A ship navigation decision-making method based on pulse neural network and space-time graph convolution fusion

Abstract

The application relates to a ship navigation decision-making method based on pulse neural network and space-time graph convolution fusion, which comprises the following steps: acquiring multi-source environment perception data, obtaining a sea area gridding feature matrix through preprocessing; adopting an event-driven mechanism of a pulse neural network and a space-time graph convolution network respectively to extract low-noise time sequence features and local space-time features in parallel; performing double-attention global and local fusion processing on and to obtain fusion features, and generating a risk heat map matrix and an energy consumption spectrum matrix based on the fusion features; taking and as constraint conditions, adopting a whale optimization algorithm to perform Pareto trade-off on multiple targets of risk, energy consumption and path to obtain an optimal route. Compared with the prior art, the application has the advantages of improving the navigation safety, energy efficiency level and decision-making real-time performance of a ship in a dynamic marine environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent ship navigation decision-making, and in particular to a ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution. Background Technology

[0002] Intelligent ship navigation and decision-making technology dynamically optimizes routes and speeds by collecting real-time marine meteorological and ship status data. Compared with traditional route planning methods that rely on human experience, it has significant advantages such as strong adaptability, multi-objective collaborative optimization, and real-time closed-loop decision-making.

[0003] However, due to the nonlinear coupling effects of various marine environmental factors such as wind, waves, currents, and channel density during navigation, traditional time-series models struggle to model the spatial structure of the sea area, and spatiotemporal graph convolutional networks cannot capture long-distance environmental propagation effects, posing a significant challenge to the real-time and accurate decision-making of ships. Insufficient environmental perception and prediction capabilities can lead to ships being unable to identify and avoid high-risk areas in complex sea conditions, significantly increasing the risk of collisions and groundings. Furthermore, the lack of dynamic zoning guidance can cause ships to blindly enter high-energy-consuming areas, resulting in a substantial increase in fuel consumption and consequently affecting navigation safety and operational economy. In existing technologies, such as those applied for CN118334623A, traditional time-series models such as LSTM and GRU are typically used to collect marine meteorological and ship status time-series data. However, these models generally lack spatial correlation information, resulting in lag in identifying high-risk areas under complex sea conditions, easily overlooking risks in distant sea areas, and creating safety hazards in route planning. Alternatively, spatiotemporal graph convolutional networks can be used to construct a graph structure by gridding the sea area, while extracting the temporal and spatial features of local sea areas to complete the representation of the marine environment and route optimization. However, the receptive field of spatiotemporal graph convolution is limited to adjacent nodes in the graph, which cannot capture the propagation effect of distant marine environmental elements. It also lacks differentiated weight enhancement for local risks and global sea area features, and its adaptive capability for route decision-making is insufficient in the face of dynamic sea conditions. The comprehensive optimization effect that takes into account navigation risks, energy consumption, and path length is limited. Summary of the Invention

[0004] The purpose of this invention is to provide a ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution, which reduces the computational burden while improving the navigation safety and real-time decision-making of ships in dynamic marine environments.

[0005] The objective of this invention can be achieved through the following technical solutions: A ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution includes the following steps: S1. Acquire multi-source environmental sensing data and obtain a gridded feature matrix of the sea area through preprocessing. ; S2. Based on the aforementioned sea area gridded feature matrix, low-noise temporal features are extracted in parallel using both the event-driven mechanism of a spiking neural network and a spatiotemporal graph convolutional network. and local spatiotemporal features ; S3, to and Global and local fusion processing with dual attention is performed to obtain fused features. and based on Generate a risk heatmap matrix and energy consumption spectrum matrix ; S4. The risk heatmap matrix is... and energy consumption spectrum matrix As a constraint, the whale optimization algorithm is used to perform Pareto trade-offs on multiple objectives such as risk, energy consumption, and route to obtain the optimal route for the ships. .

[0006] Furthermore, the multi-source environmental sensing data includes wave height. Wind speed ,wind direction Cloud cover and channel vessel density .

[0007] Furthermore, the extraction of low-noise temporal features The steps include: Spiking neural networks utilize the event-driven mechanism of LIF neurons to analyze the gridded feature matrix of the sea area. Temporal coding and noise suppression are performed to obtain low-noise temporal features while preserving key marine environmental change information. .

[0008] Furthermore, the local spatiotemporal features The extraction steps include: The sea area is gridded into a feature matrix. and the marine grid adjacency matrix characterizing the environmental features of each grid node. Input into the spatiotemporal graph convolutional network, for all adjacent grids We perform weighted aggregation to obtain the neighborhood aggregation feature, which is represented as: In the formula, This is a neighborhood aggregation feature. For a degree matrix, its diagonal elements satisfy , for The Middle Line 1 Column elements, The spatial convolution weight matrix is ​​a learnable matrix. Based on the neighborhood aggregation feature and the gridded feature matrix of the sea area Local spatiotemporal features are obtained through temporal convolution and residual fusion processing. , is represented as: In the formula, For refined features that include local spatiotemporal dependencies, For the match Features after dimensional alignment and normalization. This is a one-dimensional convolution process.

