, a set of neighborhood points with a fixed radius is taken as the center. Principal component analysis is performed on this neighborhood point set to calculate the local surface normal vector. The standard deviation of the angle between the normal vectors of each point in the neighborhood is calculated as the normal vector deflection degree. The root mean square of the vertical distance from each point in the neighborhood to the local fitting plane is calculated as the neighborhood surface undulation degree. A lower boundary threshold and an upper boundary threshold for undulation are set, and points whose neighborhood surface undulation degree values ​​are between the upper and lower boundary thresholds are extracted as shoreline feature boundary points. The set of shoreline feature boundary points is input into a pre-constructed shoreline contour growth network. A hidden state vector is generated through point-by-point state updates by a gated recurrent unit. An enhanced hidden state vector is generated by fusing neighborhood context information through an adjacent point attention convergence layer. This vector is then labeled as either shoreline continuous contour constituent points or shoreline break discontinuities by a boundary point classification output layer. The shoreline continuous contour constituent points are then connected in an ordered manner according to spatial proximity. The curvature of the arc determined by every three adjacent contour points is calculated to generate historical wharf shoreline continuous contour morphology information. The static spatial coordinates of the pre-marked target berth in this historical berthing operation are obtained and spatially registered with the shoreline contour point sequence to generate the berth center point coordinates and the berth leading edge line segment direction vector as the spatial location reference identifier for the historical target berth.

[0077] Step S222: Extract historical instantaneous flow field vector distribution data and historical real-time six-degree-of-freedom motion parameters from the historical navigation perception data set, and construct historical dynamic feasible domain constraint boundaries based on the historical instantaneous flow field vector distribution data and the historical real-time six-degree-of-freedom motion parameters.

[0078] The construction method adopts the same processing method as steps S129 to S1210. First, a flow field velocity vector field is constructed from historical instantaneous flow field vector distribution data. Then, the ship's instantaneous motion velocity vector is synthesized from the sway velocity and pitch velocity in the historical real-time six-degree-of-freedom motion parameters. The ship's instantaneous motion velocity vector is then vector-superimposed with the velocity vectors at corresponding positions in the flow field velocity vector field to generate a comprehensive disturbance velocity field. Based on the comprehensive disturbance velocity field, and combining the minimum safe distance between the ship and the wharf shoreline, the minimum safe distance between the ship and shallow water areas or obstacles, and the prohibited area where the flow velocity in the comprehensive disturbance velocity field exceeds the safe flow velocity threshold, a potential field boundary surface composed of triangular facets is generated as the historical dynamic feasible domain constraint boundary.

[0079] Step S223: Obtain the ideal berthing path trajectory manually recorded during historical berthing operations, and decompose the ideal berthing path trajectory into a historical path node sequence and a historical path node attitude angle sequence, as historical berth trend path annotation data.

[0080] The ideal berthing path trajectory is a reference path curve manually drawn by experienced pilots based on their experience and port operating procedures in this historical berthing scenario. This reference path curve is sampled at fixed arc length intervals, and the spatial coordinates of each sampling point constitute a sequence of historical path nodes. At each sampling point, the tangent direction of the reference path curve is calculated; the angle between the tangent direction and true north is the ideal ship heading angle for that node. The ideal ship heading angles of all sampling points constitute a sequence of historical path node attitude angles. The historical path node sequence and the historical path node attitude angle sequence together serve as the ground truth for labeling the berth tendency path for this training sample.

[0081] Step S224: For the historical path node sequence in the historical berth trend path annotation data, generate historical hydrodynamic disturbance data based on the historical instantaneous flow field vector distribution data and the historical real-time six-degree-of-freedom motion parameters, and use the historical path node offset under the action of the historical hydrodynamic disturbance data as historical path offset annotation data.

[0082] For each historical path node in the historical path node sequence, the instantaneous flow field velocity vector at the spatial location of that node is extracted. Combined with the real-time six-degree-of-freedom motion parameters of the target ship at that moment, the drift force generated by the hydrodynamic action on the ship is calculated using a ship motion mathematical model. The actual deviation of the ship from the ideal path caused by the drift force is the historical hydrodynamic disturbance data at that node. The negative value of the deviation is used as the historical path deviation label data, indicating the adjustment amount required to correct the deviation back to the ideal path.

[0083] Step S225: Combine the historical path offset annotation data with the historical path node sequence and the historical path node attitude angle sequence to generate historical path dynamic fine-tuning annotation data, wherein the historical path dynamic fine-tuning annotation data includes the spatial coordinate correction amount and attitude angle correction amount of the historical path nodes.

[0084] The spatial coordinate correction of historical path nodes is a negative value of the historical path offset annotation data calculated in step S224, representing the correction amount to bring the deviated path nodes back to their ideal positions. The attitude angle correction is the drift angle, i.e., the angle between the flow field velocity direction and the ship's heading. The combination of the spatial coordinate correction and attitude angle correction of historical path nodes constitutes the historical path dynamic fine-tuning annotation data, which serves as the ground truth annotation for training the path dynamic fine-tuning network.

[0085] Step S226: Combine the historical continuous contour morphology information of the wharf shoreline, the spatial location reference identifier of the historical berthing target berth, the historical dynamic feasible domain constraint boundary, and the historical berth trend path annotation data into training samples for the berth trend path generation network.

[0086] A set of training samples for the berth tendency path generation network consists of two parts: input data and labeled data. The input data includes continuous contour morphology information of the historical wharf shoreline, spatial location benchmarks of historical target berths, and historical dynamic feasible domain constraint boundaries. The labeled data consists of historical path node sequences and attitude angle sequences from the historical berth tendency path labeled data. Multiple sets of samples constitute a complete training sample set for training the berth tendency path generation network.

[0087] Step S227: Combine the historical path dynamic fine-tuning annotation data with the corresponding historical wharf shoreline continuous contour morphology information, historical instantaneous flow field vector distribution data and historical real-time six-degree-of-freedom motion parameters to form a path dynamic fine-tuning network training sample.

[0088] A set of training samples for the path dynamic fine-tuning network consists of two parts: input data and labeled data. The input data includes historical continuous contour morphology information of the wharf shoreline, historical instantaneous flow field vector distribution data, and historical real-time six-degree-of-freedom motion parameters. The labeled data includes the spatial coordinate corrections and attitude angle corrections of historical path nodes from the historical path dynamic fine-tuning labeled data. Multiple sets of samples constitute a complete training sample set for training the path dynamic fine-tuning network.

[0089] Step S230: Perform joint and alternating training on the berth tendency path generation network and the path dynamic fine-tuning network using the training sample data set to generate the trained berth tendency path generation network and the trained path dynamic fine-tuning network.

[0090] Step S231: Input the historical continuous contour morphology information of the wharf shoreline, the spatial location benchmark of the historical berthing target berth, and the historical dynamic feasible domain constraint boundary from the training samples of the berth tendency path generation network into the berth tendency path generation network to generate a predicted berth tendency path.

[0091] The berth tendency path generation network is invoked for forward propagation according to the method in step S130. The sequence of shoreline contour points from the historical continuous shoreline contour morphology information, the coordinates of the berth center point and the direction vector of the berth leading edge line segment from the spatial location benchmark of the historical berthing target berth, and the geometric parameters of the potential field boundary surface of the historical dynamic feasible domain constraint boundary are uniformly encoded into a multi-channel state tensor. The multi-channel state tensor is sequentially input into a multi-layer spatial graph convolutional layer to extract a spatial fusion feature map, then input into a bidirectional long short-term memory network layer to generate a time-series feature vector, then through an attention decoding layer to generate an attention weight distribution and weighted summation to generate a spatial coordinate prediction vector, constructing a prediction path node sequence, generating a prediction attitude angle sequence through an attitude angle prediction branch, and obtaining the predicted berth tendency path after safety distance projection correction and density homogenization processing.

[0092] Step S232: Calculate the spatial coordinate deviation of the path nodes between the predicted berth trend path and the historical path node sequence in the historical berth trend path annotation data, and calculate the attitude angle deviation between the attitude angle of the predicted path nodes in the predicted berth trend path and the attitude angle sequence of the historical path nodes.

