Ship overtaking collision warning method and system in dangerous areas

Abstract

The present invention relates to the field of navigation technology, and in particular to a method and system for warning a ship overtaking and collision in a dangerous area. The method comprises: collecting real-time status data of a ship, generating a dynamic risk field by multi-scale analysis, predicting trajectories by using a graph neural network in a non-Euclidean space, optimizing path data in combination with a multi-level Bayesian method, and using reinforcement learning for self-optimization of risk response. The present invention can significantly improve the collision avoidance capability and navigation safety of a ship in a complex marine environment through dynamic risk assessment and adaptive path optimization.

Description

Technical Field

[0001] The present invention relates to the field of navigation technology, and in particular to a method and system for early warning of ship overtaking and collision in dangerous areas. Background Art

[0002] With the rapid development of the shipping industry, maritime traffic is becoming increasingly busy, especially in ports, waterways or sea areas with dense ships, where the risk of collision between ships has increased significantly. An effective overtaking collision warning system can help ships detect potential collision risks in advance and provide appropriate avoidance plans, thereby reducing the probability of accidents and improving the safety of maritime traffic.

[0003] The existing technology (Chinese invention patent, CN117132012B, title: A multi-ship collision avoidance method, electronic device and storage medium for predicting conflict danger zone) usually adopts rule-based methods for ship collision avoidance, such as using preset international maritime collision avoidance regulations or rapidly expanding random tree (RRT) algorithm. These methods usually take the relative position and speed between ships as the main basis and use fixed rules to determine the avoidance strategy. However, such methods have the following defects:

[0004] It is unable to dynamically adapt to emergencies in complex maritime environments. The preset rules are too simple to cope with the complex situation of multi-ship collision avoidance. Path planning mainly focuses on minimizing local conflicts, which may result in the ship falling into new collision risks after resolving one risk, and lacks global optimization capabilities. It is difficult to make rapid adjustments and optimizations in real time in complex and changing maritime environments. Summary of the invention

[0005] In view of the many problems existing in the above-mentioned prior art, the present invention provides a method and system for warning of ship overtaking and collision in dangerous areas. The present invention collects real-time status data of ships, adopts multi-scale dynamic risk analysis and trajectory prediction technology in non-Euclidean space, combines multi-level Bayesian fusion method to perform risk assessment and path optimization, and uses reinforcement learning for continuous optimization. The present invention can accurately predict the future path changes of ships, dynamically adjust the avoidance strategy, and significantly improve the safety of maritime traffic.

[0006] A method for early warning of ship overtaking and collision in a dangerous area comprises the following steps:

[0007] Collect the real-time status data of the ship, including position, speed, acceleration, heading and environmental information, perform data preprocessing on the real-time status data of the ship to obtain fused sensor data; based on the fused sensor data, use multi-scale analysis to generate local scale risk field data and global scale risk field data, and form real-time updated multi-scale dynamic risk field data through dynamic fusion to reflect the current and predicted future risk distribution;

[0008] Based on multi-scale dynamic risk field data, a graph structure of non-Euclidean space is constructed, and the future trajectory of the ship is predicted using a graph neural network to generate initial non-Euclidean trajectory prediction data; bidirectional coupling correction is performed in combination with the ship's operational behavior data to obtain coupled trajectory prediction data; based on the coupled trajectory prediction data, risk transmission path analysis and adaptive path optimization are performed to generate adaptive optimization path data; the adaptive optimization path data is fused with the coupled trajectory prediction data through a multi-level Bayesian fusion method to generate the final Bayesian path data;

[0009] Based on the final Bayesian path data and non-Euclidean risk transmission data, the risk level is evaluated and classified; the execution feedback data of each risk response is recorded, and the feedback data is self-learned and continuously optimized through the reinforcement learning algorithm.

[0010] Preferably, the multi-scale analysis to generate local-scale risk field data and global-scale risk field data includes:

[0011] The local area of ​​the fused sensor data is scanned through a convolutional neural network to extract the local feature information of the distance change and speed change between ships, and generate local-scale risk field data. The historical time series of the fused sensor data is modeled using a long short-term memory network to identify the trend of ship status changes over time and generate global-scale risk field data.

[0012] Preferably, the multi-scale dynamic risk field data updated in real time through dynamic fusion includes:

[0013] The local-scale risk field data and the global-scale risk field data are converted into the risk intensity representation of the same unit, and multi-scale dynamic risk field data are generated through weighted fusion, where the calculation formula of risk intensity is:

[0014]

[0015] in, Represents the risk intensity of multi-scale dynamic risk field data; is the risk intensity of the local scale risk field data; is the risk intensity of the global-scale risk field data; and are the weight coefficients of the local scale and the global scale respectively, and satisfy , the weight coefficient is dynamically adjusted according to the speed and relative position of the ship.

[0016] Preferably, constructing a graph structure of a non-Euclidean space includes:

[0017] The ship status in the multi-scale dynamic risk field data is used as the attribute of the node, including the real-time position, speed and acceleration information of the ship; the edge weight value between the nodes is determined according to the relative distance and relative speed between the ships, and the edge weight value is calculated according to the following formula:

[0018]

[0019] in, Representation Node and nodes The edge weights between ; For Node and nodes The relative distance between It is an adjustment parameter used to control the nonlinear change degree of edge weight; is the bias parameter; it is used to adjust the threshold of edge weight.

[0020] Preferably, the bidirectional coupling correction includes:

[0021] The node positions and edge weights in the initial non-Euclidean trajectory prediction data are corrected according to the ship's operating behavior data. The operating behavior data include the heading change rate, acceleration change rate and turning speed. The coupled trajectory prediction data is generated by calculating the influence of the operating behavior data on the node positions and edge weights. The correction process adopts a two-way calculation, which not only considers the changes in the ship's own behavior, but also combines the interference of other ships' behaviors on the prediction results.

[0022] Preferably, the risk transmission path analysis includes:

[0023] Based on the node and edge weight information in the coupled trajectory prediction data, the risk propagation algorithm of the graph neural network is used to calculate the risk conduction intensity of each path node. The calculation formula of the risk conduction intensity is:

[0024]

[0025] in, Representation Node The intensity of risk transmission; For Node For Node The edge weight of For Node The risk intensity is calculated and the risk transmission path data is used to identify high-risk areas.

[0026] Preferably, the adaptive path optimization includes:

[0027] The variational adaptive gradient algorithm is used to iteratively adjust the steering angle and smoothness of each path point in the risk transmission path data to update the position of the path point. The optimized path generates adaptive optimized path data based on the local gradient information of the path point and the global risk distribution.

[0028] Preferably, the multi-level Bayesian fusion method includes:

[0029] Multiple Bayesian levels are set, and the fusion weight of each Bayesian level is set according to the specific risk priority and behavioral feature priority. The adaptive optimization path data and the coupled trajectory prediction data are weightedly fused layer by layer in turn to gradually form the final Bayesian path data.

