Dynamic assessment of collision risk and collision avoidance decision-making method for trailing suction hopper dredger

Abstract

The application discloses a method for dynamically evaluating and avoiding collision of a trailing suction dredger, and relates to the technical field of navigation collision risk evaluation. The method lays a foundation for accurate evaluation through standardization processing of full-dimensional basic data, dynamically determines a potential risk area, quantifies a collision risk degree and a response level in combination with ship maneuvering and environmental characteristics to improve the timeliness and accuracy of risk early warning, constructs a personalized operation strategy library relying on an LSTM model to make the collision avoidance instruction conform to the artificial habit and reduce the human-machine adaptation cost, and further enhances the anti-interference ability and real-time performance of the instruction through a feedforward-feedback dual-path control and a parallel DQN dynamic optimization weight, so as to solve the problems that in the prior art, the evaluation basis is not unified due to the disordered and non-standard multi-source data, the potential collision risk area identification lacks dynamic characteristics without combination with the ship maneuvering characteristics, the collision risk quantification is fuzzy, the collision avoidance instruction is separated from the artificial operation habit, the traditional control has weak anti-interference ability and the fixed weight leads to insufficient robustness.

Description

Technical Field

[0001] This invention relates to the field of navigation collision risk assessment technology, specifically to a dynamic assessment and collision avoidance decision-making method for navigation collision risks of trailing suction hopper dredgers. Background Technology

[0002] The collision risk assessment and collision avoidance decision-making system for trailing suction hopper dredgers is a core technology for ensuring their safe operation at sea. Its performance directly determines navigation safety, operational efficiency, and compliance with international collision avoidance regulations (COLREGs). Early collision avoidance technologies relied heavily on a dual-threshold early warning mechanism based on TCPA (Time to Closest Encounter) and DCPA (Distance to Closest Encounter)—an alarm was triggered when the predicted DCPA was less than 0.5 nautical miles or the TCPA was less than 10 minutes. While this method is efficient and feasible in simple scenarios such as single-ship encounters, current technology still faces three major challenges, the specific shortcomings of which are as follows:

[0003] First, the ability to analyze complex scenarios is insufficient. Existing models (including DSFM and RLDS) still rely on preset logic to handle dynamic scenarios such as multi-ship games and COLREGs rule conflicts. They cannot adapt to sudden environmental changes and are prone to one-sided risk identification or strategy adaptation deviations, making it difficult to cover the comprehensive scenario of "multi-objective + complex sea conditions" in dredging operations.

[0004] Secondly, there is a trade-off between response speed and accuracy. While potential energy field models such as DSFM can improve prediction accuracy, the calculation of three-dimensional potential energy fields takes several seconds. On the other hand, simplified algorithms that pursue response speed sacrifice accuracy. When faced with sudden threats at very close range, the response delay may exceed the safe avoidance window, which cannot meet the dual requirements of "emergency collision avoidance + continuous operation" in dredging operations.

[0005] Finally, the robustness of the strategies and the adaptability of human-machine collaboration are lacking. Reinforcement learning strategies (such as RLDS) are prone to performance degradation under extreme conditions, and the "black box" training characteristics lead to a lack of interpretability of the strategies, making it difficult to pass maritime safety verification. At the same time, existing systems do not adequately consider human-machine collaboration. The International Maritime Accident Database (GISIS) shows that 26% of dredging accidents are caused by delays in control switching, and traditional authority allocation systems cannot smoothly transition control weights. Furthermore, personalized driving strategies have poor adaptability, and only 65% ​​of captains accept the system's operating logic, which can easily lead to human-machine conflicts.

[0006] Therefore, there is an urgent need for a collision risk assessment and collision avoidance decision-making scheme that can adapt to complex scenarios, balance response speed and accuracy, and combine robustness and human-machine collaboration. This scheme can solve the above-mentioned technical bottlenecks by using multi-level fuzzy logic early warning, multi-objective reinforcement learning strategy optimization, and virtual-real collaborative verification platform, combined with dynamic permission allocation, LSTM personalized driving strategy library, and parallel DQN optimized reinforcement learning enhanced RMPC framework, thus building an all-weather, all-scenario safety decision-making shield for trailing suction hopper dredgers. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a dynamic assessment and collision avoidance decision-making method for the navigation collision risk of trailing suction hopper dredgers. This method solves the problems in existing technologies, such as inconsistent assessment basis due to the chaotic and unstandardized multi-source data, lack of dynamism in identifying potential collision risk areas without considering ship handling characteristics, ambiguous collision risk quantification and collision avoidance commands deviating from human operating habits, and insufficient robustness due to the weak anti-interference capability and fixed weights of traditional control systems.

[0008] To achieve the above objectives, the present invention provides a method for dynamic assessment and collision avoidance decision-making of collision risks for trailing suction hopper dredgers, comprising the following steps:

[0009] Standardize the basic data of the trailing suction hopper dredger's navigation and operation in all dimensions to obtain a standardized dataset. The basic data of the trailing suction hopper dredger's navigation and operation in all dimensions includes the ship's own data, environmental data and personnel operation data.