[0009] Furthermore, the obtained fusion features The steps include: Will The grids are clustered according to the coupling characteristics of environmental elements. Clusters with similar functions; Fine-grained multi-head attention mechanism is used for each cluster Fine-grained modeling of environmental interactions between homogeneous meshes yields characterization clusters. The interaction strength of environmental influences between homogeneous meshes is expressed as: In the formula, For clusters Inner grid For the grid The environmental interaction influence weight, i.e., the interaction strength, For clusters Inner grid The query vector, For clusters Inner grid The key vector, , For clusters Inner grid and grid The input features are and , , The linear projection weight matrix is ​​learnable. For attention head dimension; The interaction intensity AND value vector Multiply and sum the results to obtain the single-head attention output. Then, concatenate and aggregate the attention outputs of all heads to obtain the multi-head attention output. The multi-head attention output and input features are compared. After performing residual connections, layer normalization is performed, and then the data is input into the feedforward network and layer normalization is performed again to obtain single-grid updated features; cluster The single-grid update features of all grids within the grid are combined sequentially to obtain the updated spatiotemporal feature sequence. ; Each cluster The updated spatiotemporal feature sequence within Perform mean pooling along the time dimension to obtain each cluster. compact global representation vector ; Based on each cluster compact global representation vector Calculate cross-cluster global similarity weights And the long-distance propagation effects are weighted and superimposed using the residual structure. Above, we obtain each cluster Fusion characteristics , is represented as: In the formula, , , The global projection weight matrix, For clusters A compact global representation vector, For the feature dimensions of the global attention head, The scaling factor for dot product attention is softmax at the cluster index. Normalization is performed on the above.

[0010] Furthermore, the generated risk heatmap matrix and energy consumption spectrum matrix The steps include: Based on the fusion features The energy consumption cost and risk score of each grid are calculated, and the sea area is divided into four functional zones—high-risk high-energy consumption, high-risk low-energy consumption, low-risk high-energy consumption, and low-risk low-energy consumption—using the median of the entire sea area as a dynamic threshold. A risk heatmap matrix is ​​then output. and energy consumption spectrum matrix .

[0011] Furthermore, the optimal route of the aforementioned ship route The steps to obtain it include: The feasible area after excluding high-risk, high-energy-consuming no-fly zones is denoted as... ; In the feasible region Within, using the aforementioned risk heatmap matrix and energy consumption spectrum matrix As constraints, and with the objectives of minimizing risk, energy consumption, and path length, the following optimization problem is constructed: In the formula, For multi-objective optimization vectors, Candidate route sequence, Risk heatmap matrix Weighted integral along the route, Energy consumption spectrum matrix Weighted integral along the route, The length of each flight segment is accumulated. The risk minimization, energy consumption minimization, and path length minimization are treated as three parallel subtasks, and these three parallel subtasks are respectively treated as three subpopulations. , , The optimization problem is solved using the whale optimization algorithm, and the optimal route for the ships is output. .

[0012] Furthermore, the specific solution steps of the whale optimization algorithm include: (1) Initialize three subpopulations , , initial position Set the maximum number of iterations to ; (2) Considering multi-ship formation coordination and hard constraints, calculate the fitness value of individuals in each subpopulation, where the fitness value calculation expression is: In the formula, For individuals fitness value, For ship type adaptive weights, This is the formation deviation penalty coefficient. To constrain the penalty factors for violations, For the first The objective function value, This refers to the deviation between the actual and ideal distances between ships. For hard constraint violations; (3) Extract the inflection point location and constraint boundary information of Pareto elite individuals from each subpopulation, use it as transferable knowledge, and use the maximum mean difference to measure the distribution similarity of the optimized landscape of each subpopulation, so as to control the direction of knowledge transfer with dynamic transfer probability: In the formula, For dynamic transition probabilities, The fitness improvement rate of the subpopulation before and after iteration. The fitness improvement rate after introducing external elite knowledge, where when > hour Approaching 0, close the migration channel to avoid negative migration. > hour Approaching 1, absorbing elite knowledge to accelerate convergence; (4) Based on the dynamic migration probability Decide whether to introduce external elite knowledge and combine it with the current best individuals. This allows the three subpopulations to alternately execute the encirclement mechanism and the bubble net mechanism to balance exploration and exploitation, wherein the encirclement mechanism is as follows: The bubble net mechanism is as follows: In the formula, It is a uniform random vector. This represents the Euclidean distance between the current individual and the reference target during the prey encirclement phase. The linear convergence coefficients are: For the first The current position of an individual in a round of iteration. This represents the Euclidean distance between the current individual and the reference target during the spiral attack phase. The constant is the helical constant. These are the parameters for the spiral path; (5) Update iteration rounds The fitness values ​​are recalculated, and the Pareto elite individuals of each subpopulation are updated through non-dominated sorting. The global Pareto optimal solution set and frontier are then merged and maintained. (6) Determine the current iteration round Has the maximum number of iterations been reached? If not, return to step (3) to continue iterating; if yes, output the final Pareto optimal solution set and frontier. (7) Based on the final Pareto optimal solution set and frontier, select the optimal route that meets the current mission requirements. .

