Polar ship ice route path planning method

Abstract

The application discloses a polar ship ice area route path planning method and relates to the technical field of ship and ocean engineering. The application provides a path planning system and method suitable for ship ice area navigation. A neural network is used to replace a ship motion equation, so that the ship motion constraint can be conveniently processed in an optimal control method. An optimized path meeting the kinematic constraint of the ship is generated based on an environment map and the ship motion performance. The sea ice environment field construction based on the shipborne vision can be realized. The efficient processing of the complex kinematic constraint can be realized by using the neural network agent model. The optimized path strictly meeting the ship maneuvering performance can be generated. The problems that the existing method is difficult to simultaneously process the high-precision environment field construction and the complex kinematic constraint due to the mutual disconnection of the perception, modeling and planning links are solved.

Description

Technical Field

[0001] This invention relates to the field of shipbuilding and marine engineering technology, and in particular to a method for planning routes and paths for polar vessels in ice-covered areas. Background Technology

[0002] Sea ice is a major environmental threat to polar vessels. Collisions with sea ice can cause ships to slow down and, in severe cases, damage their structure. Ships navigating in ice-covered areas strive to avoid collisions with ice floes. Route planning based on reliable sea ice information is a fundamental prerequisite for safe navigation in ice-covered regions.

[0003] The known literature includes: Patent No. CN202510691512.5: A high ice condition avoidance path planning method considering the characteristics of ship movement in ice areas; Patent No. CN202411381190.6: Intelligent Path Planning and Control Method for Ships Navigating in Ice Zones; Patent No. CN202311173930.2: A method for sea ice identification and navigation path generation in ice-covered areas based on improved YOLOv5; Patent No. CN202310448494.9: A method for ship route planning in Arctic ice areas based on an improved gray wolf algorithm; Patent No. CN202211562357.X: A method, system and storage medium for ship path planning in ice-covered areas based on optimal control.

[0004] Existing technologies primarily focus on a single aspect of path planning, failing to establish a complete technological system integrating ship motion characteristics, sea ice environment perception, and path planning algorithms. Specifically, patents CN202510691512.5 and CN202411381190.6 mainly focus on ship motion characteristics or navigation control, but fail to effectively integrate them with real sea ice scene modeling; patent CN202311173930.2, while involving sea ice recognition, primarily targets sea ice detection with its improved YOLOv5-based visual method, without fully considering the direct impact of the ship's own kinematic constraints on path generation; patents CN202310448494.9 and CN202211562357.X focus on path planning through optimization algorithms or optimal control theory, and their planning process typically relies on sea ice environment models based on satellite remote sensing information, failing to achieve collaborative mapping and planning of "sea ice-ship" based on real-time perception by shipborne sensors. The above methods all have limitations in adaptability, safety and optimality in complex ice environments because they fail to systematically handle the coupling relationship between ship motion, sea ice mapping and route planning simultaneously. To address the aforementioned technical deficiencies, a solution is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a path planning system and method suitable for ship navigation in ice-covered areas. It uses a neural network to replace the ship's equations of motion, facilitating the handling of ship motion constraints in optimal control methods. Based on environmental maps and ship motion performance, it generates optimized paths that satisfy ship kinematic constraints. This allows for sea ice environment mapping based on shipborne vision, efficient processing of complex kinematic constraints using a neural network surrogate model, and the generation of optimized paths that strictly meet ship maneuvering performance requirements. It solves the problem of existing methods where perception, modeling, and planning are fragmented, making it difficult to simultaneously handle high-precision environmental constraints and complex kinematic constraints.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a polar vessel ice zone route planning method, comprising the following steps: Step 1: Establish a sea ice planning map based on sea ice images captured by shipborne cameras: S101. Continuously acquire sea ice images in the forward direction using camera equipment deployed on the hull, and eliminate lens distortion in the sea ice images and restore the geometric scale of the scene through image correction algorithms. S102. The corrected image is processed using a semantic segmentation deep learning model to accurately identify the pixel regions of seawater and sea ice, and generate a two-dimensional grid sea ice map containing the location information of sea ice obstacles for path planning. Step 2: Use a neural network model to handle ship kinematic constraints: S201. Construct a neural network model with a multilayer perceptron structure. Its input is the ship's state vector and control vector at the current moment, and its output is the state vector at the next predicted moment. S202. Supervised training of the neural network model is carried out using ship maneuvering data generated based on the ship longitudinal motion mathematical model, so that it can fit the dynamic response of the ship with high accuracy, thus serving as an efficient kinematic constraint surrogate model in path optimization. Step 3: Optimize the ship's navigation path using a two-layer path planning model based on the optimal control method: S301, use the HybridA algorithm to search the two-dimensional grid sea ice map to generate an initial feasible path that satisfies the kinematic constraints; S302. Using the initial feasible path as the initial guess, construct a discrete-time optimal control problem, wherein the ship motion is described by the neural network model, and the constraints include ship motion constraints and sea ice convex polygon constraints generated by the convex constraint generation algorithm. S303. A nonlinear solver is used to numerically solve the discrete-time optimal control problem, and the optimized final route is obtained and output.