[0030] Preferably, the reinforcement learning algorithm is used to self-learn and continuously optimize the feedback data including:

[0031] The risk response data is trained through the deep Q network algorithm. The risk response data includes safety indicators after execution, response time and path adjustment range. The deep Q network algorithm adjusts the path planning parameters and risk response strategies according to the positive and negative returns of the risk response data to achieve dynamic adaptability to complex marine environments.

[0032] A system for implementing the ship overtaking collision warning method in a dangerous area, the system comprising:

[0033] A multimodal sensor module is used to collect real-time status data of the ship, including position, speed, acceleration, heading and environmental information, and perform data preprocessing on the real-time status data of the ship to obtain fused sensor data;

[0034] The risk field generation module is used to generate local-scale risk field data and global-scale risk field data based on fused sensor data using multi-scale analysis, and to form real-time updated multi-scale dynamic risk field data through dynamic fusion to reflect the current and predicted future risk distribution;

[0035] A graph structure building module is used to build a graph structure of non-Euclidean space based on multi-scale dynamic risk field data, and use a graph neural network to predict the future trajectory of the ship to generate initial non-Euclidean trajectory prediction data;

[0036] A bidirectional coupling correction module is used to perform bidirectional coupling correction on the initial non-Euclidean trajectory prediction data in combination with the ship's operation behavior data to generate coupled trajectory prediction data;

[0037] The risk transmission and path optimization module is used to perform risk transmission path analysis and adaptive path optimization based on coupled trajectory prediction data to generate adaptive optimization path data;

[0038] A Bayesian fusion module is used to fuse the adaptive optimization path data with the coupled trajectory prediction data through a multi-level Bayesian fusion method to generate the final Bayesian path data;

[0039] The risk response and feedback module is used to evaluate and classify risk levels based on the final Bayesian path data and non-Euclidean risk transmission data, record the execution feedback data of each risk response, and self-learn and continuously optimize the feedback data through reinforcement learning algorithms.

[0040] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0041] The present invention realizes the comprehensive assessment of local and global risks through multi-scale dynamic risk field analysis technology, significantly improving the recognition accuracy of potential collision risks. Compared with the existing technology, the present invention introduces multi-scale analysis to dynamically integrate local and global risks, which can accurately reflect the current risk status of the ship and predict the future risk distribution.

[0042] The present invention achieves accurate estimation of the ship's future path and improves the reliability of path planning through trajectory prediction using graph neural network in non-Euclidean space. Traditional methods are usually limited to path calculation in Euclidean space, while the present invention uses graph structure for trajectory prediction, which can better reflect the nonlinear characteristics of complex marine environments.

[0043] The present invention optimizes the multi-level fusion effect of path data through a multi-level Bayesian fusion method, significantly improving the system's ability to handle different risk priorities; this technical means enables path planning to respond to local emergencies while taking into account the overall risk distribution;

[0044] The present invention improves the system's adaptive ability to complex marine environments through continuous optimization of the reinforcement learning algorithm; uses the deep Q network algorithm to self-learn the feedback data and continuously adjusts the risk response strategy, thereby improving the dynamic adaptability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the method of the present invention;

[0046] Figure 2 It is a schematic diagram of multi-scale dynamic risk analysis in the present invention;

[0047] Figure 3 A schematic diagram of trajectory prediction in non-Euclidean space in the present invention;

[0048] Figure 4 It is a schematic diagram of the multi-level Bayesian fusion method in the present invention;

[0049] Figure 5This is a system structure block diagram of the present invention. DETAILED DESCRIPTION

[0050] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0051] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0052] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0053] like Figure 1 As shown, a method for early warning of ship overtaking and collision in a dangerous area comprises the following steps:

[0054] Collect the real-time status data of the ship, including position, speed, acceleration, heading and environmental information, perform data preprocessing on the real-time status data of the ship to obtain fused sensor data; based on the fused sensor data, use multi-scale analysis to generate local scale risk field data and global scale risk field data, and form real-time updated multi-scale dynamic risk field data through dynamic fusion to reflect the current and predicted future risk distribution;

[0055] In the present invention, the ship overtaking and collision warning in dangerous areas requires the collection of real-time status data of the ship, including position, speed, acceleration, heading and environmental information. This information constitutes a comprehensive description of the ship's motion state and its surrounding environment, and is the basis for subsequent risk analysis and path optimization. In order to ensure the accuracy and timeliness of the data, the collected status data needs to be preprocessed. The core purpose of preprocessing is to eliminate data noise and solve the time synchronization problem between different sensors, so as to obtain fused sensor data. This process usually uses algorithms such as Kalman filtering or particle filtering to iteratively estimate and correct the data to improve the accuracy of the data and ensure the consistency of the data source.

[0056] Based on the fusion sensor data, the present invention uses multi-scale analysis to generate local-scale risk field data and global-scale risk field data. The core idea of ​​this multi-scale analysis is to divide the ship and its surrounding environment into different scale levels. The local scale is used to analyze the direct interaction between ships, such as the distance and speed changes between two adjacent ships; the global scale is used to analyze the risk trends in a larger range, such as the potential impact of large areas of densely populated ships or severe weather conditions on navigation. Through the convolutional neural network (CNN), the local-scale data can be feature extracted to identify close-range risk factors. The long short-term memory network (LSTM) is used for global-scale analysis, which can capture a wider range of environmental change trends based on time series data.

[0057] The result of multi-scale analysis is the generation of local-scale risk field data and global-scale risk field data, which respectively reflect the local and global risk characteristics. Next, the present invention forms real-time updated multi-scale dynamic risk field data through a dynamic fusion strategy. The principle of dynamic fusion is to weighted average the risk intensity of the local scale and the global scale to reflect the current and predicted future risk distribution. This weighting strategy is adjusted according to the speed, relative position and environmental complexity of the ship. For example, when the ship is fast and the surrounding ships are dense, the weight of the local scale is larger to capture the rapidly changing close-range risks; conversely, if the ship is in a large area of ​​open water or at a lower speed, the weight of the global scale is increased to reflect the macro risk trend.

[0058] The generation and updating of multi-scale dynamic risk field data enables the system to respond quickly to changes in the current environment. For example, when two ships are approaching and the speed difference is large, the local scale risk field can quickly identify this situation and appropriately increase the risk intensity, so that the system can activate the early warning mechanism in time. At the same time, the system can also combine historical environmental data to predict possible risk changes in the next few minutes, take avoidance measures in advance, and improve navigation safety.

[0059] Preferably, the multi-scale analysis to generate local-scale risk field data and global-scale risk field data includes:

[0060] The local area of ​​the fused sensor data is scanned through a convolutional neural network to extract the local feature information of the distance change and speed change between ships, and generate local-scale risk field data. The historical time series of the fused sensor data is modeled using a long short-term memory network to identify the trend of ship status changes over time and generate global-scale risk field data.