[0013] Furthermore, the screening step in step (7) includes: Map the mission requirements of different ship types to weight vectors By employing a ship type adaptive weighting mechanism, the selection of compromise solutions on the frontier is transformed into a weighted objective minimization problem. Combined with the aforementioned constraints, the optimal route is dynamically selected from the final Pareto optimal solution set. : In the formula, For the final Pareto optimal solution set, , , Weight vector The three components satisfy .

[0014] Furthermore, this includes a closed-loop feedback mechanism, specifically the following: exist At all times, obtain the updated actual status of the ship. and multi-source environmental sensing data and to Preprocessing is performed to obtain Sea area gridded feature matrix at time ,in The ship's position, speed, course, remaining fuel, and formation spacing deviation are used to update the next decision cycle, and serve as the initial state and constraints for route optimization in the next decision cycle. based on and Perform steps S2-S4 to obtain the result. The optimal route at any given time; For the next moment, repeat the above steps to complete the closed-loop feedback process until the ship reaches its destination.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention achieves low-power real-time environmental perception through the event-driven mechanism of spiking neural network, captures local and global spatiotemporal features of the sea area through spatiotemporal graph convolution and dual attention mechanism, and uses the whale optimization algorithm in the gridded sea area to perform multi-objective joint optimization on risk, energy consumption and path to solve the Pareto optimal route. Under the premise of ensuring navigation safety, it does not increase the onboard computing burden, and effectively controls navigation risk and energy consumption deviation, thereby improving the navigation safety and decision-making real-time performance of ships in dynamic marine environments.

[0016] (2) This invention eliminates the decision lag caused by the accumulation of prediction bias in traditional methods by using two methods: event-driven perception and dual attention prediction. At the same time, it combines the low-power perception of spiking neural networks with the prediction of spatiotemporal graph convolution. During prediction, it captures local and long-distance features of the sea area through fine-grained intra-cluster and global attention across clusters, providing high-precision constraint input for multi-objective optimization, so that the output meets the route of multiple indicators and avoids the comprehensive performance defects caused by a single objective.

[0017] (3) This invention adaptively adjusts the risk and energy consumption target weights for different ship types; and introduces periodic closed-loop feedback to continuously receive newly perceived marine environment data during navigation and dynamically update and optimize it, so that the method can adjust the route in real time with the changes in the marine environment, improve the dynamic adaptability of decision-making, and ultimately improve the overall safety of ship navigation, energy efficiency and real-time decision-making.

[0018] (4) The present invention can improve the accuracy of environmental perception and the real-time nature of navigation decision-making of ships, and has good ship type adaptability and formation coordination capabilities. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the navigation decision-making method of the present invention; Figure 2 This is a block diagram of the marine spatiotemporal feature fusion and functional zone division system of the present invention; Figure 3 This is a flowchart illustrating the multi-objective route optimization and adaptive knowledge transfer based on the whale optimization algorithm of this invention. Figure 4 This is a schematic diagram of the ship adaptive decision-making and feedback control principle of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0021] This embodiment provides a ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution, combined with... Figure 1 The overall process diagram shown illustrates that this method achieves low-power real-time environmental perception through an event-driven mechanism of a spiking neural network. It captures local and global spatiotemporal features of the sea area through spatiotemporal graph convolution and a dual attention mechanism. In a gridded sea area, it performs multi-objective joint optimization of risk, energy consumption, and path, and introduces a pre-set periodic closed-loop feedback to achieve dynamic adaptive decision-making. This method can improve the navigation safety, energy efficiency, and real-time decision-making of ships in dynamic marine environments. Specifically, the method includes the following steps: Step 1: Establish a spiking neural network for low-power event-driven marine environmental perception; Figure 1 This is a flowchart illustrating the proposed intelligent ship adaptive route planning process based on spiking neural networks and spatiotemporal graph convolution. The method employs a closed-loop architecture encompassing perception, prediction, optimization, decision-making, and feedback. It uses a left-side sea area knowledge and experience base, a historical track and ship type database, and an environmental database as global prior support. Data is input from multi-source sensors and satellite remote sensing, and after standardized preprocessing, is split into a spiking neural network (SNN) temporal feature extraction module and a spatiotemporal graph convolution (ST-GCN) spatial feature extraction module. These two feature streams are then fused globally and locally through dual attention and input into a sea area dynamic risk prediction module to generate risk and energy consumption constraint matrices. Subsequently, the matrix enters a multi-objective optimization engine based on the whale optimization algorithm (WOA), combining a maximum mean difference (MMD) adaptive knowledge transfer mechanism and a ship type adaptive weight matching mechanism to solve for the optimal navigation scheme. Finally, the ship's navigation is driven by the ship control execution module, and the navigation status feedback module transmits real-time operational data back to the system front end, triggering the next round of decision iteration. The event-driven mechanism of the spiking neural network is activated only at points of significant sea state change, effectively reducing power consumption while maintaining perception accuracy. The dual attention mechanism overcomes the limitation of traditional methods that only capture local dependencies, accurately modeling long-distance propagation effects across sea areas. The adaptive knowledge transfer mechanism selectively shares elite knowledge among subtasks, accelerating convergence and suppressing negative transfer.