[0010] Ship handling characteristics and environmental characteristics parameters are extracted from standardized datasets, and potential collision risk zones are determined based on these parameters.

[0011] Determine the collision hazard level and corresponding response level signals within the potential collision risk zone;

[0012] Based on the trained LSTM model, personalized operation strategies are determined according to different navigation scenarios, collision risk levels and corresponding response level signal combinations, and a personalized operation strategy library is established.

[0013] By employing feedforward-feedback dual-path control and parallel DQN dynamic optimization of weights, potential collision risk zones and collision hazard levels are assessed. Based on flow field forecasts, advance compensation is provided and rudder angles are corrected in real time to output the final collision avoidance command.

[0014] Furthermore, the process of determining potential collision risk zones based on ship maneuvering characteristics and environmental characteristic parameters is as follows:

[0015] Obtain feature coefficient data;

[0016] Based on the characteristic coefficient data, the potential energy value of each trailing suction hopper dredger is determined by the ship maneuvering characteristics and environmental characteristic parameters.

[0017] The potential energy value of the ship is superimposed with the potential energy value of the target ship to obtain the potential energy field superposition value. If the potential energy field superposition value is greater than the set potential energy threshold, the area where the ship and the target ship are located is marked as a risk sub-region.

[0018] The potential collision risk zone is obtained by superimposing all the risk sub-regions.

[0019] Furthermore, the ship maneuvering characteristics include standardized values ​​of mud hopper loading and ship speed; the environmental characteristic parameters include standardized values ​​of the encounter distance between the target ship and the ship and the distance between the ship and the target ship; and the characteristic coefficient data include the maneuvering response delay coefficient corresponding to the mud hopper loading, the detection error correction coefficient corresponding to visibility, and the COLREGs rule weight.

[0020] The method for obtaining the potential energy value is as follows:

[0021] The first potential energy factor is obtained by multiplying the product of the standardized value of the mud hopper loading and the standardized value of the ship speed by the ratio of the standardized value of the distance between the ship and the target ship, and then multiplying the ratio of the product of the product of the mud hopper loading and the standardized value of the ship speed to the ratio of the product ...

[0022] The second potential energy factor is obtained by multiplying the detection error correction coefficient corresponding to visibility with the COLREGs rule weight and the reciprocal of the normalized value of the encounter distance between the target ship and the ship.

[0023] The potential energy value is obtained by adding the first potential energy factor and the second potential energy factor.

[0024] Furthermore, the process of determining the collision hazard level and corresponding response level signal within the potential collision risk zone is as follows:

[0025] The channel density is obtained by the ratio of the number of vessels in the potential collision risk zone to the area of ​​the potential collision risk zone; the speed difference is obtained by the difference between the standardized speed of the vessel and the standardized speed of the target vessel.

[0026] A triangular membership function is applied to the superposition values ​​of channel density, speed difference, and potential energy field to assign membership degrees to different fuzzy levels of channel density, speed difference, and potential energy field, resulting in fuzzy feature vectors. The assignment principle is that the median value of the fuzzy level is the peak membership degree of 1, which linearly decreases to 0 towards the two boundaries.

[0027] The fuzzy feature vectors are compared according to the preset collision risk level rule base to determine the effective rules to be activated, and the final membership degree is determined according to the effective rules to be activated. The collision risk level rule base adopts the IF-THEN logical structure, where IF represents the input condition, the format of which is the same as the fuzzy feature vector, and THEN represents the output result, which is the fuzzy risk level.

[0028] The collision hazard is obtained by defuzzifying the final membership degree. Based on the preset response level correspondence rule, the collision hazard and the corresponding response level signal are obtained.

[0029] Furthermore, the process of comparing fuzzy feature vectors according to a preset collision risk level rule base, determining the effective activated rules, and determining the final membership degree based on the effective activated rules is as follows:

[0030] The matching degree between the fuzzy feature vector and each input condition in the collision risk level rule base is calculated, and the IF-THEN rules with a matching degree greater than 0 are identified and recorded as the activated valid rules.

[0031] For the valid activated rules, the final membership degree of each fuzzy risk level is calculated using the maximum membership degree synthesis method, that is, the maximum matching degree of the fuzzy risk level corresponding to all activated rules is taken.

[0032] Furthermore, the process of defuzzifying the final membership degree to obtain the collision hazard degree is as follows:

[0033] For each fuzzy risk level, assign a clear numerical range of collision hazard degree and determine the midpoint value of the range;

[0034] Determine the sum of the products of the final membership degree of each fuzzy risk level and the midpoint value of the corresponding fuzzy risk level interval, and then determine the sum of the final membership degrees of each fuzzy risk level.

[0035] The ratio of the sum of products to the sum of final membership degrees is used as the collision hazard level.

[0036] Furthermore, the process of establishing a personalized operation strategy library is as follows:

[0037] Using historical operational data as input samples and collision hazard level and corresponding response level signals as scenario risk labels, a scenario feature-operation action training dataset is constructed. The scenario features include mud hopper loading, sea state level, and collision hazard response level, while the operation actions include rudder angle adjustment range, rudder angle adjustment frequency, and braking trigger delay.