[0007] Furthermore, the specific process for generating a two-dimensional raster sea ice map is as follows: S1021. Obtain the preprocessed sea ice image and use a deep convolutional neural network as a semantic segmentation deep learning model. The semantic segmentation deep learning model preferably adopts an encoder-decoder architecture, including: Input layer: Used to receive sea ice images after correction and IPM transformation. Encoder: Used to extract deep features from sea ice images, capturing the texture, edges, and shape information of sea ice.

[0008] Decoder: Used to upsample deep feature maps to restore them to the original image resolution and perform pixel-level classification prediction; S1022. Classify and determine each pixel of the sea ice image using a semantic segmentation deep learning model. Set the category set as C. During the judgment process, the semantic segmentation deep learning model outputs a probability tensor with the same size as the input image. By determining the category of each pixel, generate a binary or multi-valued semantic mask, where the sea ice area is marked with a specific value and the seawater area is marked with another specific value. S1023. Establish a local two-dimensional raster map M centered on the ship. Specifically, set the resolution of the raster map to r and the map size to W×L. Each raster cell in the local two-dimensional raster map M corresponds to a rectangular area in physical space. S1024. Map the sea ice pixel region output by the semantic segmentation deep learning model to the local two-dimensional grid map M. The pixels identified as sea ice are transformed into the coordinate system of the local two-dimensional grid map M according to their position in the bird's-eye view image using the coordinate transformation formula. S1025. Count the number of sea ice pixels falling into each grid cell. If the proportion or number of sea ice pixels in a grid cell exceeds a preset threshold, the grid cell is determined to be in an obstacle-occupied state and assigned a value of 1. Otherwise, it is determined to be a free passage area and assigned a value of 0; S1026. Perform morphological dilation processing on the generated obstacle grid to generate a final two-dimensional grid sea ice map containing a safety buffer. The final generated two-dimensional grid sea ice map contains the precise location distribution of sea ice obstacles and distance information recovered through geometric scale.

[0009] Furthermore, the specific process of constructing a neural network model with a multilayer perceptron structure is as follows: S2011. Training samples are generated using the Monte Carlo random sampling method. The range of values ​​for ship control inputs is set to include propeller speed and rudder angle. At each sampling moment, the current state vector and control vector are randomly generated. The state vector includes longitudinal velocity, lateral velocity, yaw rate, heading angle, and position coordinates. S2012. Input the state vector and control vector into the physical mathematical model, perform one-step integration using the fourth-order Runge-Kutta method to obtain the predicted state at the next moment, integrate them to obtain the training dataset, and use min-max normalization to map all input and output data in the training dataset to the [0,1] interval. S2013. Construct a fully connected feedforward neural network to approximate the nonlinear dynamic characteristics of a ship, wherein the feedforward neural network includes: The number of neurons in the input layer corresponds to the sum of the dimensions of the current state vector and the control vector. Hidden layers: The network contains M hidden layers, each containing H neurons; The number of neurons in the output layer corresponds to the dimension of the state vector at the next time step. S2014. Stochastic gradient descent is used for backpropagation training. An initial learning rate is set, and a learning rate decay strategy is introduced to prevent oscillations near the extreme point in the later stages of training. S205. Divide the training dataset into a training set and a validation set in a 7:3 ratio. Iteratively update the weights on the training set and monitor the loss changes using the validation set. When the loss on the validation set no longer decreases, enable the early stopping mechanism to prevent overfitting and obtain the trained neural network model.