[0093] The path node spatial coordinate deviation is the average Euclidean distance between the spatial coordinates of each predicted path node in the predicted berth orientation path and the spatial coordinates of the corresponding historical path node in the historical path node sequence. The attitude angle deviation is the average absolute value of the angle differences between the attitude angles of the predicted path nodes and the attitude angles of the corresponding historical path nodes.

[0094] Step S233: Construct a berth approach path generation network loss function based on the spatial coordinate deviation of the path nodes and the attitude angle deviation, and update the network weight parameters of the berth approach path generation network through the backpropagation algorithm.

[0095] The loss function of the berth orientation path generation network is the sum of spatial coordinate deviation multiplied by the first weight coefficient and attitude angle deviation multiplied by the second weight coefficient. Minimizing this loss function is the optimization objective. The Adam optimizer is used to calculate the gradient of the loss function with respect to the trainable weight parameters and bias parameters in the multi-layer spatial graph convolutional layer, bidirectional long short-term memory network layer, and attention decoding layer. These network weight parameters are then updated by backpropagation along the gradient descent direction.

[0096] Step S234: Freeze the network weight parameters of the berth tendency path generation network in the current training round, and use the predicted berth tendency path output by the berth tendency path generation network as the initial path input in the training samples of the path dynamic fine-tuning network.

[0097] The freeze operation sets all trainable weight parameters and bias parameters in the berth tendency path generation network to a non-updateable state, retaining only the forward inference function.

[0098] Step S235: Input the historical instantaneous flow field vector distribution data and historical real-time six-degree-of-freedom motion parameters from the training samples of the path dynamic fine-tuning network into the path dynamic fine-tuning network, perform path smoothing and attitude optimization processing on the initial path input, and generate a predicted dynamic fine-tuning path.

[0099] The path dynamic fine-tuning network is invoked according to the method in step S140 to perform forward propagation and generate a predicted dynamic fine-tuning path, which includes the predicted corrected path node sequence and the predicted path node corrected attitude angle sequence.

[0100] Step S236: Calculate the deviation between the spatial coordinate correction and attitude angle correction of the historical path nodes in the predicted dynamic fine-tuning path and the historical path dynamic fine-tuning annotation data, construct the path dynamic fine-tuning network loss function, and update the network weight parameters of the path dynamic fine-tuning network through the backpropagation algorithm.

[0101] The loss function of the path dynamic fine-tuning network is the Euclidean distance error between the predicted and corrected spatial coordinates of the path nodes and the target spatial coordinates corresponding to the corrected labeled spatial coordinates, plus a weighted sum of the angle error between the predicted and corrected attitude angles of the path nodes and the target attitude angles corresponding to the corrected labeled attitude angles. Minimizing this loss function value is the optimization objective. This is used to update the trainable weights and biases in the attitude angle gating fine-tuning unit, the lateral offset compensation layer, the path curvature linkage optimization layer, the global path shape encoder, and the attitude angle fine-tuning layer of the path dynamic fine-tuning network.

[0102] Step S237: Repeatedly execute the network weight parameter update of the berth tendency path generation network and the network weight parameter update of the path dynamic fine-tuning network until the loss function of the berth tendency path generation network and the loss function of the path dynamic fine-tuning network converge to the preset training termination condition, thereby generating the trained berth tendency path generation network and the trained path dynamic fine-tuning network.

[0103] During the alternating training process, the network weight parameters of the berth-oriented path generation network are first unfrozen, while the network weight parameters of the path dynamic fine-tuning network are frozen. Steps S231 to S233 are executed to complete one update of the network weight parameters of the berth-oriented path generation network. Then, the network weight parameters of the berth-oriented path generation network are frozen, and the network weight parameters of the path dynamic fine-tuning network are unfrozen. Steps S234 to S236 are executed to complete one update of the network weight parameters of the path dynamic fine-tuning network. The above alternating training cycle is repeated, and the loss function values ​​of the berth-oriented path generation network and the path dynamic fine-tuning network are recorded after each cycle. When the fluctuation range of the berth-oriented path generation network loss function value is less than the preset first convergence threshold and the fluctuation range of the path dynamic fine-tuning network loss function value is less than the preset second convergence threshold for multiple consecutive cycles, the training is considered to have converged, and the alternating training is stopped. All network weight parameters of the berth-oriented path generation network at this time are saved as the trained berth-oriented path generation network, and all network weight parameters of the path dynamic fine-tuning network at this time are saved as the trained path dynamic fine-tuning network.

[0104] Step S310: Obtain the set of berthing motion feedback sensing data collected in real time by the shipborne sensing device during the berthing motion of the target vessel. The set of berthing motion feedback sensing data includes real-time updated quay shoreline echo point cloud data, real-time updated instantaneous flow field vector distribution data, and real-time updated real-time six-degree-of-freedom motion parameters.

[0105] During the berthing movement of the target vessel according to the automatic berthing dynamic path planning results, the shipborne lidar continuously outputs real-time updated quay echo point cloud data at a fixed frequency. Each laser echo point in the real-time updated quay echo point cloud data contains three-dimensional spatial coordinates and echo intensity values ​​in the ship's coordinate system with the current vessel position as the origin. The shipborne acoustic Doppler current profiler continuously outputs real-time updated instantaneous flow field vector distribution data at a fixed frequency. Each grid node in the real-time updated instantaneous flow field vector distribution data stores the east and north components of the current flow velocity. The shipborne inertial navigation system continuously outputs real-time updated six-degree-of-freedom motion parameters, including the current sway velocity, pitch velocity, heave velocity, roll angular velocity, pitch angular velocity, and bow angular velocity. These three types of real-time updated data are aligned to a common Coordinated Universal Time (UTC) timestamp through a ship network time synchronization protocol, together forming the berthing movement feedback sensing data set. The berthing movement feedback sensing data set reflects the constantly changing external environment and the target vessel's own motion state during the berthing movement.

[0106] Step S320: Perform online rolling correction processing on the automatic berthing dynamic path planning result based on the berthing motion feedback sensing data set to generate the online corrected automatic berthing dynamic path planning result.

[0107] Each time a new set of berthing motion feedback sensing data is received, an online rolling correction of the automatic berthing dynamic path planning results is triggered. The online rolling correction process compares the target vessel's current actual navigation state with the preset expected state in the automatic berthing dynamic path planning results, calculates the deviation, and feeds the deviation back to the pre-built path deviation compensation network to generate compensation corrections for path nodes and attitude angles that have not yet been executed, thereby updating the automatic berthing dynamic path planning results.

[0108] For example, step S321: extract real-time updated wharf shoreline echo point cloud data from the berthing motion feedback sensing data set, perform real-time shoreline geometric morphology analysis processing on the real-time updated wharf shoreline echo point cloud data, and generate real-time updated continuous contour morphology information of the wharf shoreline and real-time updated spatial location reference identifier of the berthing target berth.

[0109] The real-time shoreline geometry analysis process employs the same methods as steps S121 to S128. During the target vessel's berthing movement, the shipborne lidar continuously scans the wharf shoreline area, and the returned real-time updated shoreline echo point cloud data reflects the changes in observation perspective and the real-time relative positional relationship of the shoreline structure due to vessel movement. The real-time updated shoreline echo point cloud data is then re-processed with point cloud voxel downsampling, local point neighborhood normal vector consistency analysis, shoreline feature boundary point extraction, shoreline contour growth network inference, ordered connection of points constituting the continuous shoreline contour, and berth spatial registration. This generates real-time updated continuous shoreline contour morphology information and a real-time updated spatial position reference identifier for the target berth. The real-time updated data reflects the true relative geometric relationship between the target vessel and the wharf shoreline at the current moment.

[0110] Step S322: Extract real-time updated instantaneous flow field vector distribution data and real-time updated real-time six-degree-of-freedom motion parameters from the berthing motion feedback sensing data set, and construct a real-time updated dynamic feasible domain constraint boundary based on the real-time updated instantaneous flow field vector distribution data and the real-time updated real-time six-degree-of-freedom motion parameters.