[0038] The network structure is set as an input layer, a hidden layer and an output layer. The input layer has a feature dimension of 6, the hidden layer has 2 layers with 64 neurons per layer, and the output layer has an output dimension of 3, which correspond to the rudder angle adjustment range, rudder angle adjustment frequency and braking trigger delay, respectively.

[0039] The loss function is the minimum sum of squared errors between the model's output actions and the captain's actual actions, and the model is iterated until the loss value is ≤0.05.

[0040] Based on the trained LSTM model, personalized operation strategies are generated by combining collision risk and corresponding response level signals for different navigation scenarios, thus constructing a personalized driving strategy library.

[0041] Furthermore, the process of outputting the final collision avoidance command is as follows:

[0042] Determine the feedforward compensation rudder angle based on flow field prediction of potential collision risk zones;

[0043] The deviation between the current ship position and the safe course is multiplied by the feedback correction gain to obtain the feedback correction rudder angle;

[0044] Based on parallel DQN optimization of feedforward weights and feedback weights, the product of feedforward weights and feedforward compensation rudder angles and the product of feedback weights and feedback correction rudder angles are calculated.

[0045] Obtain the rudder angle adjustment range adapted to the current situation from the personalized operation strategy library, add the rudder angle adjustment range to the two products, and obtain the final rudder angle command;

[0046] Confirm speed adjustment instructions;

[0047] Verify the speed adjustment command and the final rudder angle command. If the nearest encounter distance is greater than the distance threshold and the nearest encounter time is greater than the time threshold, the verification is successful; otherwise, adjust the speed adjustment command and the final rudder angle command.

[0048] The output of the successfully verified speed adjustment command and final rudder angle command is the final collision avoidance command.

[0049] Furthermore, the process of determining the feedforward compensation rudder angle based on the flow field prediction of the potential collision risk zone is as follows:

[0050] The flow velocity is obtained from the flow field prediction, and the flow field compensation coefficient and correction coefficient corresponding to the flow velocity are obtained from the database.

[0051] Calculate the ratio of the current velocity magnitude to the reference current velocity, and calculate the angle between the current direction and the ship's course;

[0052] Multiply the flow field compensation coefficient, correction coefficient, ratio, and included angle to obtain the feedforward compensation rudder angle;

[0053] The process of determining the speed adjustment instruction is as follows:

[0054] Obtain the relative distance between the current vessel and the target vessel, and the shortest distance from the current vessel to the boundary of the risk zone;

[0055] When the shortest distance is less than the set shortest threshold, calculate the absolute value of the difference between the target ship's speed and the ship's speed, and use it as the relative approximation speed.

[0056] The corresponding deceleration ratio is obtained from the database based on the shortest distance. The deceleration ratio is multiplied by the ship's speed to obtain the speed adjustment command. If the verification fails, the deceleration ratio is corrected.

[0057] Furthermore, the process of optimizing the feedforward and feedback weights based on parallel DQN is as follows:

[0058] The state inputs for the parallel DQN are constructed, including the current ship position deviation, collision risk level, risk zone distance, flow field intensity, feedforward compensation rudder angle, and feedback correction rudder angle.

[0059] Parallel DQN sets up two output heads, including feedforward channel weights and feedback channel weights;

[0060] The constraint is that the sum of the feedforward channel weight and the feedback channel weight is 1;

[0061] During the training phase, DQN learns the optimal feedforward and feedback channel weights through a reward function. When validation fails, the feedforward and feedback channel weights are re-optimized. The reward function is the sum of the reward and penalty values, where:

[0062] The reward value for stable track is 2, the reward value for successful collision avoidance is 5, the penalty value for entering the risk zone is -3, and the penalty value for heading oscillation is -2.

[0063] The present invention has the following beneficial effects:

[0064] This dynamic assessment and collision avoidance decision-making method for trailing suction hopper dredgers lays a solid foundation for accurate assessment through standardized processing of comprehensive basic data. It dynamically determines potential risk zones and quantifies collision hazard and response levels by combining ship maneuvering and environmental characteristics to improve the timeliness and accuracy of risk warnings. It builds a personalized operation strategy library based on LSTM models to make collision avoidance commands conform to human habits and reduce human-machine adaptation costs. Furthermore, it enhances the anti-interference capability and real-time performance of commands through feedforward-feedback dual-path control and parallel DQN dynamic weight optimization. This method can solve the problems in existing technologies, such as inconsistent assessment basis due to the chaotic and unstandardized multi-source data, lack of dynamism in identifying potential collision risk zones without combining ship maneuvering characteristics, ambiguous collision risk quantification and collision avoidance commands deviating from human operating habits, and insufficient robustness due to weak anti-interference capability and fixed weights in traditional control systems.