[0022] The multi-source environmental sensing module receives wave height data from shipborne sensors and satellite remote sensing. Wind speed ,wind direction Cloud cover and channel vessel density The data, after missing value interpolation and Min-Max normalization preprocessing, forms a standardized gridded feature matrix of the sea area. Standardized preprocessed marine gridded feature matrix The temporal feature extraction module of the spiking neural network and the spatial feature extraction module of the spatiotemporal graph convolution are respectively input. The temporal feature extraction module of the spiking neural network uses the event-driven mechanism of LIF (Leaky Integrate-and-Fire) neurons to extract spatial features. Temporal coding and noise suppression are performed to obtain low-noise temporal features while preserving key marine environmental change information. The spatiotemporal graph convolution spatial feature extraction module uses... Adjacency matrix of sea area grid As input, neighborhood grid features are aggregated through spatial graph convolution, and the spatiotemporal evolution of environmental elements is captured through temporal convolution and residual fusion to obtain spatiotemporal features. The dual-attention global and local fusion module receives data simultaneously. and Feature fusion is achieved through intra-cluster local attention and cross-cluster global attention to obtain fused features. Based on this, a risk heatmap matrix is ​​generated. and energy consumption spectrum matrix Using the median of risk and energy consumption values ​​across the entire sea area as a dynamic threshold, the sea area is divided in real time into four functional zones: high-risk high-energy consumption, high-risk low-energy consumption, low-risk high-energy consumption, and low-risk low-energy consumption. The multi-objective optimization layer uses a risk heatmap matrix. and energy consumption spectrum matrix As a constraint, the whale optimization algorithm is used as the core evolutionary engine. Risk minimization, energy consumption minimization, and path length minimization are treated as three parallel subtasks. Elite knowledge is selectively shared among the subtasks through an adaptive knowledge transfer mechanism. The Pareto optimal front is solved, and the optimal route is output after adaptive weight matching based on ship type. The multi-source environmental sensing module operates at a preset cycle. The system receives updated wave height, wind speed, wind direction, cloud cover, and ship density data. Combined with the ship's actual position and remaining fuel information fed back from the integrated navigation system, it generates a new round of standardized gridded feature matrix, triggering the next closed-loop iteration. The closed-loop feedback mechanism is the core of the intelligent ship navigation decision-making system's continuous adaptation. It is responsible for synchronizing the ship's actual state and environmental changes back to the sensing end after each round of decision-making, eliminating the cumulative effect of prediction bias in traditional open-loop methods.

[0023] Step 2: Construct a spatiotemporal graph convolution and dual attention mechanism to fuse spatiotemporal features of the sea area; A spatiotemporal graph convolutional and dual-attention prediction model for gridded marine environmental data is constructed. The prediction model includes a spatiotemporal graph convolutional (ST-GCN) spatial feature extraction module and a dual-attention global and local fusion module. The ST-GCN spatial feature extraction module extracts local spatiotemporal features of the marine grid nodes; the dual-attention global and local fusion module fuses fine-grained local features within functional clusters and long-range global features between functional clusters. The fused features are used to generate a risk heatmap matrix R and an energy consumption spectrum matrix C, which serve as constraint inputs for a downstream multi-task optimization module. Figure 2 This is a system block diagram for the fusion of spatiotemporal features and functional zoning of marine areas. The diagram fully presents the gridded features of the marine area after standardized preprocessing. The system employs an end-to-end processing flow, from parallel input of SNN and ST-GCN branches to the final generation of risk and energy consumption heatmaps and completion of marine functional zone delineation. The system uses the output of step one as the main input, while simultaneously introducing a gridded marine adjacency matrix and predefined K functional cluster partitioning information as structured prior constraints for the model. The entire feature extraction process adopts a three-layer progressive architecture of "spatial aggregation, temporal evolution, and global fusion," sequentially capturing local spatial coupling through spatial graph convolution, modeling short-term dynamic changes through temporal convolution and residual fusion, and integrating cross-regional global dependencies through a dual attention mechanism. Finally, the risk and energy consumption prediction module outputs a full-area risk heatmap matrix and energy consumption spectrum matrix, and the dynamic threshold partitioning module divides the marine area into four functional zones in real time, providing clear constraint boundaries for subsequent multi-objective route optimization.