[0061] In the dangerous area ship overtaking collision warning of the present invention, multi-scale analysis is used to generate local scale risk field data and global scale risk field data, with the purpose of comprehensively evaluating risks from different spatial scales and time dimensions. This process mainly involves the combined application of convolutional neural networks (CNN) and long short-term memory networks (LSTM), which are responsible for local feature extraction and global time series analysis, respectively, to accurately identify potential risks and predict future risk change trends.

[0062] Convolutional neural networks (CNNs) are used to scan local areas of fused sensor data to extract key local feature information. The role of CNN in the present invention is to detect relative changes between ships, such as distance changes and speed changes. Specifically, CNN extracts feature maps of different scales through convolution operations, which reflect the movement patterns of ships in a local range. For example, when two ships approach in a narrow waterway, the convolution kernel can capture the relative speed difference and direction change between the ships, and use this information to calculate the local risk intensity and generate local scale risk field data. The multi-layer convolution structure of CNN enables it to extract complex nonlinear features layer by layer, which helps to distinguish different risk levels, such as sharp turns or rapid acceleration in normal navigation status and high-risk status. In addition, CNN can also automatically adapt to the noise in the data to avoid deviations in the overall risk judgment due to abnormalities in individual data points.

[0063] At the same time, the long short-term memory network (LSTM) is used to model the historical time series of fused sensor data to identify the trend of ship status changes over time. The main task of LSTM in the present invention is to capture the risk change pattern over a long span. Since LSTM can remember information over a long period of time and selectively remember or forget key features in historical data, it has unique advantages in global scale risk analysis. LSTM can predict the ship status and environmental change trends in the next few minutes to hours based on historical time series data, thereby generating global scale risk field data. For example, when multiple ships form a main channel at sea, LSTM can predict the future movement trend of ships by analyzing historical track data and identify high-risk areas that may cause collisions. LSTM can also incorporate this information into risk assessment by learning seasonal or periodic change patterns (such as tidal or wind direction changes) to improve the accuracy of predictions.

[0064] In the specific implementation process, the convolutional neural network first scans the fused sensor data in different regions, extracts the local feature information of each region, and combines these features to form local-scale risk field data. At the same time, the long short-term memory network outputs global-scale risk field data based on the historical change trend of the time series data. These two data sets represent the local interactions between ships and the global risk characteristics of the entire sea area, respectively. Next, the local-scale risk field data and the global-scale risk field data will be dynamically fused to generate multi-scale dynamic risk field data to provide comprehensive risk assessment results.

[0065] Through multi-scale analysis, the present invention can achieve fine-grained risk assessment and long-term trend prediction in the ship overtaking collision warning in dangerous areas. For example, in a port area with dense ships, local scale analysis can quickly identify the acceleration and turning behavior of close-range ships, and timely warn of potential collision risks; while in large-scale sea navigation, global scale analysis can predict the risk evolution trend of remote ships and help the captain formulate a safe risk avoidance strategy. This multi-scale risk assessment method improves the accuracy and response speed of the early warning system, enabling it to better adapt to complex marine environments.

[0066] Preferably, Figure 2 As shown in the figure, the multi-scale dynamic risk field data updated in real time through dynamic fusion includes:

[0067] The local-scale risk field data and the global-scale risk field data are converted into the risk intensity representation of the same unit, and multi-scale dynamic risk field data are generated through weighted fusion, where the calculation formula of risk intensity is:

[0068]

[0069] in, Represents the risk intensity of multi-scale dynamic risk field data; is the risk intensity of the local scale risk field data; is the risk intensity of the global-scale risk field data; and are the weight coefficients of the local scale and the global scale respectively, and satisfy , the weight coefficient is dynamically adjusted according to the speed and relative position of the ship.

[0070] In the ship overtaking and collision warning in dangerous areas, the key step to accurately assess and predict the risk status at sea is to form multi-scale dynamic risk field data updated in real time through dynamic fusion. The core principle of this process is to convert the local scale risk field data and the global scale risk field data into the risk intensity representation of the same unit, and weightedly fuse them according to the importance of different risk factors to generate multi-scale dynamic risk field data, thereby more comprehensively reflecting the real-time risk status of the marine environment.

[0071] The local-scale risk field data and the global-scale risk field data are standardized so that their risk intensities are comparable in the same dimension. The purpose of this step is to eliminate the dimensional differences between data of different scales and ensure that the multi-scale dynamic risk field data generated after fusion can reflect the real risk situation. The local-scale risk field data reflects the characteristic changes of ships in a local range. For example, in ports or narrow waterways, changes in distance, speed and heading between ships will significantly affect the local risk intensity; while the global-scale risk field data mainly captures risk trends in a larger range, such as the impact of environmental factors such as bad weather, tides or currents on the overall navigation safety of ships.

[0072] By weighted fusion, local-scale risk field data and global-scale risk field data are integrated into multi-scale dynamic risk field data. The risk intensity calculation formula in the fusion process is shown above. In this formula, the weight coefficient and Dynamic adjustment is the key to ensure the accuracy of risk assessment. In order to reflect the current changes in maritime risks, the weight coefficient is adjusted according to the speed and relative position of the ship. For example, when a ship is sailing at high speed in a narrow waterway, local risk factors have a greater impact on the safety of the ship, so it is necessary to increase the weight coefficient of the local scale risk field. , to enhance the sensitivity to close-range risks. When the ship is traveling at a lower speed in the open sea, it may be necessary to increase the weight coefficient of the global scale , in order to better reflect the impact of large-scale environmental changes on navigation.

[0073] In specific applications, in order to better achieve dynamic adjustment of weight coefficients, the weight adjustment strategy can be trained and optimized using machine learning algorithms (such as adaptive neural networks) in combination with the historical navigation data of ships and current environmental characteristics. For example, in certain specific sea areas, the repeatedly trained model can automatically identify high-risk areas (such as intersections where multiple ships converge) and automatically increase the local scale weight coefficients in these areas. , improving the accuracy of early warning of possible rear-end collisions. In addition, if frequent accidents caused by large-scale bad weather are found in the historical data of ships, the model will automatically increase the weight of the global scale to improve the early warning effect of the system under similar environmental conditions in the future.

[0074] The multi-scale dynamic risk field data formed by dynamic fusion can be updated in real time to reflect the current and predicted future distribution of maritime risks, thereby providing more accurate basic information for collision warning and path planning. Specifically, in the ship overtaking collision warning, the multi-scale dynamic risk field data can quickly perceive the changes in the relative positions of ships, and combine the global impact of the external environment to give the best risk response strategy. For example, when a ship in front is detected to suddenly slow down, the system will immediately adjust the risk weight, quickly update the risk intensity and issue a warning instruction to guide the ship to take necessary avoidance measures. At the same time, combined with the risk prediction of the global environment, the system can also predict the risk trend in the next few minutes to hours to ensure that effective response measures are taken before the risk occurs.