[0010] Furthermore, the specific formulas for generating ship maneuverability data based on the ship's longitudinal motion mathematical model are as follows:

[0011] In the formula, m is the hull mass, Izz is the hull moment of inertia, mx, my, and Jzz are the longitudinal additional mass, transverse additional mass, and additional moment of inertia of the hull, respectively, ΣX, ΣY, and ΣN are the longitudinal external load, transverse external load, and bow roll moment of the hull in the moving coordinate system, respectively, and the subscripts H, P, and R represent viscous hydrodynamic forces, propeller thrust, and rudder force, respectively.

[0012] Furthermore, the specific process for generating an initial feasible path that satisfies the kinematic constraints is as follows: S3011. On the generated two-dimensional grid sea ice map, the HybridA algorithm is used for initial path planning. The HybridA algorithm considers ship kinematic constraints, uses the maneuvering motion equations in the ship's longitudinal motion mathematical model for state expansion, and constructs a cost function for node search. The cost function is expressed as follows:

[0013] In the formula, the ship's state vector Let x and y be the ship's state vector, x and y be the ship's eastward and northward position coordinates relative to a point in the inertial coordinate system, and φ be the bow angle of the ship relative to true north. The control variable uk=[δk] represents the control input, i.e., the rudder angle. c1, c2, and c3 are penalty coefficients. H(xk) is the heuristic cost generated by Dijkstra's algorithm, taking into account obstacle constraints. Its main function is to guide the ship to travel in the direction with the smaller heuristic value. S3012. During the state expansion process, collision detection is performed in real time. A bounding box is created with the ship's center as the center, and sea ice pixels within the bounding box are searched. If sea ice exists, the distances dx and dy between the sea ice pixels and the ship's center are calculated. The specific calculation process is as follows:

[0014] In the formula, the distance vector from the ship's center to the pixel within the bounding box is u=[x0 y0]T, and the unit vector in the bow direction is v=[cosφ sinφ]T, where w is the ship's width, l is the ship's length, and r is the safety margin. S3013, if satisfied , If no collision occurs, then no collision will occur; otherwise, a collision will occur.

[0015] Furthermore, the specific process for obtaining the optimized final route is as follows: S3031. Using the initial feasible path as the initial guess, a convex polygon constraint region is constructed using a convex constraint generation algorithm. The non-convex obstacle constraints are transformed into a series of linear inequality constraints, as shown below:

[0016] In the formula, Pk* is the initial path point. Let the point be the tangent point between the circle centered at the initial path point and the sea ice pixel. To optimize path points, For safety margin; S3032. A neural network model is used to describe the ship's kinematic constraints, and the discrete optimal control problem is modeled as follows:

[0017] In the formula, , These are the lower and upper limits of the control variable, respectively. , These are the starting and ending state vectors, respectively; S3033. A nonlinear solver is used to numerically solve the above optimal control problem to obtain a smooth, safe final optimization path that satisfies all constraints and meets the ship's maneuverability requirements.

[0018] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This polar vessel ice zone route planning method uses a neural network to replace the ship's motion equations, which facilitates the handling of ship motion constraints in optimal control methods. Based on environmental maps and ship motion performance, it generates optimized paths that satisfy ship kinematic constraints. It can realize sea ice environment mapping based on shipborne vision, and can use neural network surrogate models to achieve efficient processing of complex kinematic constraints. It can generate optimized paths that strictly satisfy ship maneuvering performance, solving the problem that existing methods are disconnected in the perception, modeling and planning stages, making it difficult to handle high-precision environmental constraints and complex kinematic constraints at the same time. Attached Figure Description