[0065] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0066] Figure 1 This is a flowchart of the dynamic assessment and collision avoidance decision-making method for the navigation collision risk of the trailing suction hopper dredger of the present invention. Detailed Implementation

[0067] Please see Figure 1 The present invention provides a technical solution: a dynamic assessment and collision avoidance decision method for the navigation collision risk of a trailing suction hopper dredger, comprising the following steps: standardizing the basic data of the trailing suction hopper dredger's navigation and operation in all dimensions to obtain a standardized dataset, wherein the basic data of the trailing suction hopper dredger's navigation and operation in all dimensions includes the ship's own data, environmental data and personnel operation data.

[0068] Ship handling characteristics and environmental characteristics parameters are extracted from standardized datasets, and potential collision risk zones are determined based on these parameters.

[0069] The process of determining potential collision risk zones based on ship maneuvering characteristics and environmental characteristics parameters is as follows:

[0070] Obtain feature coefficient data;

[0071] Based on the characteristic coefficient data, the potential energy value of each trailing suction hopper dredger is determined by the ship maneuvering characteristics and environmental characteristic parameters.

[0072] The potential energy value of the ship is superimposed with the potential energy value of the target ship to obtain the potential energy field superposition value. If the potential energy field superposition value is greater than the set potential energy threshold, the area where the ship and the target ship are located is marked as a risk sub-region.

[0073] The potential collision risk zone is obtained by superimposing all the risk sub-regions.

[0074] Characteristic coefficient data can incorporate the ship's own maneuvering status (such as the impact of full load / empty load on maneuvering), environmental detection bias (such as errors caused by low visibility), and international collision avoidance rules into risk assessment, avoiding the one-sidedness of traditional risk zone identification that relies solely on single factors such as distance, and making risk assessment more in line with actual navigation conditions. By combining ship maneuvering characteristics (mud hull loading, standardized ship speed) and environmental characteristics (encounter distance, standardized distance between two ships) to calculate potential energy values, risk quantification can dynamically match the ship's own state and environmental changes, solving the problem of traditional risk zone identification lacking dynamism and failing to reflect the differences in maneuvering of different ships, making single-ship risk assessment more accurate.

[0075] By superimposing potential energy fields, high-risk local areas of interaction between the vessel and a single target vessel can be accurately located, avoiding misjudgments caused by large-scale, vague delineation of risk areas (such as including areas without collision risk), thus improving the accuracy of local risk identification. By integrating risk sub-regions of interaction between multiple target vessels and the vessel, a global potential collision risk range is formed, solving the problem that traditional identification only focuses on a single target vessel and cannot cover complex scenarios with multiple target vessels. This ensures the comprehensiveness of risk area identification, provides an accurate risk area basis for subsequent collision hazard calculation and collision avoidance decision-making, and reduces the possibility of missing high-risk areas.

[0076] The ship maneuvering characteristics include standardized values ​​of mud hopper loading and ship speed; the environmental characteristic parameters include standardized values ​​of the encounter distance between the target ship and the ship and the distance between the ship and the target ship; and the characteristic coefficient data include the maneuvering response delay coefficient corresponding to mud hopper loading, the detection error correction coefficient corresponding to visibility, and the COLREGs rule weight.

[0077] The method for obtaining the potential energy value is as follows:

[0078] The first potential energy factor is obtained by multiplying the product of the standardized value of the mud hopper loading and the standardized value of the ship speed by the ratio of the standardized value of the distance between the ship and the target ship, and then multiplying the ratio of the product of the product of the mud hopper loading and the standardized value of the ship speed to the ratio of the product ...

[0079] The second potential energy factor is obtained by multiplying the detection error correction coefficient corresponding to visibility with the COLREGs rule weight and the reciprocal of the normalized value of the encounter distance between the target ship and the ship.

[0080] The potential energy value is obtained by adding the first potential energy factor and the second potential energy factor.

[0081] Determine the collision hazard level and corresponding response level signals within the potential collision risk zone.

[0082] The ship maneuvering characteristics are defined as standardized values ​​of hopper loading and ship speed (the maneuverability of trailing suction hopper dredgers is significantly affected by changes in hopper loading, which is a key characteristic that distinguishes them from ordinary ships). Environmental characteristic parameters are defined as standardized values ​​of the encounter distance between the target ship and the ship itself, and standardized values ​​of the distance between the ship itself and the target ship. Characteristic coefficient data are defined as the maneuvering response delay coefficient corresponding to hopper loading, the detection error correction coefficient corresponding to visibility, and the COLREGs rule weight. This solves the problems of vague parameter boundaries and lack of focus on the core maneuvering and environmental impact factors of dredgers in traditional risk assessments. It makes the input parameters for subsequent potential energy value calculation more consistent with the actual navigation of dredgers and avoids irrelevant parameters interfering with the accuracy of risk quantification.

[0083] By coupling the ship's own maneuvering status (the mud hopper load determines maneuvering flexibility, and ship speed affects the probability of collision), the spatial distance between the two ships, and the maneuvering response delay coefficient (the coefficient is larger when the mud hopper is fully loaded, reflecting high maneuvering delay and difficulty in avoidance), the problem of traditional potential energy calculation relying only on distance and ignoring the differences in ship maneuvering ability is solved. It can dynamically reflect the risk level of the dredger under different loading states, making the quantification of risks related to single ship maneuvering more accurate.