[0075] Based on multi-scale dynamic risk field data, a graph structure of non-Euclidean space is constructed, and the future trajectory of the ship is predicted using a graph neural network to generate initial non-Euclidean trajectory prediction data; bidirectional coupling correction is performed in combination with the ship's operational behavior data to obtain coupled trajectory prediction data; based on the coupled trajectory prediction data, risk transmission path analysis and adaptive path optimization are performed to generate adaptive optimization path data; the adaptive optimization path data is fused with the coupled trajectory prediction data through a multi-level Bayesian fusion method to generate the final Bayesian path data;

[0076] like Figure 3 As shown in the figure, in the ship overtaking and collision warning in dangerous areas, the present invention gradually generates accurate risk avoidance path data by constructing a non-Euclidean spatial graph structure based on multi-scale dynamic risk field data and using graph neural network (GNN) for trajectory prediction. This process includes multiple steps, from the construction of the graph structure to the final optimization of Bayesian fusion, aiming to provide efficient warning and path planning for the safe navigation of ships in complex environments.

[0077] Based on the multi-scale dynamic risk field data, a graph structure of non-Euclidean space is constructed. The core of this step is to convert the real-time status of the ship (such as position, speed, acceleration, etc.) and multi-scale risk information into the attributes of the nodes and edges of the graph. Specifically, each node in the graph structure represents a ship, and the weight of the edge represents the degree of risk association between ships, which can be calculated by relative distance, relative speed and risk intensity. For example, if the distance between two ships is very close and the relative speed is large, the edge weight will increase accordingly, reflecting a higher risk of collision. The non-Euclidean characteristics of this graph structure enable the system to better adapt to complex maritime environments, especially when the distribution of ships is irregular and the risk distribution is uneven, the non-Euclidean graph can more accurately represent the risk relationship between each other.

[0078] Using the constructed graph structure, the present invention uses a graph neural network to predict the future trajectory of the ship and generate initial non-Euclidean trajectory prediction data. The role of the graph neural network is to share and propagate information between nodes (ships) through a multi-layer message passing mechanism, thereby capturing the dynamic characteristics of ship interactions in the marine environment. Through multi-layer convolution operations, GNN can make high-precision predictions of the future position of the ship while considering the global network structure. For example, when a ship in front slows down or turns, the graph neural network can quickly perceive the risk change and predict possible risk avoidance paths, so that the system can issue an early warning.

[0079] The initial non-Euclidean trajectory prediction data and the ship's operating behavior data are bidirectionally coupled and corrected to obtain more accurate coupled trajectory prediction data. The core of the bidirectional coupling correction is to not only consider the ship's own operating behavior (such as turning rate, acceleration or deceleration), but also combine the potential impact of the behavior of surrounding ships on the current ship. For example, when a ship is sailing in a narrow channel, the speed change behavior of the ship in front will directly affect the trajectory prediction result of the ship behind. The bidirectional coupling correction can dynamically adjust these prediction deviations, thereby improving the accuracy of trajectory prediction.

[0080] After obtaining the coupled trajectory prediction data, the system further performs risk conduction path analysis and adaptive path optimization to generate adaptive optimized path data. Risk conduction path analysis identifies high-risk paths by calculating the risk conduction strength between ships. The conduction strength can be calculated based on the risk characteristics of each node (ship) and the weight of the edge. This step helps determine which paths are most likely to cause collisions or dangerous situations. Adaptive path optimization is based on the risk conduction path, and locally adjusts and globally optimizes each node of the path to find a safe and efficient avoidance path. The optimization process can be achieved through a variational adaptive gradient algorithm, which adjusts the steering angle and speed change of the path point according to the gradient information of the current path to ensure the smoothness and safety of the path.

[0081] The present invention fuses the adaptive optimization path data with the coupled trajectory prediction data through a multi-level Bayesian fusion method to generate the final Bayesian path data. The principle of Bayesian fusion is to weight and integrate data from different sources according to their credibility and priority, and the Bayesian hierarchy can be fused step by step according to risk priority and behavior priority. For example, if the ships in a certain area are dense and the speed is high, the Bayesian fusion process will give a higher weight to the adaptive optimization path based on local risks, while the weight of the global risk prediction path will be relatively reduced to better deal with high-risk short-term emergencies. Through multi-level fusion, the final Bayesian path data can not only reflect the high-risk situation in the short term, but also take into account the overall long-term trend, thereby providing the best navigation plan for the ship.

[0082] Preferably, constructing a graph structure of a non-Euclidean space includes:

[0083] The ship status in the multi-scale dynamic risk field data is used as the attribute of the node, including the real-time position, speed and acceleration information of the ship; the edge weight value between the nodes is determined according to the relative distance and relative speed between the ships, and the edge weight value is calculated according to the following formula:

[0084]

[0085] in, Representation Node and nodes The edge weights between ; For Node and nodes The relative distance between It is an adjustment parameter used to control the nonlinear change degree of edge weight; is the bias parameter; it is used to adjust the threshold of edge weight.

[0086] In the present invention, in order to realize the ship overtaking collision warning in the dangerous area, constructing the graph structure of non-Euclidean space is a key step. The purpose of this step is to represent the mutual relationship and risk information between ships through the graph structure, and to quantify the strength of these relationships through the edge weight value. Specifically, each node in the graph structure represents a ship, and the attributes of the node include the real-time status information of the ship, such as position, speed and acceleration, while the edge weight is used to represent the potential risk association degree between ships.

[0087] In the process of constructing the non-Euclidean graph structure, the ship status information in the multi-scale dynamic risk field data needs to be used as the attribute of the node first. The ship status information here is not just the static position, but also includes the dynamic speed and acceleration, which can more comprehensively reflect the movement trend of the ship. For example, if the speed and acceleration of two ships are similar, it means that their movement patterns are similar, and the edge weight may be small, indicating relative safety; when the acceleration of a ship suddenly increases and approaches another ship, the edge weight will increase accordingly, indicating that the risk of collision is increasing.

[0088] Next, the edge weights between nodes are calculated based on the relative distance and relative speed between the ships. The specific calculation formula is shown above. The core principle of this formula is to convert the distance information between ships into a standardized weight value, so that the edge weight value can vary between 0 and 1. The exponential function here is used to introduce nonlinear changes so that when the relative distance is close to the threshold When the relative distance is large, the edge weight value increases rapidly, reflecting a higher risk; when the relative distance is large, the edge weight value is close to zero, indicating that the risk impact between ships is small. The severity of this nonlinear change can be controlled, for example, a larger The value can make the edge weight change more sensitive, which is suitable for narrow waterways or high-density ship areas; while smaller The value can smooth the changes in edge weights and is more suitable for open ocean areas.