[0111] The construction method employs the same processing approach as steps S129 to S1210. The real-time updated instantaneous flow field vector distribution data reflects the actual flow field conditions in the berthing area at the current moment, and the real-time updated six-degree-of-freedom motion parameters reflect the actual motion state of the target vessel at the current moment. A real-time updated comprehensive disturbance velocity field is generated by vector superposition of the real-time updated flow field velocity vector field and the real-time updated instantaneous vessel velocity vector. This, combined with preset safety boundary rules, generates a real-time updated dynamic feasible region constraint boundary. This real-time updated dynamic feasible region constraint boundary more accurately describes the safety constraints of the waters surrounding the target vessel at the current moment.

[0112] Step S323: Obtain the actual navigation status of the target vessel at the current moment. The actual navigation status includes the actual spatial coordinates of the target vessel and the actual heading angle of the target vessel.

[0113] The actual spatial coordinates of the target vessel are obtained through the shipborne satellite navigation and positioning receiver, and the actual heading angle of the target vessel is obtained through the shipborne inertial navigation system.

[0114] Step S324: Calculate the position deviation vector between the actual spatial coordinates of the target vessel and the spatial coordinates of the next path node in the automatic berthing dynamic path planning result; calculate the angular deviation scalar between the actual heading angle of the target vessel and the path node correction attitude angle of the next path node in the automatic berthing dynamic path planning result.

[0115] The position deviation vector is the vector difference between the spatial coordinates of the next path node and the actual spatial coordinates of the target vessel in the automatic berthing dynamic path planning result. The angle deviation scalar is the difference between the path node correction attitude angle of the next path node and the actual heading angle of the target vessel, normalized to the interval [-π, π].

[0116] Step S325: The position deviation vector, the angle deviation scalar, the real-time updated continuous contour morphology information of the wharf shoreline, and the real-time updated dynamic feasible domain constraint boundary are input into the multi-layer spatial graph convolutional layer of the pre-constructed path deviation compensation network for spatial feature extraction processing to generate a deviation compensation spatial fusion feature map.

[0117] The structure of the path deviation compensation network is the same as that of the multi-layer spatial graph convolutional layer in the berth tendency path generation network. The three components of the position deviation vector and the angular deviation scalar are used as the ship deviation state features, which are jointly encoded into a multi-channel deviation state tensor along with the shoreline contour point sequence from the real-time updated continuous shoreline contour morphology information and the potential field boundary surface geometric parameters of the real-time updated dynamic feasible domain constraint boundary. The multi-layer spatial graph convolutional layer extracts spatial features from this multi-channel deviation state tensor to generate a deviation compensation spatial fusion feature map.

[0118] Step S326: The deviation compensation spatial fusion feature map is processed by performing temporal dependency modeling on the bidirectional long short-term memory network layer of the path deviation compensation network to generate a deviation compensation time series feature vector.

[0119] The deviation compensation spatial fusion feature map is expanded into a sequence along the berthing direction and then input into a bidirectional long short-term memory network layer to generate a deviation compensation time series feature vector containing bidirectional contextual information.

[0120] Step S327: Input the deviation compensation time series feature vector into the attention decoding layer of the path deviation compensation network for global feature interaction processing to generate path deviation compensation vector and attitude deviation compensation vector.

[0121] The attention decoding layer adopts the same multi-head self-attention mechanism structure as the berth tendency path generation network. After the deviation compensation time series feature vector is processed by the attention decoding layer, it outputs two branches: one branch outputs the path deviation compensation vector, which contains the spatial coordinate compensation amount for each path node that has not yet been executed in the automatic berthing dynamic path planning result; the other branch outputs the attitude deviation compensation vector, which contains the attitude angle compensation amount for each path node that has not yet been executed.

[0122] Step S328: Based on the path deviation compensation vector, perform spatial coordinate offset adjustment processing on the path node correction sequence that has not yet been executed in the automatic docking dynamic path planning result; based on the attitude deviation compensation vector, perform attitude angle offset adjustment processing on the path node correction attitude angle sequence that has not yet been executed in the automatic docking dynamic path planning result, and generate the online corrected automatic docking dynamic path planning result.

[0123] In the automatic berthing dynamic path planning results, the path nodes that have not yet been executed are those whose node numbers are greater than the nearest path node number already reached by the target vessel. For these path nodes, their spatial coordinates are increased by the corresponding spatial coordinate compensation amount in the path deviation compensation vector to obtain the adjusted spatial coordinates, and their path node corrected attitude angles are increased by the corresponding attitude angle compensation amount in the attitude deviation compensation vector to obtain the adjusted attitude angles. The adjusted path node correction sequence and the path node corrected attitude angle sequence constitute the online-corrected automatic berthing dynamic path planning results. This online rolling correction process is continuously executed during the target vessel's berthing motion, triggering an online correction each time a new set of berthing motion feedback sensing data is acquired.

[0124] For example, the method may further include: step S410: obtaining the water pressure distribution data of the outer hull plate of the target vessel at the current berthing time, wherein the water pressure distribution data of the outer hull plate includes the measured water pressure values ​​collected at multiple measuring points along the outer hull plate of the target vessel.

[0125] The water pressure distribution data of the hull plating is collected by pressure sensors installed at various measuring points on the hull plating of the target vessel. Each measuring point outputs the normal pressure value of the water on the hull surface at a fixed frequency. The measured water pressure values ​​from multiple measuring points together constitute the water pressure distribution data of the hull plating. The spatial coordinates of each measuring point in the hull coordinate system are pre-defined by the vessel design drawings.

[0126] Step S420: Map the water pressure distribution data of the hull plate to the three-dimensional geometric model of the target ship's hull plate according to the spatial coordinates of each measuring point, and generate a water pressure spatial distribution field of the hull plate. The water pressure spatial distribution field of the hull plate includes the water pressure intensity and the normal vector of the water pressure direction on each hull plate grid cell.

[0127] The three-dimensional hull plate geometric model of the target vessel is a pre-constructed hull surface geometric model composed of multiple quadrilateral or triangular mesh elements. Using the Kriging space interpolation algorithm, the measured water pressure values ​​at each discrete measurement point are interpolated to the center point of each mesh element in the three-dimensional hull plate geometric model, obtaining the water pressure intensity of each mesh element. The normal vector of the water pressure direction is the unit normal vector pointing outwards from the hull surface at the center point of each mesh element. The water pressure intensity and the normal vector of the water pressure direction together constitute the spatial distribution field of water pressure on the hull plate.

[0128] Step S430: Based on the spatial distribution field of water pressure on the hull plate, for each hull plate grid cell, the water pressure intensity is multiplied by the area of ​​the grid cell to obtain the water pressure vector. The direction of the water pressure vector is the normal vector of the water pressure direction. The water pressure vectors of all grid cells are summed and integrated to generate the resultant vector of the asymmetric hydrodynamic lateral force on the target ship hull. The resultant vector of the asymmetric hydrodynamic lateral force includes the direction and intensity of the resultant lateral force.

[0129] For the k-th grid cell, its area is Ak, the water pressure intensity is Pk, and the normal vector of the water pressure direction is nk. The water pressure vector of this grid cell is Fk = Pk × Ak × nk. Integrating and summing the water pressure vectors of all grid cells along the hull surface yields the total resultant water pressure vector Ftotal = sum(Fk). Ftotal is decomposed into two components along the longitudinal and transverse directions of the hull. The transverse component is the resultant vector of the asymmetric hydrodynamic lateral force, Flat. The asymmetric hydrodynamic lateral force is the lateral thrust generated by the asymmetric distribution of water pressure on both sides of the hull.

[0130] Step S440: The resultant force vector of the asymmetric hydrodynamic lateral force is superimposed with the flow field velocity vector in the real-time flow field vector distribution data to generate a corrected hydrodynamic environment comprehensive force vector field. The hydrodynamic environment comprehensive force vector field includes the comprehensive hydrodynamic direction and intensity at each hull plate grid unit.

[0131] The resultant vector of asymmetric hydrodynamic lateral forces, Flat, is distributed to each hull plate grid cell. This vector is then superimposed with the original water pressure vector at each grid cell. Finally, the vector is combined with the real-time flow velocity vector corresponding to the spatial location of each grid cell, converted to dynamic pressure using Bernoulli's equation, to generate the comprehensive hydrodynamic vector at each hull plate grid cell. The comprehensive hydrodynamic vectors of all grid cells constitute the comprehensive force vector field of the hydrodynamic environment.