[0084] By combining the detection error correction coefficient corresponding to visibility (such as adjusting the coefficient when visibility is low to correct the impact of detection data deviation on risk assessment), the COLREGs rule weight (reflecting the collision avoidance responsibility and risk level under different encounter scenarios, such as the weight of the encounter situation being higher than that of the overtaking situation), and the encounter distance between the two ships, the problem of traditional potential energy calculation ignoring environmental detection errors and not being adapted to international collision avoidance rules, resulting in risk quantification not conforming to actual navigation rules and environmental limitations, is solved. This makes the environment-rule-related risk quantification more in line with navigation regulations and actual detection conditions.

[0085] By combining the ship's own maneuvering risk (first factor) and the environmental-rule adaptation risk (second factor), a comprehensive single-ship potential energy quantification result is formed. This solves the problem that traditional risk quantification only focuses on a single dimension (such as only looking at distance) and cannot cover the multi-dimensional risks of maneuvering, environment and rules. It provides an accurate and comprehensive risk quantification basis for the subsequent superposition of the potential energy fields of the ship and the target ship and the determination of potential collision risk areas, avoiding misjudgment of risk areas due to the lack of risk dimensions (such as missing high-risk areas in low visibility + full load conditions).

[0086] The channel density is obtained by the ratio of the number of vessels in the potential collision risk zone to the area of ​​the potential collision risk zone; the speed difference is obtained by the difference between the standardized speed of the vessel and the standardized speed of the target vessel.

[0087] A triangular membership function is applied to the superposition values ​​of channel density, speed difference, and potential energy field to assign membership degrees to different fuzzy levels of channel density, speed difference, and potential energy field, resulting in fuzzy feature vectors. The assignment principle is that the median value of the fuzzy level is the peak membership degree of 1, which linearly decreases to 0 towards the two boundaries.

[0088] The fuzzy feature vectors are compared according to the preset collision risk level rule base to determine the effective rules to be activated, and the final membership degree is determined according to the effective rules to be activated. The collision risk level rule base adopts the IF-THEN logical structure, where IF represents the input condition, the format of which is the same as the fuzzy feature vector, and THEN represents the output result, which is the fuzzy risk level.

[0089] The collision hazard is obtained by defuzzifying the final membership degree. Based on the preset response level correspondence rule, the collision hazard and the corresponding response level signal are obtained.

[0090] First, the channel density is calculated by the ratio of the number of ships to the area within the potential collision risk zone, and the speed difference is calculated by the difference between the standardized values ​​of the ship's speed and the target ship's speed. This adds two key risk factors: channel congestion and speed difference (both of which directly affect the probability of collision; for example, the denser the channel and the greater the speed difference, the higher the collision risk). This solves the problem of traditional assessments relying solely on a single dimension of distance or potential energy values, allowing risk assessments to cover multiple dimensions such as spatial congestion, speed interaction, and potential energy superposition, which is more in line with the complex causes of collision risks in actual navigation.

[0091] Secondly, a triangular membership function (with the middle value being the peak membership value of 1, which linearly decreases to 0 towards both sides) is applied to assign membership degrees to channel density, speed difference, and potential field superposition value to obtain fuzzy feature vectors. Through standardized membership degree calculation rules, the problems of lack of unified standards and subjective membership degree assignment in traditional fuzzification processing are solved, ensuring that the fuzzification results of different types of parameters are comparable, providing a unified basis for subsequent rule matching, and improving the objectivity of risk quantification.

[0092] Furthermore, based on the collision risk level rule base with a preset IF-THEN logical structure (the input format is consistent with the fuzzy feature vector), the fuzzy feature vector is compared, and valid rules with a matching degree > 0 are activated. The final membership degree of each fuzzy risk level is calculated using the maximum membership degree synthesis method. Through the standardized rule base and clear synthesis method, the problems of traditional risk judgment relying on human experience, lack of reasoning rules, and poor consistency of results are solved. This makes the risk reasoning process reproducible and verifiable, ensures the accuracy of valid rule activation in different scenarios, and reduces human subjective error.

[0093] Finally, the final membership degree is defuzzified (by assigning a range of hazard values, calculating the ratio of the sum of the products of 'membership degree × midpoint value of the range' to the sum of membership degrees to determine the hazard level, and then corresponding to the response level), transforming the fuzzy risk level into a quantified collision hazard (such as a specific value from 0 to 100), and matching it with a clear response level signal. This solves the problem that traditional assessments only provide qualitative results (such as high risk), cannot intuitively determine the degree of risk, and cannot guide subsequent specific collision avoidance actions. The quantified results make the magnitude of risk clearer, and the response level signal can directly connect to the generation of subsequent personalized operation strategies and collision avoidance commands, providing a basis for timely and accurate collision avoidance decisions.

[0094] The process of comparing fuzzy feature vectors according to a preset collision risk level rule base, determining the effective rules to be activated, and determining the final membership degree based on the effective rules to be activated is as follows:

[0095] The matching degree between the fuzzy feature vector and each input condition in the collision risk level rule base is calculated, and the IF-THEN rules with a matching degree greater than 0 are identified and recorded as the activated valid rules.