[0019] Figure 1 A schematic diagram of the overall method flow of the present invention is shown; Figure 2 A schematic diagram of the image correction algorithm of the present invention is shown; Figure 3a This is a schematic diagram of the original sea ice image for the present invention; Figure 3b This is a schematic diagram of the original sea ice image after tilt correction according to the present invention. Figure 3c A schematic diagram showing the sea ice image after semantic segmentation processing according to the present invention. Figure 4 This is a structural diagram of the neural network proxy model of the present invention; Figure 5 This is a flowchart of the convex constraint generation algorithm of the present invention; Figure 6a This diagram illustrates the process of calculating the actual driving distance required to reach the destination from different positions using the convex constraint generation algorithm of this invention. Figure 6b This is a diagram illustrating the actual driving distance required to reach the destination from different positions, calculated by the convex constraint generation algorithm of this invention. Figure 7 This is a diagram illustrating the optimization process of the convex constraint generation algorithm of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0021] like Figure 1-7 As shown, the polar vessel ice zone route planning method includes the following steps: Step 1: Establish a sea ice planning map based on sea ice images captured by shipborne cameras: S101. Continuously acquire sea ice images in the forward direction using camera equipment deployed on the hull, and eliminate lens distortion in the sea ice images and restore the geometric scale of the scene through image correction algorithms. S102. The corrected image is processed using a semantic segmentation deep learning model to accurately identify the pixel regions of seawater and sea ice, and generate a two-dimensional grid sea ice map containing the location information of sea ice obstacles for path planning. The specific process for generating a two-dimensional raster sea ice map is as follows: S1021. Obtain the preprocessed sea ice image and use a deep convolutional neural network as the semantic segmentation deep learning model. The semantic segmentation deep learning model preferably adopts an encoder-decoder architecture, including: Input layer: Used to receive sea ice images after correction and IPM transformation. Encoder: Used to extract deep features from sea ice images, capturing the texture, edges, and shape information of sea ice.

[0022] Decoder: Used to upsample deep feature maps to restore them to the original image resolution and perform pixel-level classification prediction; S1022. Classify and determine each pixel of the sea ice image using a semantic segmentation deep learning model. Set the category set as C. During the judgment process, the semantic segmentation deep learning model outputs a probability tensor with the same size as the input image. By determining the category of each pixel, generate a binary or multi-valued semantic mask, where the sea ice area is marked with a specific value and the seawater area is marked with another specific value. S1023. Establish a local two-dimensional raster map M centered on the ship. Specifically, set the resolution of the raster map to r and the map size to W×L. Each raster cell in the local two-dimensional raster map M corresponds to a rectangular area in physical space. S1024. Map the sea ice pixel region output by the semantic segmentation deep learning model to the local two-dimensional grid map M. The pixels identified as sea ice are transformed into the coordinate system of the local two-dimensional grid map M according to their position in the bird's-eye view image using the coordinate transformation formula. S1025. Count the number of sea ice pixels falling into each grid cell. If the proportion or number of sea ice pixels in a grid cell exceeds a preset threshold, the grid cell is determined to be in an obstacle-occupied state and assigned a value of 1. Otherwise, it is determined to be a free passage area and assigned a value of 0; S1026. Perform morphological dilation on the generated obstacle grid to generate a final two-dimensional grid sea ice map containing a safety buffer. The final generated two-dimensional grid sea ice map contains the precise location distribution of sea ice obstacles and distance information recovered through geometric scaling.

[0023] Step 2: Use a neural network model to handle ship kinematic constraints: S201. Construct a neural network model with a multilayer perceptron structure. Its input is the ship's state vector and control vector at the current moment, and its output is the state vector at the next predicted moment. The specific process of constructing a neural network model with a multilayer perceptron structure is as follows: S2011. The Monte Carlo random sampling method is used to generate training samples. The range of values ​​for ship control input is set to include propeller speed and rudder angle. At each sampling moment, the current state vector and control vector are randomly generated. The state vector includes longitudinal velocity, lateral velocity, yaw rate, heading angle and position coordinates. S2012. Input the state vector and control vector into the physical mathematical model, perform one-step integration using the fourth-order Runge-Kutta method to obtain the predicted state at the next moment, integrate them to obtain the training dataset, and use min-max normalization to map all input and output data in the training dataset to the [0,1] interval. S2013. Construct a fully connected feedforward neural network to approximate the nonlinear dynamic characteristics of a ship. The feedforward neural network includes: The number of neurons in the input layer corresponds to the sum of the dimensions of the current state vector and the control vector. Hidden layers: The network contains M hidden layers, each containing H neurons; The number of neurons in the output layer corresponds to the dimension of the state vector at the next time step. S2014. Stochastic gradient descent is used for backpropagation training. An initial learning rate is set, and a learning rate decay strategy is introduced to prevent oscillations near the extreme point in the later stages of training. S205. Divide the training dataset into a training set and a validation set in a 7:3 ratio. Iteratively update the weights on the training set and monitor the loss changes using the validation set. When the loss on the validation set no longer decreases, enable the early stopping mechanism to prevent overfitting and obtain the trained neural network model.