[0089] The edge weight values ​​calculated in this way can dynamically reflect the risk level between ships. For example, in a busy waterway, if a ship's speed suddenly increases while the ship in front of it is slowing down, the relative speed difference between the two ships will increase, and the corresponding edge weight value will also increase. In this case, the system will increase the risk assessment value of the edge to warn of the possible risk of rear-end collision. At the same time, if the two ships are far apart and heading in opposite directions, the edge weight will be close to zero, indicating that the risk association between them is low, which can reduce unnecessary consumption of computing resources and focus on the paths that are really risky.

[0090] By constructing a graph structure in non-Euclidean space, the present invention can more accurately quantify the risk associations between ships in a complex marine environment, especially when the distribution of ships is uneven or the risk distribution has complex nonlinear characteristics, the non-Euclidean graph can better represent these complex relationships. The construction of this graph structure not only provides high-quality input data for subsequent graph neural network trajectory prediction, but also can dynamically adjust the risk weights between ships, thereby improving the sensitivity and accuracy of the early warning. In summary, the process of constructing a non-Euclidean graph structure not only realizes the quantification of complex relationships between ships, but also provides a flexible risk analysis framework for collision warning, significantly enhancing the real-time and accuracy of the system.

[0091] Preferably, the bidirectional coupling correction includes:

[0092] The node positions and edge weights in the initial non-Euclidean trajectory prediction data are corrected according to the ship's operating behavior data. The operating behavior data include the heading change rate, acceleration change rate and turning speed. The coupled trajectory prediction data is generated by calculating the influence of the operating behavior data on the node positions and edge weights. The correction process adopts a two-way calculation, which not only considers the changes in the ship's own behavior, but also combines the interference of other ships' behaviors on the prediction results.

[0093] In the present invention, the core of the two-way coupling correction is to dynamically adjust the initial non-Euclidean trajectory prediction data based on the ship's operational behavior data to improve the accuracy of trajectory prediction. This process not only takes into account the behavior changes of the target ship itself, but also combines the interference effect of the behavior of other surrounding ships on the prediction results. This ensures that the predicted trajectory can be closer to the actual situation, which helps to achieve more accurate collision warning and path planning in complex marine environments.

[0094] The first step of the two-way coupling correction is to collect and analyze the ship's operational behavior data, mainly including the heading change rate, acceleration change rate and turning speed. The heading change rate reflects the speed at which the ship changes its course, that is, the turning agility of the ship; the acceleration change rate indicates the change in the ship's speed, such as the magnitude of acceleration or deceleration; and the turning speed indicates how fast the ship changes its course within a certain period of time. These data directly affect the ship's motion trajectory and path selection, and to a certain extent determine the level of collision risk. For example, when a ship's acceleration suddenly increases, it means that its speed changes rapidly in a short period of time, which may pose a potential collision threat to the ships in front of or around it.

[0095] In the specific implementation process, the bidirectional coupling correction requires dynamic adjustment of the node positions and edge weights in the initial non-Euclidean trajectory prediction data. The node position here represents the predicted position of the ship, while the edge weight represents the degree of risk association between ships. The correction process uses bidirectional calculation, which not only considers the impact of the ship's own behavior changes on the node position, but also combines the interference of other ships' behaviors on the edge weights. Specifically, when the rate of change of the heading or acceleration of a ship is large, the system will make appropriate corrections to its predicted position to reflect this change. For example, if a ship in front suddenly turns, the trajectory prediction of the ship behind should take this interference factor into account, thereby correcting the edge weight to increase the sensitivity to the risk in that direction.

[0096] The correction method of two-way calculation can be further explained as follows: the first step is self-influence correction, that is, the offset of the trajectory prediction result is calculated by the change of the ship's own operating behavior data. For example, if the acceleration of the ship increases, the node position in the predicted trajectory should be adjusted a certain distance forward to reflect the impact of the increase in speed. The second step is cross-interference correction, which corrects the edge weight of the target ship by analyzing the operating behavior data of other ships around. The correction of edge weights usually adopts a risk propagation-based model to calculate the degree of influence of changes in the behavior of other ships on the current ship. Assuming that there is a ship approaching at high speed, the system will automatically increase the edge weight value of the ship to the target ship to improve the accuracy of the prediction and the reliability of risk warning.

[0097] This bidirectional coupling correction is effective in collision warning in complex maritime environments. For example, in port areas where many ships gather, the traditional one-way prediction method may not be able to capture all potential risks due to the frequent changes in the operating behaviors of ships. However, through bidirectional coupling correction, the trajectory prediction can be updated at each time step, fully considering the impact of other ships on the current ship, thereby greatly improving the accuracy of the warning. This correction process can also be applied to dynamically changing environments, such as sudden deterioration of weather conditions or sudden changes in tidal flows. These factors can be incorporated into the correction model through analysis of operational behavior data, further enhancing the system's adaptability and flexibility.

[0098] Preferably, the risk transmission path analysis includes:

[0099] Based on the node and edge weight information in the coupled trajectory prediction data, the risk propagation algorithm of the graph neural network is used to calculate the risk conduction intensity of each path node. The calculation formula of the risk conduction intensity is:

[0100]

[0101] in, Representation Node The intensity of risk transmission; For Node For Node The edge weight of For Node The risk intensity is calculated and the risk transmission path data is used to identify high-risk areas.

[0102] In the ship overtaking and collision warning in dangerous areas, risk transmission path analysis is a risk propagation algorithm based on graph neural network (GNN) to calculate the risk transmission intensity of each path node to identify high-risk areas at sea. This method processes the weight information of nodes (ships) and edges (associations between ships) in the coupled trajectory prediction data to determine the risk propagation path and intensity in the graph structure, providing a reliable risk assessment basis for the early warning system.

[0103] The principle of risk transmission path analysis is to recursively propagate and accumulate risks in the graph structure based on the risk information of nodes and edges. Each node represents a ship, and the risk intensity of the node reflects the potential danger faced by the ship in the current environment. The edge weight indicates the degree of risk association between ships. The higher the weight value, the more significant the risk transmission between the two ships. By calculating the risk transmission of all nodes, high-risk areas that may cause collisions can be identified.

[0104] Specifically, the calculation formula for risk transmission intensity is shown above. The meaning of this formula is that the node The risk transmission intensity of the nodes is determined by the risk situation of the surrounding nodes and their relationship with the node The edge weights between them determine that the risk not only comes from its own status, but is also affected by other ships.

[0105] In the specific implementation, the system first determines the initial risk intensity of each node based on the status of each ship in the coupled trajectory prediction data (such as position, speed, acceleration, etc.) , and calculate the edge weights based on the relative positions and speeds between ships . Then, the graph neural network is used to calculate the risk transmission of all nodes. The graph neural network iteratively updates the risk intensity of the node at each layer. Through multi-layer propagation and accumulation, the final risk transmission intensity of each node not only considers the impact of directly adjacent nodes, but also indirectly reflects the risk situation of more distant nodes. This multi-layer propagation mechanism can capture the transmission characteristics of risks in complex maritime environments, especially in areas with densely populated ships, and can effectively identify the concentration points of risk transmission paths.