[0096] For the valid activated rules, the final membership degree of each fuzzy risk level is calculated using the maximum membership degree synthesis method, that is, the maximum matching degree of the fuzzy risk level corresponding to all activated rules is taken.

[0097] The matching degree between fuzzy feature vectors and rule base input conditions is calculated to determine valid rules with a matching degree > 0. This solves the problem of traditional risk reasoning relying on manual experience to select rules and lacking clear activation criteria. By quantifying the matching degree (rather than subjective judgment), a matching degree > 0 is used as the activation threshold for valid rules. This ensures that all IF-THEN rules related to the current navigation scenario (channel density, speed difference, potential field superposition value) can be identified, avoiding the one-sidedness of risk reasoning caused by manual omission or misselection of rules. This makes the input rules for risk judgment more comprehensive and provides a foundation for accurate calculation of membership degrees in the future.

[0098] Traditional membership degree synthesis lacks a unified method (such as arbitrarily using the averaging method or weighting method), and the results are unstable and lack physical meaning. The maximum membership degree synthesis method can highlight the risk signal of the rule that best fits the current scenario (that is, the risk level weight corresponding to the rule with the highest matching degree is the largest), avoid irrelevant or weakly related rules from interfering with the final membership degree, make the membership degree result more in line with the actual risk scenario, and provide reliable intermediate data for subsequent defuzzification to obtain accurate collision risk (such as specific values ​​from 0 to 100), ensuring the consistency and reproducibility of risk level judgment under different scenarios.

[0099] The process of defuzzifying the final membership degree to obtain the collision hazard degree is as follows:

[0100] For each fuzzy risk level, assign a clear numerical range of collision hazard degree and determine the midpoint value of the range;

[0101] Determine the sum of the products of the final membership degree of each fuzzy risk level and the midpoint value of the corresponding fuzzy risk level interval, and then determine the sum of the final membership degrees of each fuzzy risk level.

[0102] The ratio of the sum of products to the sum of final membership degrees is used as the collision hazard level.

[0103] First, by assigning clear collision risk value ranges to each fuzzy risk level and determining the midpoint value of the range, the problem of the disconnect between fuzzy risk levels and quantified values ​​in traditional defuzzification is solved. By setting clear numerical boundaries for high / medium / low fuzzy levels (e.g., 80-100 for high risk, 40-80 for medium risk, and 0-40 for low risk) and midpoint values ​​of the ranges (e.g., 90, 60, 20), the originally abstract fuzzy risks have a unified and quantifiable reference standard, avoiding risk judgment bias caused by the fuzzy definition of fuzzy levels (e.g., different people's understanding of high risk).

[0104] Secondly, by determining the sum of the products of the final membership degree and the corresponding midpoint value of each fuzzy risk level, as well as the sum of the final membership degrees of each fuzzy risk level, the problem of traditional defuzzification relying solely on membership degree or a single risk level and ignoring the comprehensive impact of multiple levels is solved. By multiplying the membership degree (the weight of the risk level) with the midpoint value of the interval (the quantitative benchmark of the risk level), the contribution of different fuzzy risk levels can be fully integrated (e.g., a membership degree of 0.6 for high risk and 0.4 for medium risk can simultaneously reflect the impact of both types of risks), avoiding the loss of risk information of other levels due to only taking the highest membership degree level, and making risk calculation more comprehensive.

[0105] Finally, the ratio of the sum of products to the sum of final membership degrees is used as the collision hazard level. This solves the problem that traditional defuzzification results are coarse (e.g., only outputting high risk) and cannot accurately reflect the degree of risk. By using a weighted average, continuous and specific collision hazard values ​​(e.g., 78, 52, etc.) are obtained, rather than discrete levels. This can accurately distinguish the risk differences within the same fuzzy level (e.g., 85 and 95 in high risk, the former is low risk and the latter is extremely high risk). This provides a precise quantitative basis for matching response signals according to the preset response level rules and generating collision avoidance commands that are appropriate for the risk level. It avoids decisions that are too strict (low risk is misjudged as high risk, increasing unnecessary operations) or too lenient (high risk is misjudged as medium risk, delaying the collision avoidance opportunity) due to fuzzy hazard levels.

[0106] Based on the trained LSTM model, personalized operation strategies are determined according to different navigation scenarios, collision risk levels, and corresponding response level signal combinations, and a personalized operation strategy library is established.

[0107] Using historical operational data as input samples and collision hazard level and corresponding response level signals as scenario risk labels, a scenario feature-operation action training dataset is constructed. The scenario features include mud hopper loading, sea state level, and collision hazard response level, while the operation actions include rudder angle adjustment range, rudder angle adjustment frequency, and braking trigger delay.

[0108] The network structure is set as an input layer, a hidden layer and an output layer. The input layer has a feature dimension of 6, the hidden layer has 2 layers with 64 neurons per layer, and the output layer has an output dimension of 3, which correspond to the rudder angle adjustment range, rudder angle adjustment frequency and braking trigger delay, respectively.