[0132] Step S450: Based on the comprehensive hydrodynamic environment force vector field, perform non-uniform expansion correction processing on the potential field boundary surface of the dynamic feasible domain constraint boundary. Offset the potential field boundary surface corresponding to the region where the comprehensive hydrodynamic environment force vector intensity exceeds the preset force intensity threshold to the outside to generate an asymmetric dynamic feasible domain constraint boundary. Replace the dynamic feasible domain constraint boundary with the asymmetric dynamic feasible domain constraint boundary. Recall the berth tendency path generation network to perform multi-constraint path planning processing on the continuous contour morphology information of the wharf shoreline, the spatial location benchmark, and the asymmetric dynamic feasible domain constraint boundary to generate the corrected initial berth tendency path.

[0133] For each vertex of a triangular facet on the potential field boundary surface, the intensity of the combined hydrodynamic force vector at the corresponding hull plate mesh element is found. If this intensity exceeds a preset force intensity threshold, the vertex is translated outward along the direction of the outward normal vector of the hull plate at that point, with the translation amount proportional to the magnitude of the force intensity exceeding the threshold. After expansion, all modified triangular facet vertices reconstruct the asymmetric dynamic feasible region constraint boundary. The asymmetric dynamic feasible region constraint boundary replaces the dynamic feasible region constraint boundary, and the berth tendency path generation network is called again in step S130 to generate the corrected initial berth tendency path.

[0134] Step S460: Replace the initial berth orientation path with the corrected initial berth orientation path, re-input the path dynamic fine-tuning network for path smoothing and attitude optimization, generate the corrected automatic berthing dynamic path planning result, regenerate the corrected berthing control command set based on the corrected automatic berthing dynamic path planning result, and drive the target vessel to adjust its berthing trajectory along the corrected automatic berthing dynamic path planning result.

[0135] Following step S140, the modified initial berth orientation path is input into the path dynamic fine-tuning network to generate the modified automatic berthing dynamic path planning result. Following step S150, the modified berthing control command set is generated and sent to the propulsion system and steering system for execution.

[0136] Step S510: Obtain three-dimensional laser scanning point cloud data of the wharf structure surrounding the target berth. The three-dimensional laser scanning point cloud data includes the surface shape point cloud of the wharf fender structure and the spatial position point cloud of the wharf mooring bollards.

[0137] The 3D laser scanning point cloud data is pre-collected by a ground-based fixed laser scanner deployed at the berth's edge before berthing operations. The ground-based fixed laser scanner performs a high-resolution panoramic scan of the berth's fender structure and mooring bollards surrounding the target berth, outputting 3D laser scanning point cloud data containing the 3D spatial coordinates and echo intensity values ​​of each scanned point. A point cloud semantic segmentation algorithm is then used to divide the 3D laser scanning point cloud data into a subset representing the surface shape of the berth's fender structure and a subset representing the spatial location of the mooring bollards.

[0138] Step S520: Perform elastic deformation range analysis on the point cloud of the surface shape of the wharf fender structure in the three-dimensional laser scanning point cloud data, and combine the elastic deformation characteristic parameters of the wharf fender material to generate the boundary of the compressible deformation area of ​​the wharf fender structure. The boundary of the compressible deformation area includes the spatial range corresponding to the maximum compressive deformation of the wharf fender structure under the squeezing action of the target ship during berthing.

[0139] The elastic deformation characteristics of the quay fender material are obtained from the technical specifications provided by the fender manufacturer, including the elastic modulus, maximum compressibility, and original geometric thickness of the fender material. For each fender surface point in the point cloud of the quay fender structure's surface shape, the compression direction is taken as the normal vector direction, and the maximum compressive deformation obtained by multiplying the original geometric thickness of the fender by the maximum compressibility is taken as the compressive displacement. The point is then translated along the opposite direction of the normal vector by the maximum compressive deformation distance. The translated point is the surface position of the fender under maximum compression. All the surface points corresponding to the maximum compression state constitute the inner boundary surface of the quay fender structure, and the surface formed by the original quay fender structure's surface shape point cloud is the outer boundary surface. The region between the inner and outer boundary surfaces is the boundary of the compressible deformation region of the quay fender structure.

[0140] Step S530: Perform mooring cable accessibility analysis on the spatial location point cloud of the dock mooring bollard in the three-dimensional laser scanning point cloud data. Based on the maximum extension length of the target ship's mooring cable, generate a connection accessibility spatial domain for the mooring cable. The connection accessibility spatial domain includes a spherical region with the dock mooring bollard as the center and the maximum extension length as the radius.

[0141] Each mooring bollard in the spatial location point cloud stores its spatial coordinates. For each mooring bollard, the maximum extension length of the target vessel's mooring cable is determined by the vessel's mooring equipment technical parameters. The reachable spatial domain of the mooring cable connection is a spherical region centered on the spatial coordinates of that mooring bollard and with the maximum extension length as its radius. If multiple mooring bollards exist, the reachable spatial domain of the mooring cable connection is the union of all spherical regions.

[0142] Step S540: Perform terminal path segment constraint analysis on the boundary of the compressible deformation area of ​​the wharf fender structure and the reachable space of the mooring cable with the initial berth approach path. Calculate whether the last few path nodes of the initial berth approach path are within the overlap range of the boundary of the compressible deformation area and the reachable space. When the last few path nodes of the initial berth approach path exceed the overlap range of the boundary of the compressible deformation area and the reachable space, perform path node replanning on the terminal path segment of the initial berth approach path. Path nodes that exceed the range are backed up in the opposite direction of the initial berth approach path and corrected towards the overlap range to generate the initial berth approach path after terminal path segment correction.

[0143] A predetermined number of path nodes at the end of the initial berth orientation path sequence are selected as the final path segment. For each path node in the final path segment, it is determined whether the spatial coordinates of the path node simultaneously satisfy two conditions: it is located between the inner and outer boundary surfaces of the compressible deformation zone boundary of the wharf fender structure, and it is located within the reachable space of the mooring cables. Satisfying both conditions indicates that the path node can utilize the fender's elasticity to buffer the impact of ship berthing, and it is also within the reachable range of the mooring cables. If any path node in the final path segment does not meet the above overlap conditions, then the path node is out of range. The out-of-range path node is retracted in the opposite direction of the initial berth orientation path, with the retraction direction being the opposite of the path tangent direction at that path node. After retraction, using the out-of-range constraint, the retracted path node is projected and corrected towards the nearest boundary within the overlap range. After correction, the continuity between path nodes is maintained, generating the corrected initial berth orientation path for the final path segment.

[0144] Step S550: Replace the initial berth approach path with the corrected terminal path segment, re-input the path dynamic fine-tuning network for path smoothing and attitude optimization, and generate an automatic berthing dynamic path planning result containing the corrected path node sequence and the corrected attitude angle sequence of the path nodes. Based on the automatic berthing dynamic path planning result containing the corrected path node sequence and the corrected attitude angle sequence of the path nodes, generate a corrected berthing control command set to drive the target vessel to perform the final berthing movement according to the corrected terminal path segment.

[0145] Following step S140, the initial berth orientation path after the terminal path segment correction is input into the path dynamic fine-tuning network. Steps S141 to S1410 are then executed sequentially to generate the corrected automatic berthing dynamic path planning result. Following step S150, the corrected berthing control command set is generated and sent to the propulsion system and steering system.

[0146] Step S610: Obtain the automatic identification system broadcast data of surrounding vessels received by the target vessel during the berthing process. The automatic identification system broadcast data of surrounding vessels includes the spatial position coordinates of surrounding vessels, the speed vector of surrounding vessels, and the predicted track lines of surrounding vessels.

[0147] The broadcast data from the Automatic Identification System (AIS) of surrounding vessels is received by the AIS transceiver installed on the target vessel. Surrounding vessels broadcast their static and dynamic information at fixed time intervals. The dynamic information includes the vessel's GPS coordinates, speed relative to ground, and heading relative to ground. The static information includes the vessel's identifier and type. The speed and heading relative to ground of the surrounding vessels are combined into a speed vector. The received multiple broadcast data are then used to calculate the predicted track of the surrounding vessels using a Kalman filter algorithm.