[0024] (1) Prediction Model For gridded sea areas, the environmental state of each grid node is progressively influenced by three mechanisms: local spatial coupling propagation of environmental elements between adjacent grids, short-term dynamic evolution of environmental state along the time axis, and long-distance cross-regional global dependence between functional clusters of different sea areas. Spatial graph convolution, temporal convolution, and dual attention mechanisms are used to describe the spatiotemporal prediction model of the sea area, respectively. 1) Spatial graph convolution: grid nodes Environmental characteristics are represented in the spatial dimension by an adjacency matrix. The defined connectivity relation applies to all its adjacent grids. The features are weighted and aggregated to capture the local propagation effects of environmental elements such as wind field and waves between adjacent sea areas. This aggregation process is uniformly represented on the entire map using the symmetric normalized Laplacian operator. (1) in (2) The standardized preprocessed marine gridded feature matrix serves as the common input to both the SNN and ST-GCN parallel branches. The SNN branch outputs low-noise temporal features. , It is the identity matrix. For the sea area grid adjacency matrix, for The degree matrix, This is a symmetric normalized Laplacian operator used to eliminate scale bias in feature aggregation between high-degree nodes and low-degree nodes. The spatial convolution weight matrix is ​​a learnable matrix. 2) Temporal Convolution and Residual Fusion: The neighborhood aggregation features output by the spatial graph convolution are fed into a one-dimensional convolution along the time dimension. The spatial features of consecutive time steps are weighted and combined through the convolution kernel, thereby capturing the dynamic evolution trend of environmental elements within a short time window. (3) in, The convolutional kernel slides along the time axis, weighting and fusing the local spatial aggregated features of adjacent time steps. The refined features processed by temporal convolution are then combined with the original input of the ST-GCN branch after dimensional alignment by two-dimensional convolution. The elements are added together using residual joins, and the expression is as follows: (4) in, For refined features that include local spatiotemporal dependencies, For the original input of the ST-GCN branch Features after dimensional alignment and normalization. The residual structure, while preserving the original input details, alleviates the gradient vanishing problem in deep network training; 3) Dual Attention Feature Fusion: While ST-GCN can capture the local spatiotemporal dependencies of adjacent grids, it is limited by the fixed local state of the static adjacency matrix and cannot model cross-regional propagation effects over long distances. Therefore, firstly... The grids are clustered according to the coupling characteristics of environmental elements. Several functionally similar clusters are used, and within each cluster, a fine-grained multi-head attention mechanism is employed to model the environmental interactions between homogeneous meshes in detail, represented as follows: (5) in, Clusters Inner grid Query vectors and grids The key vector, Let the learnable linear projection weight matrix be denoted as the cluster. The spatiotemporal feature sequence of each grid node after fine-grained multi-head attention update is as follows: ,in For clusters The number of grids within, For feature dimensions. For the attention head dimension, divide by This is used to scale the variance of the dot product values ​​to 1 to prevent gradient saturation. ∈[0,1] quantifies the interaction intensity of environmental influences between grids within a homogeneous sea area. Wherein, The steps to obtain it are as follows: Interaction intensity AND value vector Multiply and sum the results to obtain the single-head attention output. Then, concatenate and aggregate the attention outputs of all heads to obtain the multi-head attention output. Finally, combine the multi-head attention output with the input features. After performing residual connections, layer normalization is applied. The data is then input into the feedforward network and normalized again to obtain single-grid updated features. Finally, the clusters are... The single-grid update features of all grids within the grid are combined sequentially to obtain .

[0025] Subsequently, mean pooling is performed on the refined features within each cluster along the time dimension to obtain the compact global representation vector of that cluster. Calculate cross-cluster global similarity weights Furthermore, the effects of long-distance propagation are weighted and superimposed onto the original local fine features through the residual structure: (6) in, , The global projection weight matrix, These are residual retention terms, used to ensure that all information about the fine local features within the original cluster is not lost. To contribute variables across clusters, the long-distance influence of other clusters is introduced through a weighted summation. This is especially relevant when there is no direct ocean dynamic coupling channel between two clusters. The contribution of unrelated clusters is automatically suppressed to zero, thus fundamentally avoiding the problem of noise propagating between arbitrary grids in the whole-ocean attention model. Based on fusion features The energy consumption cost and risk score of each grid are calculated, and the sea area is divided into four functional zones in real time using the median of the entire sea area as a dynamic threshold. A risk heat map matrix is ​​then output. and energy consumption spectrum matrix The fusion features calculated by equation (6) Right now Figure 1 The output of the dual attention global and local fusion module is used as the input for subsequent risk and energy consumption predictions.