[0024] S202. Supervised training of the neural network model is carried out using ship maneuvering data generated based on the ship longitudinal motion mathematical model, so that it can fit the dynamic response of the ship with high accuracy, thus serving as an efficient kinematic constraint surrogate model in path optimization. The specific formulas for generating ship maneuverability data based on the ship's longitudinal motion mathematical model are as follows:

[0025] In the formula, m is the hull mass, Izz is the hull moment of inertia, mx, my, and Jzz are the longitudinal additional mass, transverse additional mass, and additional moment of inertia of the hull, respectively, ΣX, ΣY, and ΣN are the longitudinal external load, transverse external load, and bow roll moment of the hull in the moving coordinate system, respectively, and the subscripts H, P, and R represent viscous hydrodynamic forces, propeller thrust, and rudder force, respectively.

[0026] Step 3: Optimize ship navigation path using a two-layer path planning model based on optimal control method: S301, use the HybridA algorithm to search in a two-dimensional grid sea ice map to generate an initial feasible path that satisfies kinematic constraints; The specific process for generating an initial feasible path that satisfies the kinematic constraints is as follows: S3011. On the generated two-dimensional raster sea ice map, the HybridA algorithm is used for initial path planning. The HybridA algorithm considers ship kinematic constraints, uses the maneuvering motion equations from the ship's longitudinal motion mathematical model for state expansion, and constructs a cost function for node search. The cost function is expressed as follows:

[0027] In the formula, the ship's state vector Let x and y be the ship's state vector, x and y be the ship's eastward and northward position coordinates relative to a point in the inertial coordinate system, and φ be the bow angle of the ship relative to true north. The control variable uk=[δk] represents the control input, i.e., the rudder angle. c1, c2, and c3 are penalty coefficients. H(xk) is the heuristic cost generated by Dijkstra's algorithm, taking into account obstacle constraints. Its main function is to guide the ship to travel in the direction with the smaller heuristic value. S3012. During the state expansion process, collision detection is performed in real time. A bounding box is created with the ship's center as the center, and sea ice pixels within the bounding box are searched. If sea ice exists, the distances dx and dy between the sea ice pixels and the ship's center are calculated. The specific calculation process is as follows:

[0028] In the formula, the distance vector from the ship's center to the pixel within the bounding box is u=[x0 y0]T, and the unit vector in the bow direction is v=[cosφ sinφ]T, where w is the ship's width, l is the ship's length, and r is the safety margin. S3013, if satisfied , If no collision occurs, then no collision will occur; otherwise, a collision will occur.

[0029] S302. Using the initial feasible path as the initial guess, construct a discrete-time optimal control problem, where the ship motion is described by a neural network model, and the constraints include ship motion constraints and sea ice convex polygon constraints generated by the convex constraint generation algorithm. S303. A nonlinear solver is used to numerically solve the discrete-time optimal control problem, and the optimized final route is obtained and output.

[0030] The specific process for obtaining the optimized final route is as follows: S3031. Using the initial feasible path as the initial guess, a convex polygon constraint region is constructed using a convex constraint generation algorithm. The non-convex obstacle constraints are transformed into a series of linear inequality constraints, as shown below:

[0031] In the formula, Pk* is the initial path point. Let the point be the tangent point between the circle centered at the initial path point and the sea ice pixel. To optimize path points, For safety margin; S3032. A neural network model is used to describe the ship's kinematic constraints, and the discrete optimal control problem is modeled as follows:

[0032] In the formula, , These are the lower and upper limits of the control variable, respectively. , These are the starting and ending state vectors, respectively. S3033. A nonlinear solver is used to numerically solve the above optimal control problem to obtain a smooth, safe final optimization path that satisfies all constraints and meets the ship's maneuverability requirements.