[0106] For example, if multiple ships are travelling in a similar direction and speed on a channel and one of them suddenly slows down or changes direction, the risk intensity of the ship will be At the same time, due to the high edge weights between ships, this risk will be quickly transmitted to neighboring nodes (ships), causing the risk transmission intensity of these ships to increase accordingly. At this time, the system can identify high-risk paths in the area and issue avoidance warnings to relevant ships, thereby preventing potential rear-end or collision accidents.

[0107] Through this risk transmission path analysis, the present invention can conduct real-time assessment of the complex dynamic environment at sea and find out which paths or areas have greater collision risks. This method can not only provide timely warnings before risks occur, but also provide key basis for path optimization and risk avoidance strategies. For the complex mutual influence between multiple ships, the risk propagation algorithm of the graph neural network can reflect the risk correlation degree between each ship in multi-layer propagation, significantly improving the sensitivity and accuracy of collision warning.

[0108] Preferably, the adaptive path optimization includes:

[0109] The variational adaptive gradient algorithm is used to iteratively adjust the steering angle and smoothness of each path point in the risk transmission path data to update the position of the path point. The optimized path generates adaptive optimized path data based on the local gradient information of the path point and the global risk distribution.

[0110] In the dangerous area ship overtaking collision warning of the present invention, the purpose of adaptive path optimization is to dynamically adjust the navigation path of the ship according to the analysis results of the risk transmission path data to avoid high-risk areas and ensure navigation safety. The core of adaptive path optimization is to use the variational adaptive gradient algorithm to iteratively adjust the steering angle of the path point and the smoothness of the path to update the position of the path point, thereby generating a safer and more efficient adaptive optimization path.

[0111] The principle of adaptive path optimization is to optimize the position of path points through the gradient descent method. First, high-risk areas in the path are identified based on the risk transmission path data. These areas usually have large risk transmission intensity or dense high-weight edges. For these high-risk areas, the adaptive path optimization algorithm will make adjustments based on the local gradient information of each path point. The gradient information reflects the risk increment of the path point, that is, how the risk intensity changes if the path point moves in a certain direction. By calculating the local gradient, the optimal direction to reduce the path risk can be found so that the path points avoid high-risk areas as much as possible. In addition, the algorithm will also combine the global risk distribution to optimize the path as a whole, so that it can meet the risk hedging needs while maintaining the smoothness and rationality of the path.

[0112] The variational adaptive gradient algorithm plays a key role in the optimization process. The algorithm iteratively updates the steering angle and position of each path point to minimize the overall risk on the path while maintaining the smoothness of the path. This optimization process is divided into two steps: first, local adjustment, each path point is moved slightly according to the risk gradient information around it to avoid abrupt sharp turns or unreasonable course changes; second, global smoothing, by adjusting the distance and steering angle between adjacent path points, the path is made smoother and more natural as a whole, reducing overly frequent path adjustments. This method can not only improve the seaworthiness of the path, but also reduce the difficulty of ship maneuvering, thereby enhancing the feasibility of path planning.

[0113] For example, in a narrow waterway where many ships converge, if the system detects that a certain area on the path ahead has a high risk conduction intensity (perhaps due to the deceleration or sharp turns of some ships), the adaptive path optimization algorithm will automatically adjust the ship's navigation direction to move to a low-risk area, while ensuring the smoothness of the steering process to avoid drastic changes in direction that affect navigation stability. In addition, the variational adaptive gradient algorithm can also make real-time adjustments to the path based on dynamic changes in the environment. For example, when a ship approaches a risk area, the system will automatically increase the frequency and amplitude of path adjustments to better adapt to the changing risk distribution, thereby minimizing the possibility of collision.

[0114] The effect of adaptive path optimization is to improve the adaptability and safety of ships in complex marine environments. Through dynamic iterative adjustment of path points, the system can respond to risk changes in a short time and find a safe and efficient avoidance path. This method is more flexible than traditional fixed path planning because it not only takes into account the current motion state of the ship, but also combines real-time risk data and environmental information to provide more intelligent decision support for ship overtaking and avoidance. In addition, adaptive path optimization can also take into account the operability and risk avoidance needs of the path, making the optimized path more in line with actual navigation conditions, thereby greatly reducing the risk of ship collision.

[0115] Preferably, Figure 4 As shown, the multi-level Bayesian fusion method includes:

[0116] Multiple Bayesian levels are set, and the fusion weight of each Bayesian level is set according to the specific risk priority and behavioral feature priority. The adaptive optimization path data and the coupled trajectory prediction data are weightedly fused layer by layer in turn to gradually form the final Bayesian path data.

[0117] In the ship overtaking and collision warning in dangerous areas, the multi-level Bayesian fusion method comprehensively processes the adaptive optimization path data and the coupled trajectory prediction data through layer-by-layer weighted fusion to generate the final Bayesian path data. The core of this method is to set multiple Bayesian levels, and the fusion weight of each level is dynamically adjusted according to the specific risk priority and behavior feature priority, so as to achieve the weight allocation of different data sources and achieve the effect of gradually optimizing the path.

[0118] The principle of the multi-level Bayesian fusion method is based on the hierarchical structure of Bayesian reasoning, and the data is weighted layer by layer during the fusion process. Each Bayesian level represents a data fusion priority or a risk assessment scale, such as local risk and global risk, the urgency and long-term nature of adaptive path adjustment, etc. By setting different weights for each level, the degree of influence of the fusion of multiple data can be controlled, so that the final fusion result can accurately reflect the risk status under the current navigation conditions.

[0119] Specifically, the first step is to set multiple Bayesian levels. Usually three to five levels are set to adapt to different risk situations and path adjustment needs. The fusion weight of each level is set according to the specific risk priority and behavior feature priority. For example, if the ship's operating behavior (such as sharp turns or rapid deceleration) shows high-risk behavior, the system will give higher weights to low-level stages (such as local short-term risk levels) and relatively lower weights to global long-term risk levels. This can give priority to solving current emergency risk problems and improve the system's response speed to emergencies.

[0120] The system performs weighted fusion of the adaptive optimization path data and the coupled trajectory prediction data layer by layer in turn. The weighted fusion process updates the probability distribution of the input data through the Bayesian formula to reflect the contribution of the weight of each layer to the final result. For example, in a certain Bayesian level, if the path data shows a high-risk area, the fusion process will give priority to the risk of this area and weight its impact on path planning. Then, in the next level, the output results of the previous level are fused again with the global trajectory prediction data and adjusted according to the new weight settings. This layer-by-layer fusion method helps to gradually refine the data during the fusion process and avoid information loss or calculation bias caused by simply weighting all the data at one time.