[0026] Step 3: Solve multi-objective optimization problems based on the whale optimization algorithm, considering risk, energy consumption, and path. Figure 3This document presents a framework diagram for multi-objective route optimization and adaptive knowledge transfer based on the whale optimization algorithm. The diagram comprehensively illustrates the entire process of multi-objective route optimization centered on the improved whale optimization algorithm, as well as the collaborative operation mechanism of the adaptive knowledge transfer mechanism among sub-tasks. The engine first uses the risk heatmap matrix and energy consumption spectrum matrix output from the upstream prediction layer as constraints to construct a feasible region after excluding high-risk, high-energy-consumption no-navigation zones, transforming route planning into a joint optimization problem of risk, energy consumption, and path. The engine employs a subpopulation parallel evolutionary architecture, mapping the three optimization objectives to independent subpopulations. It balances global exploration and local development capabilities through alternating execution of encirclement and bubble net mechanisms. After each iteration, the similarity of the optimization landscape among sub-tasks is measured using the maximum mean difference (MMD), dynamically controlling the direction and intensity of elite knowledge transfer, effectively avoiding negative transfer problems while accelerating convergence. A comprehensive fitness function, including formation deviation penalties and constraint violation penalties, is designed to meet the requirements of multi-ship formation collaboration. After multiple iterations, the Pareto optimal solution set is continuously maintained and output, providing high-quality candidate solutions for subsequent ship type adaptive decision-making. The method employed uses the risk heatmap matrix R and the energy consumption spectrum matrix C as constraint boundaries, excluding high-risk, high-energy-consumption no-fly zones within the feasible region. The internal framework allows routes to make Pareto trade-offs among three objectives: risk, energy consumption, and route. The optimization problem can be represented as: (7) in Candidate route sequence, for The weighted integral of the matrix along the flight path. for The weighted integral of the matrix along the flight path. This is the sum of the flight segment lengths.

[0027] The whale optimization algorithm has a wide population distribution and multiple convergence directions. In the early stages of iteration, it needs to maintain global exploration to cover the feasible region, while in the later stages, it needs to gradually shrink to approach the Pareto front. If the step size and shrinkage rate are reasonably controlled and the search is guided by the current best individual, the exploration and development can be balanced by alternately executing encirclement and bubble net mechanisms. (8) (9) in The vector is a uniform random vector, and a decreases linearly from the initial value of 2 to 0; For Euclidean distance, The constant is the helical constant. ; Let T be defined as the current iteration round of the WOA algorithm, and the three subpopulations. (risk), (Energy consumption) (Path) Updates are performed using equations (8) and (9) respectively. After each iteration, the inflection points and constraint boundary information of Pareto elite individuals are extracted from each subpopulation as transferable knowledge. The maximum mean difference is used to measure the distribution similarity of the optimized landscapes among subtasks, and the direction of knowledge transfer is controlled by dynamic transfer probability. (10) in The fitness improvement rate of the subpopulation before and after iteration. The fitness improvement rate after introducing external knowledge. > hour Approaching 0, negative migration is avoided by closing migration channels; when > hour To approach 1, convergence is accelerated by incorporating elite knowledge. Since multi-ship formation coordination is considered, formation deviation penalty and constraint violation penalty are introduced into the fitness function: (11) in For ship type adaptive weights, This is the formation deviation penalty coefficient. This refers to the deviation between the actual and ideal distances between ships. To constrain the penalty factors for violations, The value represents the hard constraint violation. The Pareto optimal solution set is obtained by iteratively solving equation (11), and the frontier is maintained.

[0028] Step 4: Establish an adaptive weighting system for ship types and a closed-loop feedback mechanism to obtain the optimal route; Figure 4 This diagram illustrates the principle of ship adaptive decision-making and feedback control, clearly showing the complete control chain from the Pareto optimal solution set to the final route execution and closed-loop feedback. The ship adaptive decision-maker, based on a preset ship type weight vector, selects the optimal route from the Pareto solution set that meets the current mission requirements and sends it to the ship control system. During ship execution, real-time collected navigation status and updated environmental data are fed back to the system front-end through a feedback channel, triggering the next round of decision iteration, forming a continuously adaptive closed-loop control. The ship type adaptive weight mechanism is essentially a goal priority adjustment method that runs through the entire chain, mapping the mission requirements of different ship types to weight vectors, transforming the selection of compromise solutions on the Pareto front into a weighted goal minimization problem. For the weight vector, The optimal flight path for the current period Wire, This is the Pareto solution set output from step three.

[0029] (12) (13) The closed-loop feedback decision-making adopts a rolling optimization approach. At time t, the risk heatmap matrix output from step two is used. and energy consumption spectrum matrix As a constraint, the Pareto optimal solution set output in step three is used as the candidate set, and the optimal route for the current decision cycle is selected according to equation (13). And it is then issued to the ship's control system for execution. At any given moment, the system collects updated ship status and original environmental observation data, and re-executes environmental preprocessing, SNN temporal feature extraction, ST-GCN spatial feature extraction, dual attention feature fusion, and multi-objective route optimization, causing the decision window to continuously scroll forward over time. This closed-loop scrolling method decomposes the entire voyage optimization into multiple short-cycle local optimization sub-problems, thereby reducing the scale of a single optimization and meeting the real-time decision-making requirements of the shipborne edge platform. At any given time, the ship's status update is represented as follows: (14) in, The actual ship status collected for the integrated ship navigation system includes actual position, speed, heading, and remaining fuel. The ship's status is used to provide feedback for the next round of decision-making.