[0033] This invention provides a path planning system and method suitable for ship navigation in ice-covered areas. It uses a neural network to replace the ship's equation of motion, which facilitates the handling of ship motion constraints in the optimal control method. Based on the environmental map and ship motion performance, it generates an optimized path that meets the ship's kinematic constraints. It can realize sea ice environment mapping based on shipborne vision, and can use a neural network surrogate model to achieve efficient processing of complex kinematic constraints. It can generate an optimized path that strictly meets the ship's maneuvering performance. It solves the problem that existing methods are fragmented in the perception, modeling and planning stages, and are difficult to handle high-precision environmental constraints and complex kinematic constraints at the same time.

[0034] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0035] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for planning routes and paths for polar vessels in ice-covered areas, characterized in that, Includes the following steps: Step 1: Establish a sea ice planning map based on sea ice images captured by shipborne cameras: S101. Continuously acquire sea ice images in the forward direction using camera equipment deployed on the hull, and eliminate lens distortion in the sea ice images and restore the geometric scale of the scene through image correction algorithms. S102. The corrected image is processed using a semantic segmentation deep learning model to accurately identify the pixel regions of seawater and sea ice, and generate a two-dimensional grid sea ice map containing the location information of sea ice obstacles for path planning. Step 2: Use a neural network model to handle ship kinematic constraints: S201. Construct a neural network model with a multilayer perceptron structure. Its input is the ship's state vector and control vector at the current moment, and its output is the state vector at the next predicted moment. S202. Supervised training of the neural network model is performed using ship maneuvering data generated based on the mathematical model of ship maneuvering motion. Step 3: Optimize the ship's navigation path using a two-layer path planning model based on the optimal control method: S301, use the HybridA algorithm to search the two-dimensional grid sea ice map to generate an initial feasible path that satisfies the kinematic constraints; S302. Using the initial feasible path as the initial guess, construct a discrete-time optimal control problem, wherein the ship motion is described by the neural network model, and the constraints include ship motion constraints and sea ice convex polygon constraints generated by the convex constraint generation algorithm. S303. A nonlinear solver is used to numerically solve the discrete-time optimal control problem, and the optimized final route is obtained and output.

2. The polar vessel ice zone route planning method according to claim 1, characterized in that, The specific process for generating a two-dimensional raster sea ice map is as follows: S1021. Obtain the preprocessed sea ice image and use a deep convolutional neural network as a semantic segmentation deep learning model. The semantic segmentation deep learning model preferably adopts an encoder-decoder architecture, including: Input layer: Used to receive sea ice images after correction and IPM transformation. Encoder: Used to extract deep features from sea ice images, capturing the texture, edges, and shape information of sea ice. Decoder: Used to upsample deep feature maps to restore them to the original image resolution and perform pixel-level classification prediction; S1022. Classify and determine each pixel of the sea ice image using a semantic segmentation deep learning model. Set the category set as C. During the judgment process, the semantic segmentation deep learning model outputs a probability tensor with the same size as the input image. By determining the category of each pixel, generate a binary or multi-valued semantic mask, where the sea ice area is marked with a specific value and the seawater area is marked with another specific value. S1023. Establish a local two-dimensional raster map M centered on the ship. Specifically, set the resolution of the raster map to r and the map size to W×L. Each raster cell in the local two-dimensional raster map M corresponds to a rectangular area in physical space. S1024. Map the sea ice pixel region output by the semantic segmentation deep learning model to the local two-dimensional grid map M. The pixels identified as sea ice are transformed into the coordinate system of the local two-dimensional grid map M according to their position in the bird's-eye view image using the coordinate transformation formula. S1025. Count the number of sea ice pixels falling into each grid cell. If the proportion or number of sea ice pixels in a grid cell exceeds a preset threshold, the grid cell is determined to be in an obstacle-occupied state and assigned a value of 1. Otherwise, it is determined to be a free passage area and assigned a value of 0; S1026. Perform morphological dilation processing on the generated obstacle grid to generate a final two-dimensional grid sea ice map containing a safety buffer. The final generated two-dimensional grid sea ice map contains the precise location distribution of sea ice obstacles and distance information recovered through geometric scale.