[0148] Step S620: Convert the spatial coordinates of the surrounding vessels in the broadcast data of the Automatic Identification System for Surrounding Vessels to a relative coordinate system with the current position of the target vessel as the origin, and generate a relative position distribution map of the surrounding vessels. The relative position distribution map of the surrounding vessels includes the distance and azimuth of each surrounding vessel relative to the target vessel.

[0149] A relative coordinate system is established with the target vessel's spatial coordinates as the origin and its heading angle as the longitudinal axis. The GPS spatial coordinates of each surrounding vessel are transformed to this relative coordinate system through translation and rotation. The relative positions of each surrounding vessel after transformation are expressed as distance and relative azimuth. The relative distances and relative azimuths of all surrounding vessels constitute a distribution map of their relative positions.

[0150] Step S630: Based on the speed vectors of the surrounding vessels and the predicted tracks of the surrounding vessels, extrapolate and predict the tracks of the surrounding vessels at future times to generate a future track prediction zone for the surrounding vessels. The future track prediction zone for the surrounding vessels includes the water space area that the surrounding vessels may occupy within the future time window.

[0151] Using the current moment as the time origin, for each surrounding vessel, its current spatial position is used as the starting point. A preset prediction time window is then pushed forward along its predicted trajectory direction at its current speed to obtain the predicted spatial positions of that surrounding vessel at various future moments. At each predicted moment, a circular danger zone is delineated with that predicted spatial position as the center and the safe collision avoidance distance of surrounding vessels as the radius. The time union of all circular danger zones at all predicted moments constitutes the future trajectory prediction zone for that surrounding vessel. The spatial union of all future trajectory prediction zones for surrounding vessels constitutes the future trajectory prediction zone for all surrounding vessels.

[0152] Step S640: Perform spatiotemporal conflict detection processing on the predicted future trajectories of surrounding vessels and the initial berth approach path, determine whether there are conflicting path nodes on the path node sequence of the initial berth approach path that overlap with the predicted future trajectories of surrounding vessels in both space and time, mark the conflicting path nodes with spatiotemporal overlap as path nodes to be avoided, and extract the path node number and spatial coordinates of the path nodes to be avoided in the initial berth approach path.

[0153] For each path node in the initial berth approach path sequence, the expected arrival time of the target vessel at that path node is calculated based on its path node number and the path node spacing. At this expected arrival time, the spatial coordinates of the path node are checked to see if they fall within any circular danger zone of the predicted future tracks of surrounding vessels at that moment. If they do, the path node is determined to have a spatiotemporal conflict and is marked as a path node to be avoided. The path node number and spatial coordinates are extracted for all path nodes to be avoided.

[0154] Step S650: Based on the spatial coordinates of the path nodes to be avoided and the spatial boundary of the predicted future course of surrounding vessels, generate a lateral avoidance offset vector and a longitudinal avoidance offset vector for each path node to be avoided. Use the lateral avoidance offset vector and the longitudinal avoidance offset vector to perform offset correction processing on the spatial coordinates of the path nodes to be avoided, and generate a path node sequence after avoidance correction.

[0155] For each path node to be avoided, the lateral avoidance offset vector is perpendicular to the tangent of the initial berth approach path at that path node and points away from the center of the predicted future course zone for surrounding vessels. Its magnitude is the safe distance between the path node and the boundary of the predicted future course zone for surrounding vessels. The longitudinal avoidance offset vector is the spatial offset corresponding to advancing or delaying the arrival time of the path node along the tangent of the path path at that path node, to further reduce the probability of conflict. The original spatial coordinates of the path nodes are added to the lateral and longitudinal avoidance offset vectors to obtain the offset-corrected spatial coordinates. All corrected path nodes are then merged with unmarked path nodes to form the corrected path node sequence.

[0156] Step S660: Merge the path node sequence after collision avoidance correction with the original path node sequence that is not marked as a path node to be avoided in the original path node number order to generate an initial berth approach path containing the path node sequence after collision avoidance correction. Replace the initial berth approach path with the initial berth approach path and re-input it into the path dynamic fine-tuning network for path smoothing and attitude optimization processing to generate an automatic berthing dynamic path planning result containing the path node sequence after collision avoidance correction and the path node corrected attitude angle sequence. Generate a corrected berthing control command set based on the automatic berthing dynamic path planning result to drive the target vessel to perform berthing movement along the corrected path to avoid surrounding vessels.

[0157] In step S140, the initial berth approach path input path dynamic fine-tuning network containing the collision avoidance correction path node sequence is used to generate the corrected automatic berthing dynamic path planning result, and in step S150, a berthing control command set is generated.

[0158] Step S710: Obtain multiple historical successful berthing trajectory data recorded by the target vessel during historical berthing operations. The historical successful berthing trajectory data includes historical initial berth tendency paths and historical automatic berthing dynamic path planning results.

[0159] Historical successful berthing trajectory data is stored in the ship berthing history database, with each historical successful berthing trajectory corresponding to a successfully completed berthing operation. The historical initial berth orientation path is the version of the initial berth orientation path generated in step S130 of this berthing operation before the final correction execution. The historical automatic berthing dynamic path planning result is the automatic berthing dynamic path planning result generated and successfully executed in step S140 of this berthing operation. All historical successful berthing trajectory data together constitute a sample library for learning prior knowledge of berthing paths.

[0160] Step S720: Perform path shape clustering analysis on the historical initial berth trend paths in the multiple historical successful berthing trajectory data, group historical initial berth trend paths with similar path geometry into the same path shape category, generate a set of berth trend path shape categories, perform statistical feature extraction on the historical initial berth trend paths in each berth trend path shape category, and extract the category center path node sequence and category path node spatial distribution variance for each berth trend path shape category.

[0161] The path shape clustering analysis uses dynamic time-warped distance as a metric for the similarity between paths. For any two historical initial berth approach paths, the path node sequences of each path are dynamically time-warped and aligned. The cumulative Euclidean distance between the aligned corresponding path nodes is calculated as the dynamic time-warped distance between the two paths. Using the dynamic time-warped distance matrix between all pairs of historical initial berth approach paths as input, a hierarchical agglomerative clustering algorithm is used for bottom-up clustering. In the initialization of the hierarchical agglomerative clustering algorithm, each path forms its own cluster. In each iteration, the two clusters with the smallest inter-cluster dynamic time-warped distance are merged into a new cluster until the inter-cluster distance exceeds a preset inter-cluster distance threshold, at which point clustering stops, resulting in a set of berth approach path shape categories. For each berth approach path shape category, after dynamic time-warping alignment of all historical initial berth approach paths within that category, the arithmetic mean of the spatial coordinates of each path node is calculated at the corresponding path node, yielding the cluster center path node sequence for that berth approach path shape category. The variance of the spatial coordinates of each path node is calculated at the corresponding path node to obtain the spatial distribution variance of the path node for the berth's directional path shape category. The sequence of path nodes at the category center represents the typical geometric shape of the path in that category, and the spatial distribution variance of the path node represents the degree of dispersion of each path within that category at the corresponding node.

[0162] Step S730: Construct the category center path node sequence and the category path node spatial distribution variance into a berth tendency path prior knowledge base. The berth tendency path prior knowledge base contains typical path templates and path distribution ranges for target ships tending to berth at target berths under different berthing scenarios.

[0163] The berth tendency path prior knowledge base is stored using a relational data table structure. Each berth tendency path shape category corresponds to a row in the data table. The record fields include a category identifier, serialized data of the category's central path node sequence, serialized data of the spatial distribution variance of the category's path nodes, and the corresponding berthing scenario feature label. The berthing scenario feature label includes the berth type identifier of the target berth and the current water area's flow field disturbance intensity level. The berth tendency path shape category with the highest matching degree can be retrieved using the berth type identifier and the flow field disturbance intensity level.

[0164] Step S740: Obtain the berthing scene feature information of the current berthing operation. The berthing scene feature information includes the berth type identifier of the target berth and the current water flow disturbance intensity level.

[0165] The berth type identifier for the target berth is obtained from the port terminal database. Berth types include berths along the shore, breakwater berths, and jetty berths. The current flow field disturbance intensity level is classified into three levels—weak disturbance, moderate disturbance, and strong disturbance—based on the average velocity modulus of each grid node in the current instantaneous flow field vector distribution data.