[0030] Meanwhile, the raw environmental observation data collected in the new cycle are represented as follows: (15) in, , , , and These represent wave height, wind speed, wind direction, cloud cover, and channel vessel density, respectively. Missing value interpolation and Min-Max normalization are performed on the updated original environmental observation data to obtain: (16) in This represents the interpolation process for missing values. This represents Min-Max normalization. This is the standardized environmental feature matrix for the new cycle. This matrix is ​​input into the SNN temporal feature extraction branch and the ST-GCN spatial feature extraction branch, respectively. After dual attention global and local fusion, the two feature paths are regenerated to generate the risk heatmap matrix and energy consumption spectrum matrix. (17) Where A is the sea area grid adjacency matrix, and F_predict represents the environmental prediction model composed of the SNN temporal feature extraction branch, the ST-GCN spatial feature extraction branch, and the dual attention global and local fusion module. The next round of optimization recalculates the fitness based on the updated ship status, risk heatmap matrix, and energy consumption spectrum matrix. (18) in For the updated ship formation spacing deviation, This is the formation deviation penalty coefficient. To constrain the penalty factors for violations, This refers to the amount of violation of hard constraints. The weighting principle for different ship types is as follows: ocean-going container ships High-speed rescue boat Oceanographic research vessel The weighting is applied across three layers: the perception layer adaptively filters the environmental sensitivity characteristics of different ship types, the prediction layer adjusts the partition threshold according to the seaworthiness level of the ship type, and the optimization layer determines the compromise direction on the Pareto front. By solving the closed-loop iterative solution of equations (12) to (18), the optimal route for each decision cycle is obtained by minimizing the weighted objective function and applied to the ship control system in sequence.

[0031] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0032] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0033] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0034] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0035] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0036] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0037] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution, characterized in that, Includes the following steps: S1. Acquire multi-source environmental sensing data and obtain a gridded feature matrix of the sea area through preprocessing. ; S2. Based on the aforementioned sea area gridded feature matrix, low-noise temporal features are extracted in parallel using both the event-driven mechanism of a spiking neural network and a spatiotemporal graph convolutional network. and local spatiotemporal features ; S3, to and Global and local fusion processing with dual attention is performed to obtain fused features. and based on Generate a risk heatmap matrix and energy consumption spectrum matrix ; S4. The risk heatmap matrix is... and energy consumption spectrum matrix As a constraint, the whale optimization algorithm is used to perform Pareto trade-offs on multiple objectives such as risk, energy consumption, and route to obtain the optimal route for the ships. .

2. The ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution as described in claim 1, characterized in that, The multi-source environmental sensing data includes wave height. Wind speed ,wind direction Cloud cover and channel vessel density .

3. The ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution as described in claim 1, characterized in that, The low-noise time-series features were extracted. The steps include: Spiking neural networks utilize the event-driven mechanism of LIF neurons to analyze the gridded feature matrix of the sea area. Temporal coding and noise suppression are performed to obtain low-noise temporal features while preserving key marine environmental change information. .

4. The ship navigation decision-making method based on the fusion of spiking neural network and spatiotemporal graph convolution as described in claim 1, characterized in that, The local spatiotemporal features The extraction steps include: The sea area is gridded into a feature matrix. and the marine grid adjacency matrix characterizing the environmental features of each grid node. Input into the spatiotemporal graph convolutional network, for all adjacent grids We perform weighted aggregation to obtain the neighborhood aggregation feature, which is represented as: In the formula, This is a neighborhood aggregation feature. For a degree matrix, its diagonal elements satisfy , for The Middle Line 1 Column elements, The spatial convolution weight matrix is ​​a learnable matrix. Based on the neighborhood aggregation feature and the gridded feature matrix of the sea area Local spatiotemporal features are obtained through temporal convolution and residual fusion processing. , is represented as: In the formula, For refined features that include local spatiotemporal dependencies, For the match Features after dimensional alignment and normalization. This is a one-dimensional convolution process.

5. The ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution as described in claim 1, characterized in that, The obtained fusion features The steps include: Will The grids are clustered according to the coupling characteristics of environmental elements. Clusters with similar functions; Fine-grained multi-head attention mechanism is used for each cluster Fine-grained modeling of environmental interactions between homogeneous meshes yields characterization clusters. The interaction strength of environmental influences between homogeneous meshes is expressed as: In the formula, For clusters Inner grid For the grid The environmental interaction influence weight, i.e., the interaction strength, For clusters Inner grid The query vector, For clusters Inner grid The key vector, , For clusters Inner grid and grid The input features are and , , The linear projection weight matrix is ​​learnable. For attention head dimension; The interaction intensity AND value vector Multiply and sum the results to obtain the single-head attention output. Then, concatenate and aggregate the attention outputs of all heads to obtain the multi-head attention output. The multi-head attention output and input features are compared. After performing residual connections, layer normalization is performed, and then the data is input into the feedforward network and layer normalization is performed again to obtain single-grid updated features; cluster The single-grid update features of all grids within the grid are combined sequentially to obtain the updated spatiotemporal feature sequence. ; Each cluster The updated spatiotemporal feature sequence within Perform mean pooling along the time dimension to obtain each cluster. compact global representation vector ; Based on each cluster compact global representation vector Calculate cross-cluster global similarity weights And the long-distance propagation effects are weighted and superimposed using the residual structure. Above, we obtain each cluster Fusion characteristics , is represented as: In the formula, , , The global projection weight matrix, For clusters A compact global representation vector, For the feature dimensions of the global attention head, The scaling factor for dot product attention is softmax at the cluster index. Normalization is performed on the above.