[0121] To better understand this method, an example can be given: suppose the current path planning needs to pass through a port area with dense ships. At this time, the local level risk assessment will be given a high weight, because the collision risk in the local dense area is greater, and the system needs to pay more attention to the impact of small operational changes of the ship on the path safety. In this case, when fusing local risk data, the system will give priority to processing local emergency avoidance needs from the adaptive path data, and appropriately reduce the impact of global risk data in global level fusion to ensure that local high-risk areas are avoided first.

[0122] The effect of the multi-level Bayesian fusion method is that it can not only balance short-term and long-term risk avoidance needs at different risk levels, but also dynamically adjust the fusion weights to cope with complex changes in the maritime environment. For example, when a ship is about to enter a high-risk narrow channel, the system can automatically increase the low-level local risk fusion weights, so that the path planning pays more attention to immediate risk avoidance needs; and when the ship is in a relatively safe open sea, the system will enhance the optimization of long-term navigation paths to improve overall navigation efficiency.

[0123] Based on the final Bayesian path data and non-Euclidean risk transmission data, the risk level is evaluated and classified; the execution feedback data of each risk response is recorded, and the feedback data is self-learned and continuously optimized through the reinforcement learning algorithm.

[0124] Preferably, the reinforcement learning algorithm is used to self-learn and continuously optimize the feedback data including:

[0125] The risk response data is trained through the deep Q network algorithm. The risk response data includes safety indicators after execution, response time and path adjustment range. The deep Q network algorithm adjusts the path planning parameters and risk response strategies according to the positive and negative returns of the risk response data to achieve dynamic adaptability to complex marine environments.

[0126] In the ship overtaking and collision warning in dangerous areas, evaluating and dividing the risk level based on the final Bayesian path data and non-Euclidean risk transmission data is one of the core steps to realize the warning system. This process determines the risk level of different areas and gives specific risk response measures by analyzing the complex relationship between path data and risk transmission. In addition, the system records the execution feedback data of each risk response, and uses reinforcement learning algorithms to self-learn and continuously optimize these feedback data to improve the system's adaptability to complex marine environments.

[0127] The risk level is assessed based on the final Bayesian path data and non-Euclidean risk transmission data. These data contain the risk distribution of the current ship and its surroundings. First, based on the risk transmission strength and edge weights of the nodes in the non-Euclidean risk transmission data, the system can identify high-risk areas and their possible risk transmission paths. These areas may be concentrated in narrow waterways with dense ships or in sea areas with poor meteorological conditions. Combined with the final Bayesian path data, the system can comprehensively judge the risk level of the ship when it travels along the path, such as dividing the path into three risk areas of low, medium and high, and setting corresponding risk response measures for each. Such risk division helps to take measures in advance when the ship approaches a high-risk area to avoid unexpected accidents.

[0128] On this basis, the system will record the execution feedback data of each risk response, including safety indicators, response time, and path adjustment after risk response. These data are not only an evaluation of the current risk response effect, but also an important basis for subsequent reinforcement learning algorithm training. Through reinforcement learning algorithms, especially the deep Q network (DQN) algorithm, the system can self-learn and continuously optimize the feedback data to gradually improve the accuracy of early warning and path planning.

[0129] The deep Q network algorithm is a reinforcement learning method based on the value function, which can continuously optimize the decision-making strategy of the system through positive and negative returns (rewards or penalties) to the risk response data. In the present invention, the deep Q network will evaluate the response effect of the early warning system according to different indicators of the feedback data (such as safety, response time and path adjustment amplitude). If the system's response time is short and the safety index is high, it means that the risk response effect of this path planning is good, then the system will give positive rewards to the corresponding strategy; if the response time is long or the safety index is low, a negative reward will be given. This reward mechanism enables the system to gradually learn which path planning strategies are optimal in the process of continuous risk response, so as to better adapt to the dynamic changes in the complex marine environment.

[0130] For example, when a ship is sailing in a complex marine environment, the system may need to weigh multiple path options. Suppose the system chooses a path that avoids high-risk areas, and through deep Q network evaluation, it is learned that the adjustment of this path is large, but after execution, the safety of navigation is significantly improved. The system will record this strategy as a successful case in the experience library. The next time a similar situation is encountered, the system will give priority to this strategy to reduce the risk of collision. In another example, if the system chooses a relatively smooth path but fails to avoid the newly emerging high-risk area in time, the system will give a negative reward to this decision, prompting it to handle similar scenarios more carefully in subsequent training.

[0131] Through the self-learning and continuous optimization of the reinforcement learning algorithm, the system can not only dynamically adjust the path planning parameters, but also improve the risk response strategy, so that it can gradually adapt to various complex and changing maritime navigation environments. This adaptive ability makes the system not limited to a single preset response measure, but also able to adjust the strategy according to the actual situation, improving the flexibility and accuracy of the early warning system.

[0132] like Figure 5 As shown, a system for implementing the ship overtaking collision warning method in a dangerous area, the system comprises:

[0133] The multimodal sensor module is used to collect the real-time status data of the ship, including position, speed, acceleration, heading and environmental information, and preprocess the real-time status data of the ship to obtain fused sensor data; collect the real-time status data of the ship (position, speed, acceleration, heading and environmental information), and preprocess the data to generate fused sensor data. This module can integrate multi-source data to form a comprehensive understanding of the current status of the ship.

[0134] The risk field generation module is used to generate local-scale risk field data and global-scale risk field data based on fused sensor data using multi-scale analysis, and to form real-time updated multi-scale dynamic risk field data through dynamic fusion to reflect the current and predicted future risk distribution; based on fused sensor data, multi-scale analysis is used to generate local and global-scale risk field data. Through the dynamic fusion method, the risk field is updated in real time to reflect the current and future risk distribution. This enables the system to capture instantaneous local risks and long-term trend global risks.

[0135] The graph structure construction module is used to construct the graph structure of non-Euclidean space based on multi-scale dynamic risk field data, and use the graph neural network to predict the future trajectory of the ship to generate initial non-Euclidean trajectory prediction data; use multi-scale dynamic risk field data to construct the graph structure of non-Euclidean space, and predict the future trajectory of the ship through the graph neural network (GNN). This module achieves more accurate trajectory prediction by modeling complex spatial relationships.

[0136] The bidirectional coupling correction module is used to perform bidirectional coupling correction on the initial non-Euclidean trajectory prediction data in combination with the ship's operational behavior data to generate coupled trajectory prediction data; combined with the ship's operational behavior data (such as turning rate, acceleration change, etc.), the initial trajectory is bidirectionally corrected to generate coupled trajectory prediction data that is more in line with the actual navigation situation. The correction process takes into account the impact of the ship's own behavior and the behavior of surrounding ships on the prediction.

[0137] The risk transmission and path optimization module is used to perform risk transmission path analysis and adaptive path optimization based on coupled trajectory prediction data to generate adaptive optimization path data; risk transmission analysis is performed based on coupled trajectory prediction data, and the steering angle and position of the path points are optimized using the adaptive path optimization method to generate an adaptive optimization path. This process enables the system to flexibly adjust the navigation path according to the risk distribution.