[0166] Step S750: Based on the berth type identifier and flow field disturbance intensity level in the berthing scene feature information, retrieve the berth trend path shape category with the highest matching degree in the berth trend path prior knowledge base, extract the category center path node sequence of the berth trend path shape category as the prior path template of the current berthing scene, encode the spatial coordinates of the path node sequence of the prior path template into a path prior feature vector, and input it into the attention decoding layer of the berth trend path generation network as an additional global prior attention bias term introduced in the multi-head self-attention mechanism.

[0167] The retrieval matching degree uses a weighted Euclidean distance metric. The current berth type identifier and flow field disturbance intensity level are encoded into a numerical vector format identical to the scene feature labels in the knowledge base records. The weighted Euclidean distance between this vector and the scene feature label vectors of each record in the knowledge base is calculated, and the record with the smallest distance is selected as the berth tendency path shape category with the highest matching degree. The spatial coordinates of all path nodes in the category center path node sequence of this category are sequentially flattened into a one-dimensional vector, which is the path prior feature vector. In the multi-head self-attention mechanism of the attention decoding layer, the original attention score matrix S = Q × KT / sqrt(dk) is transformed into a new attention score matrix S' = S + λ × Vprior after introducing a global prior attention bias term. Here, Vprior is the bias matrix embedded into the path prior feature vector using a trainable linear projection matrix to the same dimension as the attention score matrix, and λ is the prior weight adjustment coefficient. This bias term skews the attention distribution towards the geometric shape direction of the prior path, generating an attention weight distribution guided by the prior path shape.

[0168] Step S760: By introducing the global prior attention bias term into the attention decoding layer, the time series feature vector is subjected to global feature interaction processing with prior constraints to generate an attention weight distribution guided by prior path shape. The attention weight distribution is used to perform weighted summation processing on the corresponding spatial positions of the spatial fusion feature map to generate a path node spatial coordinate prediction vector with prior path shape constraints.

[0169] In the multi-head self-attention mechanism of the attention decoding layer, the attention score matrix S' = Q×KT / sqrt(dk) + λ×Vprior is introduced after a global prior attention bias term is introduced. Here, Q is the query matrix, K is the key matrix, dk is the key vector dimension, Vprior is the bias matrix of the same dimension as Q×KT, mapped from the path prior feature vector generated in step S750 by a trainable linear projection matrix, and λ is the prior weight adjustment coefficient. This bias term injects the shape information of the prior path template into the attention score in the form of an additive bias, causing the attention distribution to focus on regions consistent with the spatial location of the prior path template. Each row of the attention score matrix is ​​normalized using Softmax to obtain the attention weight matrix A' = softmax(S'). The value matrix V is then weighted and summed using this attention weight matrix to obtain the context-aware aggregated feature vector for each path node to be generated. This context-aware aggregated feature vector is input into a feedforward fully connected network, which outputs the spatial coordinate prediction vector for each path node to be generated. Since the distribution of attention weights is guided by the shape of the prior path template, the weighted summation process incorporates the spatial distribution information of the prior path into the aggregated feature vector. As a result, the generated path node spatial coordinate prediction vector can inherit the geometric morphological features of the prior path, thus avoiding the generation of abnormal path nodes that deviate from the empirically optimal path shape in similar berthing scenarios.

[0170] Step S770: Construct a path node sequence under prior constraints based on the path node spatial coordinate prediction vector, generate the corresponding target vessel reference attitude angle sequence under prior constraints through the attitude angle prediction branch of the berth approach path generation network, combine them into an initial berth approach path under prior constraints, replace the initial berth approach path with the initial berth approach path under prior constraints, input the path dynamic fine-tuning network for path smoothing and attitude optimization processing, and generate an automatic berthing dynamic path planning result containing a path node correction sequence under prior constraints and a path node correction attitude angle sequence.

[0171] The initial berth orientation path under prior constraints is input into the path dynamic fine-tuning network according to the method in step S140. The following processing steps are executed in sequence: construction of spatiotemporal sequence samples of path nodes, flow field interpolation, splicing of hydrodynamic coupling features, attitude angle gating fine-tuning, lateral offset compensation, path curvature linkage optimization, global path shape encoding and constraint, and attitude angle fine-tuning. The result is an automatic berthing dynamic path planning result containing the path node correction sequence under prior constraints and the path node corrected attitude angle sequence.

[0172] Figure 2 This application illustrates a deep learning-based automatic docking dynamic path planning system 100, comprising a processor 1001 and a memory 1003. The processor 1001 and memory 1003 are connected, for example, via a bus 1002. Optionally, the deep learning-based automatic docking dynamic path planning system 100 may further include a transceiver 1004, which can be used for data interaction between this deep learning-based automatic docking dynamic path planning system and other deep learning-based automatic docking dynamic path planning systems, such as sending and / or receiving data. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of this deep learning-based automatic docking dynamic path planning system 100 does not constitute a limitation on the embodiments of this application.

[0173] The memory 1003 is used to store program code for executing the embodiments of this application, and its execution is controlled by the processor 1001. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0174] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application, without departing from the technical concept of this application, also fall within the protection scope of the embodiments of this application.

Claims

1. A dynamic path planning method for automatic docking based on deep learning, characterized in that, The method includes: Acquire a set of real-time navigation perception data of the target vessel during berthing operations. The set of real-time navigation perception data includes the shoreline echo point cloud data collected by the shipborne sensing equipment, the instantaneous flow field vector distribution data of the waters surrounding the target vessel, and the real-time six-degree-of-freedom motion parameters of the target vessel's hull attitude. The shoreline echo point cloud data is processed by shoreline geometry analysis to generate continuous shoreline contour information and spatial location reference identifier of the target berthing berth. At the same time, the dynamic feasible domain constraint boundary of the target vessel in the berthing water is constructed based on the instantaneous flow field vector distribution data and the real-time six-degree-of-freedom motion parameters. The pre-built berth approach path generation network is invoked to perform multi-constraint path planning processing on the continuous contour morphology information of the wharf shoreline, the spatial location reference identifier and the dynamic feasible domain constraint boundary, to generate an initial berth approach path for the target vessel to approach the target berth from its current position. The initial berth approach path includes a continuously distributed sequence of path nodes and a sequence of target vessel reference attitude angles at each path node. The initial berth orientation path, the continuous contour morphology information of the wharf shoreline, the instantaneous flow field vector distribution data, and the real-time six-degree-of-freedom motion parameters are input into a pre-constructed path dynamic fine-tuning network for path smoothing and attitude optimization, generating automatic berthing dynamic path planning results. The automatic berthing dynamic path planning results include a corrected path node sequence and a path node corrected attitude angle sequence. A set of berthing control commands is generated based on the automatic berthing dynamic path planning results. The set of berthing control commands is used to drive the propulsion system and steering gear system of the target ship to perform berthing movements along the automatic berthing dynamic path planning results.