[0109] The loss function is the minimum sum of squared errors between the model's output actions and the captain's actual actions, and the model is iterated until the loss value is ≤0.05.

[0110] Based on the trained LSTM model, personalized operation strategies are generated by combining collision risk and corresponding response level signals for different navigation scenarios, thus constructing a personalized driving strategy library.

[0111] By employing feedforward-feedback dual-path control and parallel DQN dynamic optimization of weights, potential collision risk zones and collision hazard levels are assessed. Based on flow field forecasts, advance compensation is provided and rudder angles are corrected in real time to output the final collision avoidance command.

[0112] Using historical operational data as input, collision risk level and response level as scenario risk labels, scenario features include mud hopper loading capacity and sea state level, and operational actions include rudder angle adjustment range, this approach solves the problem that traditional strategy library training data is detached from the characteristics of dredgers and does not associate with risk scenarios. By incorporating scenario features such as mud hopper loading capacity (a core maneuvering influencing factor that distinguishes dredgers from ordinary ships) and sea state level, as well as specific operational actions such as rudder angle adjustment frequency and braking trigger delay, and simultaneously associating risk labels (collision risk level + response level), this approach ensures that the training data can accurately map the correspondence between the actual dredger scenario, risk state, and manual operation, laying a realistic foundation for subsequent personalized strategy generation and avoiding the inapplicability of general ship strategies in dredger scenarios.

[0113] The model features a 6-dimensional input layer, 2 hidden layers with 64 neurons each, and a 3-dimensional output layer corresponding to specific operational actions. This design solves the problems of unstable training results caused by the vague network structure and lack of standardized parameters in traditional models. The clear input dimensions (covering 6 key scene features) ensure that the model can comprehensively capture factors affecting operations. The hidden layer parameters (2 layers with 64 neurons) balance the model's learning ability and computational efficiency. The 3-dimensional output layer accurately corresponds to the rudder angle adjustment range, frequency, and braking delay (the core operational dimensions of collision avoidance for dredgers). This avoids incomplete scene feature capture and large deviations in operation prediction caused by unreasonable structure, ensuring that the model can effectively learn the mapping rules from scene to operation.

[0114] By minimizing the error between the model output and the actual operation of experienced captains, the model is forced to learn human operation logic. The loss threshold (≤0.05) strictly controls the model accuracy, ensuring that the operation actions output by the trained model are highly consistent with the captain's operating habits, greatly reducing the difficulty of human-machine adaptation (the captain does not need to adapt to unfamiliar commands) and avoiding operation delays or misjudgments caused by commands not conforming to human operation logic.

[0115] Based on the diverse actual scenarios of dredgers (such as full hull + sea state 5 + emergency response level, empty hull + sea state 2 + warning response level, etc.), targeted strategies are generated. Subsequent collision avoidance decisions can directly call the operation strategy adapted to the current scenario without temporary calculation, thus improving the efficiency of instruction generation. At the same time, the personalization of the strategy ensures that the operation actions under different risk scenarios are both safe and in line with human habits, avoiding the problem of general strategies being conservative or aggressive in specific scenarios (such as high-risk emergency collision avoidance).

[0116] The process of outputting the final collision avoidance command is as follows:

[0117] Determine the feedforward compensation rudder angle based on flow field prediction of potential collision risk zones;

[0118] The deviation between the current ship position and the safe course is multiplied by the feedback correction gain to obtain the feedback correction rudder angle;

[0119] Based on parallel DQN optimization of feedforward weights and feedback weights, the product of feedforward weights and feedforward compensation rudder angles and the product of feedback weights and feedback correction rudder angles are calculated.

[0120] Obtain the rudder angle adjustment range adapted to the current situation from the personalized operation strategy library, add the rudder angle adjustment range to the two products, and obtain the final rudder angle command;

[0121] Confirm speed adjustment instructions;

[0122] Verify the speed adjustment command and the final rudder angle command. If the nearest encounter distance is greater than the distance threshold and the nearest encounter time is greater than the time threshold, the verification is successful; otherwise, adjust the speed adjustment command and the final rudder angle command.

[0123] The output of the successfully verified speed adjustment command and final rudder angle command is the final collision avoidance command.

[0124] By using flow field forecasting to calculate the rudder angle in advance to offset water flow interference, the problem of traditional collision avoidance commands not considering foreseeable environmental interference and being prone to course deviation due to flow field issues is solved. This allows for proactive prevention, reduces the pressure of subsequent feedback corrections, and enhances the command's anti-interference capabilities.

[0125] By using flow field forecasting to calculate the rudder angle in advance to offset water flow interference, the problem of traditional collision avoidance commands not considering foreseeable environmental interference and being prone to course deviation due to flow field is solved. This allows for prevention before problems occur, reduces the pressure of subsequent feedback correction, and improves the command's anti-interference foundation. By calling upon a strategy library built to fit human habits, the problem of traditional commands being out of the captain's operating logic and having high human-machine adaptation costs is solved. This ensures that rudder angle adjustments conform to human operating feel (such as rudder angle amplitude and frequency), reduces captain's resistance to execution, and improves the efficiency of command implementation.