[0138] The Bayesian fusion module is used to fuse the adaptive optimization path data with the coupled trajectory prediction data through a multi-level Bayesian fusion method to generate the final Bayesian path data; through a multi-level Bayesian fusion method, the adaptive optimization path data and the coupled trajectory prediction data are weightedly fused layer by layer to generate the final Bayesian path data. This can take into account the urgency of local risks and the stability of global navigation.

[0139] The risk response and feedback module is used to evaluate and classify risks based on the final Bayesian path data and non-Euclidean risk transmission data, record the execution feedback data of each risk response, and self-learn and continuously optimize the feedback data through reinforcement learning algorithms. Risk assessment and classification are performed based on the final Bayesian path data and non-Euclidean risk transmission data. The feedback data of risk response is recorded, and the feedback data is continuously optimized through reinforcement learning algorithms (such as deep Q networks) to improve the dynamic adaptability of the system.

[0140] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware.

[0141] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A ship overtaking collision warning method in a dangerous area, characterized in that: The following steps are involved: Collecting real-time status data of the ship, including position, speed, acceleration, heading and environmental information, and performing data preprocessing on the real-time status data of the ship to obtain fused sensor data; Based on the fusion sensor data, multi-scale analysis is used to generate local-scale risk field data and global-scale risk field data, and real-time updated multi-scale dynamic risk field data is formed through dynamic fusion to reflect the current and predicted future risk distribution; Multi-scale analysis generates local-scale risk field data and global-scale risk field data including: The local area of ​​the fused sensor data is scanned by a convolutional neural network to extract the local feature information of the distance change and speed change between ships, and generate local scale risk field data; the historical time series of the fused sensor data is modeled using a long short-term memory network to identify the trend of ship status changes over time and generate global scale risk field data; The multi-scale dynamic risk field data updated in real time through dynamic fusion includes: The local-scale risk field data and the global-scale risk field data are converted into the risk intensity representation of the same unit, and multi-scale dynamic risk field data are generated through weighted fusion, where the calculation formula of risk intensity is: in, Represents the risk intensity of multi-scale dynamic risk field data; is the risk intensity of the local scale risk field data; is the risk intensity of the global-scale risk field data; and are the weight coefficients of the local scale and the global scale respectively, and satisfy ,The weight coefficient is dynamically adjusted according to the speed and relative position of the ship; Based on multi-scale dynamic risk field data, a graph structure of non-Euclidean space is constructed, and the future trajectory of the ship is predicted using a graph neural network to generate initial non-Euclidean trajectory prediction data; bidirectional coupling correction is performed in combination with the ship's operational behavior data to obtain coupled trajectory prediction data; based on the coupled trajectory prediction data, risk transmission path analysis and adaptive path optimization are performed to generate adaptive optimization path data; the adaptive optimization path data is fused with the coupled trajectory prediction data through a multi-level Bayesian fusion method to generate the final Bayesian path data; The construction of the graph structure of non-Euclidean space includes: taking the ship status in the multi-scale dynamic risk field data as the attribute of the node, including the real-time position, speed and acceleration information of the ship; determining the edge weight value between the nodes according to the relative distance and relative speed between the ships, and the edge weight value is calculated according to the following formula: in, Representation Node and nodes The edge weights between ; For Node and nodes The relative distance between It is an adjustment parameter used to control the nonlinear change degree of edge weight; is the offset parameter; it is used to adjust the threshold of edge weight; The two-way coupling correction includes: correcting the node positions and edge weights in the initial non-Euclidean trajectory prediction data according to the ship's operating behavior data. The operating behavior data includes the heading change rate, acceleration change rate and turning speed. By calculating the influence of the operating behavior data on the node positions and edge weights, the coupled trajectory prediction data is generated. The correction process adopts a two-way calculation, that is, considering the changes in the ship's own behavior and combining the interference of other ship behaviors on the prediction results; The risk transmission path analysis includes: based on the node and edge weight information in the coupled trajectory prediction data, the risk transmission intensity of each path node is calculated using the risk propagation algorithm of the graph neural network. The calculation formula of the risk transmission intensity is: in, Representation Node The intensity of risk transmission; For Node For Node The edge weight of For Node The calculated risk transmission path data is used to identify high-risk areas; Adaptive path optimization includes: using a variational adaptive gradient algorithm to iteratively adjust the steering angle and smoothness of each path point in the risk transmission path data to update the position of the path point. The optimized path generates adaptive optimization path data based on the local gradient information of the path point and the global risk distribution; The multi-level Bayesian fusion method includes: setting multiple Bayesian levels, setting the fusion weight of each Bayesian level according to the specific risk priority and behavior feature priority, and weightedly fusing the adaptive optimization path data and the coupled trajectory prediction data layer by layer in turn to gradually form the final Bayesian path data; Based on the final Bayesian path data and non-Euclidean risk transmission data, the risk level is evaluated and classified; the execution feedback data of each risk response is recorded, and the feedback data is self-learned and continuously optimized through the reinforcement learning algorithm; The reinforcement learning algorithm is used for self-learning and continuous optimization of feedback data, including: training risk response data through the deep Q network algorithm. The risk response data includes safety indicators, response time and path adjustment range after execution. The deep Q network algorithm adjusts the path planning parameters and risk response strategies according to the positive and negative returns of the risk response data to achieve dynamic adaptability to complex marine environments.

2. A system for implementing the ship overtaking collision warning method in dangerous areas according to claim 1, characterized in that: The system comprises: A multimodal sensor module is used to collect real-time status data of the ship, including position, speed, acceleration, heading and environmental information, and perform data preprocessing on the real-time status data of the ship to obtain fused sensor data; The risk field generation module is used to generate local-scale risk field data and global-scale risk field data based on fused sensor data using multi-scale analysis, and to form real-time updated multi-scale dynamic risk field data through dynamic fusion to reflect the current and predicted future risk distribution; A graph structure building module is used to build a graph structure of non-Euclidean space based on multi-scale dynamic risk field data, and use a graph neural network to predict the future trajectory of the ship to generate initial non-Euclidean trajectory prediction data; A bidirectional coupling correction module is used to perform bidirectional coupling correction on the initial non-Euclidean trajectory prediction data in combination with the ship's operation behavior data to generate coupled trajectory prediction data; The risk transmission and path optimization module is used to perform risk transmission path analysis and adaptive path optimization based on coupled trajectory prediction data to generate adaptive optimization path data; A Bayesian fusion module is used to fuse the adaptive optimization path data with the coupled trajectory prediction data through a multi-level Bayesian fusion method to generate the final Bayesian path data; The risk response and feedback module is used to evaluate and classify risk levels based on the final Bayesian path data and non-Euclidean risk transmission data, record the execution feedback data of each risk response, and self-learn and continuously optimize the feedback data through reinforcement learning algorithms.

Images