2. The deep learning-based automatic docking dynamic path planning method according to claim 1, characterized in that, The process involves analyzing the shoreline geometric shape of the wharf shoreline echo point cloud data to generate continuous shoreline contour information and spatial location reference markers for the target berthing position. Simultaneously, based on the instantaneous flow field vector distribution data and the real-time six-degree-of-freedom motion parameters, a dynamic feasible domain constraint boundary for the target vessel in the berthing waters is constructed, including: The point cloud echo data of the wharf shoreline is subjected to point cloud voxel downsampling processing to generate a sparse shoreline point cloud set. The sparse shoreline point cloud set retains the spatial point position information that represents the shoreline structure skeleton in the wharf shoreline echo point cloud data. The sparse shoreline point cloud set is subjected to local point neighborhood normal vector consistency analysis to generate a set of local geometric morphology parameters for the normal vector deflection degree and neighborhood surface undulation degree of each voxel unit. Based on the degree of neighborhood surface undulation in the set of local geometric morphology parameters, the sparse shoreline point cloud set is processed to extract shoreline feature boundary points, generating a set of shoreline feature boundary points. Each shoreline feature boundary point in the set of shoreline feature boundary points carries spatial coordinate information and local normal vector pointing information. The set of shoreline feature boundary points is input into the gated recurrent unit of the pre-constructed shoreline contour growth network to perform point-by-point state update processing along the spatial distribution of the set of shoreline feature boundary points, thereby generating the hidden state vector of each shoreline feature boundary point. The neighbor attention convergence layer of the shoreline contour growth network performs feature convergence processing on the hidden state vector of each shoreline feature boundary point and the hidden state vector of its neighboring shoreline feature boundary points in its local neighborhood, generating an enhanced hidden state vector containing neighborhood context association information. The enhanced hidden state vector of each shoreline feature boundary point is classified by the boundary point classification output layer of the shoreline contour growth network. Each shoreline feature boundary point is marked as a shoreline continuous contour constituent point or a shoreline break discontinuity point, generating a set of shoreline continuous contour constituent points. The continuous shoreline contour points in the set of shoreline continuous contour points are connected in an orderly manner according to their spatial proximity to generate continuous shoreline contour morphology information of the wharf shoreline. The continuous shoreline contour morphology information of the wharf shoreline includes the sequence of shoreline contour points that constitute the wharf shoreline and the curvature of the connecting line segments between adjacent shoreline contour points. Obtain the pre-calibrated static spatial coordinates of the target berth, and perform spatial registration processing between the static spatial coordinates of the berth and the shoreline contour point sequence in the continuous contour morphology information of the wharf shoreline to generate a spatial position reference identifier for the target berth. The spatial position reference identifier for the target berth includes the coordinates of the berth center point and the direction vector of the berth leading edge line segment. Based on the instantaneous flow field vector distribution data, a flow field velocity vector field of the waters surrounding the target vessel is constructed; based on the target vessel's sway velocity, pitch velocity, and bow roll velocity in the real-time six-degree-of-freedom motion parameters, the instantaneous motion velocity vector of the target vessel is determined; the instantaneous motion velocity vector of the target vessel is vector-superimposed with the flow field velocity vector at the corresponding spatial position in the flow field velocity vector field to generate a comprehensive disturbance velocity field of the target vessel in the berthing waters; Based on the comprehensive disturbance velocity field and combined with the preset safety boundary rules, a dynamic feasible domain constraint boundary for the target vessel in the berthing water is generated. The dynamic feasible domain constraint boundary includes the potential field boundary surface around the target vessel that is prohibited from being crossed.

3. The deep learning-based automatic docking dynamic path planning method according to claim 1, characterized in that, The pre-built berth approach path generation network is invoked to perform multi-constraint path planning processing on the continuous contour morphology information of the wharf shoreline, the spatial location reference identifier, and the dynamic feasible domain constraint boundary, generating an initial berth approach path for the target vessel to approach the target berth from its current position. The initial berth approach path includes a continuously distributed sequence of path nodes and a sequence of reference attitude angles of the target vessel at each path node, including: The target vessel's current position coordinates, the target vessel's current heading angle, the berth center point coordinates and berth leading edge line segment direction vector in the spatial position reference identifier, the potential field boundary surface geometric parameters of the dynamic feasible domain constraint boundary, and the shoreline contour point sequence in the continuous contour morphology information of the wharf shoreline are uniformly encoded into a multi-channel state tensor with the same spatial reference. The multi-channel state tensor is input into the multi-layer spatial graph convolutional layer of the berth tendency path generation network for spatial feature extraction processing, generating a spatial fusion feature map that includes the ship's own state and the constraints of the aquatic environment. The spatial fusion feature map is modeled with temporal dependencies by the bidirectional long short-term memory network layer of the berth tendency path generation network to generate a time series feature vector along the berthing direction. The time series feature vector along the berthing direction contains the temporal context information of the current state for the selection of future path nodes. The time series feature vector is input into the attention decoding layer of the berth tendency path generation network. The multi-head self-attention mechanism of the attention decoding layer performs global feature interaction processing on the time series feature vector to generate the attention weight distribution of each path node to be generated. The attention weight distribution is used to perform weighted summation on the corresponding spatial positions of the spatial fusion feature map to generate a spatial coordinate prediction vector for each path node to be generated. The spatial coordinate prediction vector contains the predicted spatial coordinates of the path node. A path node sequence is constructed based on the spatial coordinate prediction vector. The spatial coordinate difference vector between adjacent path nodes in the path node sequence is analyzed and processed by the attitude angle prediction branch of the berth tendency path generation network to generate a target ship reference attitude angle sequence at each path node. The target ship reference attitude angle sequence includes the ship's predicted heading angle at each path node. The path node sequence and the target ship reference attitude angle sequence are processed by point-by-point distance calculation with the potential field boundary surface of the dynamic feasible domain constraint boundary. The path nodes that are less than the preset safety distance margin from the potential field boundary surface are reprojected to the safe area outside the potential field boundary surface to generate the corrected path node sequence. The corrected path node sequence is subjected to path node density homogenization processing. New path nodes are inserted between path node pairs where the spacing between adjacent path nodes exceeds the preset node spacing upper limit to generate a path node sequence with uniform density. The uniformly dense sequence of path nodes and the corresponding sequence of target vessel reference attitude angles are combined to form an initial berth approach path. The initial berth approach path includes a continuously distributed sequence of path nodes and a sequence of target vessel reference attitude angles at each path node.

4. The deep learning-based automatic docking dynamic path planning method according to claim 1, characterized in that, The process involves inputting the initial berth orientation path, the continuous contour morphology information of the wharf shoreline, the instantaneous flow field vector distribution data, and the real-time six-degree-of-freedom motion parameters into a pre-constructed path dynamic fine-tuning network for path smoothing and attitude optimization. This generates an automatic berthing dynamic path planning result containing a corrected path node sequence and a corrected path node attitude angle sequence, including: Obtain the path node sequence of the initial berth orientation path and the target ship reference attitude angle sequence at each path node, and construct a spatiotemporal sequence sample of the path nodes. The spatiotemporal sequence sample of the path nodes includes the spatial coordinates of each path node, the corresponding target ship reference attitude angle, and the path node number. The flow field velocity vector in the instantaneous flow field vector distribution data is spatially interpolated according to the spatial coordinates of the path nodes to generate an instantaneous flow velocity vector at each path node. The instantaneous flow velocity vector includes the flow velocity direction and the flow velocity magnitude. The instantaneous flow velocity vector and the real-time six-degree-of-freedom motion parameters are concatenated to generate a hydrodynamic coupling feature vector at each path node; The hydrodynamic coupling feature vector and the target ship reference attitude angle of the path node are input into the attitude angle gating fine-tuning unit of the path dynamic fine-tuning network. The fine-tuning offset of the target ship reference attitude angle is calculated through the gating mechanism of the attitude angle gating fine-tuning unit, and the initial corrected attitude angle of the path node is generated. The attitude angle of the path node is initially corrected and the spatial coordinates of the path node are transformed. Combined with the curvature of the connecting line segment between adjacent shoreline contour points in the continuous contour morphology information of the wharf shoreline, the minimum distance interval between the outer contour of the ship and the shoreline contour at the path node is calculated. The minimum distance interval is input into the lateral offset compensation layer of the path dynamic fine-tuning network to calculate the lateral offset correction amount of each path node in the direction perpendicular to the path tangent. Based on the lateral offset correction amount, the spatial coordinates of the path node are laterally translated to generate the preliminary correction result of the spatial coordinates of the path node. The preliminary correction results of the spatial coordinates of the path nodes and the preliminary correction attitude angles of the path nodes are input into the path curvature linkage optimization layer of the path dynamic fine-tuning network. The local curvature change features of the path node sequence are extracted through convolution operation. The preliminary correction results of the spatial coordinates of the path nodes are smoothed according to the local curvature change features to generate the secondary correction results of the spatial coordinates of the path nodes. The global path shape encoder of the path dynamic fine-tuning network performs global shape encoding on the secondary correction result of the spatial coordinates of the path nodes to generate global path shape latent variables, which contain the overall morphological features of the initial berth tendency path. The global path shape latent variable and the overall shape features of the shoreline contour point sequence in the continuous contour morphology information of the wharf shoreline are subjected to shape consistency constraint processing. The spatial coordinates of the path node spatial coordinates are fine-tuned under global shape constraints to generate a path node correction sequence. The path node correction sequence and the path node preliminary correction attitude angle are input into the attitude angle fine-tuning layer of the path dynamic fine-tuning network. The attitude angle fine-tuning layer recalculates the path tangent direction based on the coordinate difference vector of adjacent path nodes in the path node correction sequence, and performs fine-tuning processing on the path node preliminary correction attitude angle according to the path tangent direction and the hydrodynamic coupling feature vector to generate the path node correction attitude angle sequence.