3. The polar vessel ice zone route planning method according to claim 1, characterized in that, The specific process of constructing a neural network model with a multilayer perceptron structure is as follows: S2011. Training samples are generated using the Monte Carlo random sampling method. The range of values ​​for ship control inputs is set to include propeller speed and rudder angle. At each sampling moment, the current state vector and control vector are randomly generated. The state vector includes longitudinal velocity, lateral velocity, yaw rate, heading angle, and position coordinates. S2012. Input the state vector and control vector into the physical mathematical model, perform one-step integration using the fourth-order Runge-Kutta method to obtain the predicted state at the next moment, integrate them to obtain the training dataset, and use min-max normalization to map all input and output data in the training dataset to the interval [0, 1]. S2013. Construct a fully connected feedforward neural network to approximate the nonlinear dynamic characteristics of a ship, wherein the feedforward neural network includes: The number of neurons in the input layer corresponds to the sum of the dimensions of the current state vector and the control vector. Hidden layers: The network contains M hidden layers, each containing H neurons; The number of neurons in the output layer corresponds to the dimension of the state vector at the next time step. S2014. Stochastic gradient descent is used for backpropagation training. An initial learning rate is set, and a learning rate decay strategy is introduced to prevent oscillations near the extreme point in the later stages of training. S205. Divide the training dataset into a training set and a validation set in a 7:3 ratio. Iteratively update the weights on the training set and monitor the loss changes using the validation set. When the loss on the validation set no longer decreases, enable the early stopping mechanism to prevent overfitting and obtain the trained neural network model.

4. The polar vessel ice zone route planning method according to claim 1, characterized in that, The specific formulas for generating ship maneuverability data based on the ship's longitudinal motion mathematical model are as follows: In the formula, m is the hull mass, Izz is the hull moment of inertia, mx, my, and Jzz are the longitudinal additional mass, transverse additional mass, and additional moment of inertia of the hull, respectively, ΣX, ΣY, and ΣN are the longitudinal external load, transverse external load, and bow roll moment of the hull in the moving coordinate system, respectively, and the subscripts H, P, and R represent viscous hydrodynamic forces, propeller thrust, and rudder force, respectively.

5. The polar vessel ice zone route planning method according to claim 1, characterized in that, The specific process for generating an initial feasible path that satisfies the kinematic constraints is as follows: S3011. On the generated two-dimensional grid sea ice map, the Hybrid A algorithm is used for initial path planning. The Hybrid A algorithm considers ship kinematic constraints, uses the maneuvering motion equations from the ship's longitudinal motion mathematical model for state expansion, and constructs a cost function for node search. The cost function is expressed as follows: In the formula, the ship's state vector Let x and y be the ship's state vector, x and y be the ship's eastward and northward position coordinates relative to a point in the inertial coordinate system, and φ be the bow angle of the ship relative to true north. The control variable uk = [δk] represents the control input, i.e., the rudder angle. c1, c2, and c3 are penalty coefficients. H(xk) is the heuristic cost generated by Dijkstra's algorithm that takes obstacle constraints into account. S3012. During the state expansion process, collision detection is performed in real time. A bounding box is created with the ship's center as the center. Sea ice pixels within the bounding box are searched. If sea ice exists, the distances dx and dy between the sea ice pixels and the ship's center are calculated. The specific calculation process is as follows: In the formula, the distance vector from the ship's center to the pixel within the bounding box is u=[x0y0]T, and the unit vector in the bow direction is v = [cosφsinφ]T, where w is the ship's width, l is the ship's length, and r is the safety margin. S3013, if satisfied , If no collision occurs, then no collision will occur; otherwise, a collision will occur.

6. The polar vessel ice zone route planning method according to claim 1, characterized in that, The specific process for obtaining the optimized final route is as follows: S3031. Using the initial feasible path as the initial guess, a convex polygon constraint region is constructed using a convex constraint generation algorithm. The non-convex obstacle constraints are transformed into a series of linear inequality constraints, as shown below: In the formula, Pk* is the initial path point. Let the point be the tangent point between the circle centered at the initial path point and the sea ice pixel. To optimize path points, For safety margin; S3032. A neural network model is used to describe the ship's kinematic constraints, and the discrete optimal control problem is modeled as follows: In the formula, , These are the lower and upper limits of the control variable, respectively. #imgpt14# represents the starting and ending state vectors, respectively; S3033. A nonlinear solver is used to numerically solve the above optimal control problem to obtain a smooth, safe final optimization path that satisfies all constraints and meets the ship's maneuverability requirements.

Images