6. The ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution as described in claim 1, characterized in that, The generated risk heatmap matrix and energy consumption spectrum matrix The steps include: Based on the fusion features The energy consumption cost and risk score of each grid are calculated, and the sea area is divided into four functional zones—high-risk high-energy consumption, high-risk low-energy consumption, low-risk high-energy consumption, and low-risk low-energy consumption—using the median of the entire sea area as a dynamic threshold. A risk heatmap matrix is ​​then output. and energy consumption spectrum matrix .

7. The ship navigation decision-making method based on the fusion of spiking neural network and spatiotemporal graph convolution as described in claim 1, characterized in that, The optimal route of the ship's route The steps to obtain it include: The feasible area after excluding high-risk, high-energy-consuming no-fly zones is denoted as... ; In the feasible region Within, using the aforementioned risk heatmap matrix and energy consumption spectrum matrix As constraints, and with the objectives of minimizing risk, energy consumption, and path length, the following optimization problem is constructed: In the formula, For multi-objective optimization vectors, Candidate route sequence, Risk heatmap matrix Weighted integral along the route, Energy consumption spectrum matrix Weighted integral along the route, The length of each flight segment is accumulated. The risk minimization, energy consumption minimization, and path length minimization are treated as three parallel subtasks, and these three parallel subtasks are respectively treated as three subpopulations. , , The optimization problem is solved using the whale optimization algorithm, and the optimal route for the ships is output. .

8. The ship navigation decision-making method based on the fusion of spiking neural network and spatiotemporal graph convolution as described in claim 7, characterized in that, The specific solution steps of the whale optimization algorithm include: (1) Initialize three subpopulations , , initial position Set the maximum number of iterations to ; (2) Considering multi-ship formation coordination and hard constraints, calculate the fitness value of individuals in each subpopulation, where the fitness value calculation expression is: In the formula, For individuals fitness value, For ship type adaptive weights, This is the formation deviation penalty coefficient. To constrain the penalty factors for violations, For the first The objective function value, This refers to the deviation between the actual and ideal distances between ships. For hard constraint violations; (3) Extract the inflection point location and constraint boundary information of Pareto elite individuals from each subpopulation, use it as transferable knowledge, and use the maximum mean difference to measure the distribution similarity of the optimized landscape of each subpopulation, so as to control the direction of knowledge transfer with dynamic transfer probability: In the formula, For dynamic transition probabilities, The fitness improvement rate of the subpopulation before and after iteration. The fitness improvement rate after introducing external elite knowledge, where when > hour Approaching 0, close the migration channel to avoid negative migration. > hour Approaching 1, absorbing elite knowledge to accelerate convergence; (4) Based on the dynamic migration probability Decide whether to introduce external elite knowledge and combine it with the current best individuals. This allows the three subpopulations to alternately execute the encirclement mechanism and the bubble net mechanism to balance exploration and exploitation, wherein the encirclement mechanism is as follows: The bubble net mechanism is as follows: In the formula, It is a uniform random vector. This represents the Euclidean distance between the current individual and the reference target during the prey encirclement phase. The linear convergence coefficients are: For the first The current position of an individual in a round of iteration. This represents the Euclidean distance between the current individual and the reference target during the spiral attack phase. The constant is the helical constant. These are the parameters for the spiral path; (5) Update iteration rounds The fitness values ​​are recalculated, and the Pareto elite individuals of each subpopulation are updated through non-dominated sorting. The global Pareto optimal solution set and frontier are then merged and maintained. (6) Determine the current iteration round Has the maximum number of iterations been reached? If not, return to step (3) to continue iterating; if yes, output the final Pareto optimal solution set and frontier. (7) Based on the final Pareto optimal solution set and frontier, select the optimal route that meets the current mission requirements. .

9. A ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution as described in claim 8, characterized in that, The screening steps in step (7) include: Map the mission requirements of different ship types to weight vectors By employing a ship type adaptive weighting mechanism, the selection of compromise solutions on the frontier is transformed into a weighted objective minimization problem. Combined with the aforementioned constraints, the optimal route is dynamically selected from the final Pareto optimal solution set. : In the formula, For the final Pareto optimal solution set, , , Weight vector The three components satisfy .

10. A ship navigation decision-making method based on the fusion of spiking neural networks and spatiotemporal graph convolution as described in claim 1, characterized in that, This includes a closed-loop feedback mechanism, specifically the following: exist At all times, obtain the updated actual status of the ship. and multi-source environmental sensing data and to Preprocessing is performed to obtain Sea area gridded feature matrix at time ,in Used to update the ship's position, speed, course, remaining fuel, and formation spacing deviation for the next cycle, and as the initial state and constraints for route optimization in the next decision cycle; based on and Perform steps S2-S4 to obtain the result. The optimal route at any given time; For the next moment, repeat the above steps to complete the closed-loop feedback process until the ship reaches its destination.

Images