5. The deep learning-based automatic docking dynamic path planning method according to claim 4, characterized in that, The process of obtaining the path node sequence of the initial berth orientation path and the target vessel reference attitude angle sequence at each path node, and constructing a spatiotemporal sequence sample of the path nodes, includes: The data structure of the initial berth orientation path is analyzed, and the sequence of continuously distributed path nodes in the initial berth orientation path is extracted. Each path node in the path node sequence is assigned a unique path node number according to its arrangement order in the sequence. Extract the target ship reference attitude angle from the target ship reference attitude angle sequence corresponding to each path node. The target ship reference attitude angle is the ship heading angle value predicted synchronously by the berth tendency path generation network when generating the path node. Each path node's spatial coordinates, path node number, and target ship reference attitude angle are combined to form a basic spatiotemporal data unit. The spatial coordinates of the basic spatiotemporal data unit include the horizontal coordinate component, vertical coordinate component, and vertical height coordinate component in a three-dimensional spatial coordinate system. The timestamp information of the real-time six-degree-of-freedom motion parameters in the real-time navigation perception data is obtained, and a time alignment mapping relationship is established based on the timestamp information and the path node number of the path node. The timestamp information is then appended to the corresponding basic spatiotemporal data unit. Spatial distance calculation is performed between adjacent basic spatiotemporal data units to obtain the Euclidean spatial distance between adjacent path nodes. The Euclidean spatial distance is then added as a spatial relationship feature between path nodes to the attribute field of the previous basic spatiotemporal data unit. The time interval between adjacent basic spatiotemporal data units is calculated to obtain the time step difference between adjacent path nodes. The time step difference is then added to the attribute field of the previous basic spatiotemporal data unit as a feature of the temporal relationship between path nodes. The basic spatiotemporal data units carrying path node numbers, target ship reference attitude angles, timestamp information, spatial relationship characteristics between path nodes, and temporal relationship characteristics between path nodes are arranged in ascending order according to the path node numbers to generate spatiotemporal sequence samples of path nodes. The spatiotemporal sequence samples of path nodes are used as the input data stream of the path dynamic fine-tuning network.

6. The deep learning-based automatic docking dynamic path planning method according to claim 1, characterized in that, The method further includes: Acquire a set of historical navigation perception data recorded during historical berthing operations. The set of historical navigation perception data includes historical wharf shoreline echo point cloud data, historical instantaneous flow field vector distribution data, and historical real-time six-degree-of-freedom motion parameters. A training sample dataset is constructed based on the historical navigation perception dataset, which includes historical berth orientation path annotation data and historical path dynamic fine-tuning annotation data. The berth approach path generation network and the path dynamic fine-tuning network are jointly and alternately trained using the training sample data set to generate the trained berth approach path generation network and the trained path dynamic fine-tuning network.

7. The deep learning-based automatic docking dynamic path planning method according to claim 6, characterized in that, The step of constructing a training sample data set based on the historical navigation perception data set, wherein the training sample data set includes historical berth tendency path annotation data and historical path dynamic fine-tuning annotation data, including: Historical wharf shoreline echo point cloud data is extracted from the historical navigation perception data set. Historical shoreline geometric morphology analysis processing is performed on the historical wharf shoreline echo point cloud data to generate continuous contour morphology information of historical wharf shoreline and spatial location reference identifier of historical berthing target berth. Historical instantaneous flow field vector distribution data and historical real-time six-degree-of-freedom motion parameters are extracted from the historical navigation perception data set, and historical dynamic feasible domain constraint boundaries are constructed based on the historical instantaneous flow field vector distribution data and the historical real-time six-degree-of-freedom motion parameters; Obtain the ideal berthing path trajectory manually recorded during historical berthing operations, and decompose the ideal berthing path trajectory into a historical path node sequence and a historical path node attitude angle sequence, as historical berth trend path annotation data; For the historical path node sequence in the historical berth trend path annotation data, historical hydrodynamic disturbance data is generated based on the historical instantaneous flow field vector distribution data and the historical real-time six-degree-of-freedom motion parameters. The historical path node offset under the action of the historical hydrodynamic disturbance data is used as historical path offset annotation data. The historical path offset annotation data is combined with the historical path node sequence and the historical path node attitude angle sequence to generate historical path dynamic fine-tuning annotation data, which includes the spatial coordinate correction and attitude angle correction of the historical path nodes. The historical continuous contour morphology information of the wharf shoreline, the spatial location benchmark of the historical berthing target berth, the historical dynamic feasible domain constraint boundary, and the historical berth trend path annotation data are combined into berth trend path generation network training samples. The historical path dynamic fine-tuning annotation data, along with the corresponding historical continuous contour morphology information of the wharf shoreline, historical instantaneous flow field vector distribution data, and historical real-time six-degree-of-freedom motion parameters, are combined to form the training samples for the path dynamic fine-tuning network.

8. The deep learning-based automatic docking dynamic path planning method according to claim 6, characterized in that, The step of jointly and alternately training the berth approach path generation network and the path dynamic fine-tuning network using the training sample data set to generate the trained berth approach path generation network and the trained path dynamic fine-tuning network includes: The historical continuous contour morphology information of the wharf shoreline, the spatial location benchmark of the historical berthing target berth, and the historical dynamic feasible domain constraint boundary in the training samples of the berth tendency path generation network are input into the berth tendency path generation network to generate a predicted berth tendency path. Calculate the spatial coordinate deviation of the path nodes between the predicted berth trend path and the historical path node sequence in the historical berth trend path annotation data, and calculate the attitude angle deviation between the attitude angle of the predicted path nodes in the predicted berth trend path and the attitude angle sequence of the historical path nodes. The loss function of the berth approach path generation network is constructed based on the spatial coordinate deviation of the path nodes and the attitude angle deviation, and the network weight parameters of the berth approach path generation network are updated through the backpropagation algorithm. Freeze the network weight parameters of the berth tendency path generation network in the current training round, and use the predicted berth tendency path output by the berth tendency path generation network as the initial path input in the training samples of the path dynamic fine-tuning network. The historical instantaneous flow field vector distribution data and historical real-time six-degree-of-freedom motion parameters in the training samples of the path dynamic fine-tuning network are input into the path dynamic fine-tuning network. The initial path input is then processed for path smoothing and attitude optimization to generate a predicted dynamic fine-tuning path. Calculate the deviation between the spatial coordinate correction and attitude angle correction of the historical path nodes in the predicted dynamic fine-tuning path and the historical path dynamic fine-tuning annotation data, construct the path dynamic fine-tuning network loss function, and update the network weight parameters of the path dynamic fine-tuning network through the backpropagation algorithm. Repeatedly update the network weight parameters of the berth-oriented path generation network and the network weight parameters of the path dynamic fine-tuning network until the loss function of the berth-oriented path generation network and the loss function of the path dynamic fine-tuning network converge to the preset training termination condition, thereby generating the trained berth-oriented path generation network and the trained path dynamic fine-tuning network.

9. The deep learning-based automatic docking dynamic path planning method according to claim 1, characterized in that, The method further includes: The berthing motion feedback sensing data set is obtained in real time by the shipborne sensing device during the berthing motion of the target vessel. The berthing motion feedback sensing data set includes real-time updated quay shoreline echo point cloud data, real-time updated instantaneous flow field vector distribution data, and real-time updated real-time six-degree-of-freedom motion parameters. The automatic berthing dynamic path planning result is subjected to online rolling correction processing based on the berthing motion feedback perception data set, and an online corrected automatic berthing dynamic path planning result is generated.

10. A dynamic path planning system for automatic docking based on deep learning, characterized in that, The method includes a processor and a computer-readable storage medium storing machine-executable instructions that, when executed by the processor, implement the deep learning-based automatic docking dynamic path planning method according to any one of claims 1-9.

Images