[0126] The system defines speed adjustment commands to supplement the limitations of relying solely on rudder angle adjustments for collision avoidance. Especially in high-risk scenarios, it extends the TCPA (Time to Closest Encounter) by slowing down, addressing the limitations of traditional collision avoidance methods that are simplistic and lack sufficient safety redundancy, thus further reducing the probability of collisions. The system verifies speed and rudder angle commands (meeting DCPA and TCPA thresholds); if not, adjustments are made. Hard constraint verification addresses the lack of compliance verification and potential safety hazards associated with traditional commands, preventing collision risks or excessive energy consumption after command execution. Successfully verified commands are output, ensuring that the final collision avoidance command simultaneously meets safety, adaptability to human habits, and anti-interference requirements. This solves the problem of traditional commands being directly output without any fallback and prone to execution deviations, providing a reliable basis for collision avoidance operations for trailing suction hopper dredgers.

[0127] The process of determining the feedforward compensation rudder angle based on the flow field prediction of the potential collision risk zone is as follows:

[0128] The flow velocity is obtained from the flow field prediction, and the flow field compensation coefficient and correction coefficient corresponding to the flow velocity are obtained from the database.

[0129] Calculate the ratio of the current velocity magnitude to the reference current velocity, and calculate the angle between the current direction and the ship's course;

[0130] Multiply the flow field compensation coefficient, correction coefficient, ratio, and included angle to obtain the feedforward compensation rudder angle;

[0131] The process of determining the speed adjustment instruction is as follows:

[0132] Obtain the relative distance between the current vessel and the target vessel, and the shortest distance from the current vessel to the boundary of the risk zone;

[0133] When the shortest distance is less than the set shortest threshold, calculate the absolute value of the difference between the target ship's speed and the ship's speed as the relative approximation speed.

[0134] The corresponding deceleration ratio is retrieved from the database based on the shortest distance. The deceleration ratio is then multiplied by the ship's speed to obtain a speed adjustment command. If the verification fails, the deceleration ratio is corrected.

[0135] By storing the velocity-coefficient correspondence in a database, the flow field compensation coefficient and correction coefficient are ensured to accurately match the current flow velocity (e.g., a larger compensation coefficient corresponds to a higher flow velocity), avoiding insufficient compensation (unable to offset strong current interference) or excessive compensation (causing track fluctuations) caused by subjective coefficients, thus laying the foundation for subsequent accurate compensation. The velocity ratio converts the actual flow velocity into a ratio relative to a benchmark (e.g., a benchmark flow velocity of 1 m / s and an actual flow velocity of 2 m / s would result in a ratio of 2), ensuring a unified calculation benchmark for the compensation amplitude under different flow velocity scenarios; the heading angle (e.g., 90° for crossflow and 0° for downstream) quantifies the direction of water flow interference on the ship's heading, providing a basis for the accurate calculation of subsequent compensation direction and amplitude (a larger compensation rudder angle is required for crossflow).

[0136] Through multi-factor coupled calculations, the feedforward compensation rudder angle can dynamically match the flow velocity intensity (ratio), the interference direction (angle), and the scenario adaptability (coefficient), ensuring that the heading deviation caused by the flow field is accurately offset, improving the command's anti-interference capability, and reducing the pressure of subsequent feedback correction. The relative distance (reflecting the degree of approach between the two ships) and the shortest distance to the risk zone (reflecting the distance between itself and the risk boundary) are used as the core basis for speed adjustment, ensuring that the adjustment action conforms to the actual spatial risks and avoiding meaningless energy waste or safety hazards.

[0137] The minimum distance < a set threshold clearly defines the trigger condition for speed adjustment (adjustment only occurs when entering the critical range of the risk zone to avoid premature intervention); the relative approach speed (e.g., a value of 3 if the target ship is 3 knots faster than the ship) quantifies the speed at which the two ships approach each other, providing an urgent basis for calculating the subsequent deceleration magnitude (the faster the approach, the greater the deceleration required). Database matching of deceleration ratios (e.g., a 20% deceleration ratio corresponding to a minimum distance of 0.5 nautical miles) ensures that the ratios are supported by data, avoiding subjective bias; if verification fails, the ratio is corrected (e.g., if verification shows that TCPA is still <15 seconds, the deceleration ratio is increased) to form a closed-loop optimization, ensuring that speed adjustment commands both meet safety constraints (DCPA and TCPA meet standards) and energy-saving requirements (avoiding excessive deceleration).

[0138] The process of optimizing feedforward and feedback weights based on parallel DQN is as follows:

[0139] The state inputs for the parallel DQN are constructed, including the current ship position deviation, collision risk level, risk zone distance, flow field intensity, feedforward compensation rudder angle, and feedback correction rudder angle.

[0140] Parallel DQN sets up two output heads, including feedforward channel weights and feedback channel weights;

[0141] The constraint is that the sum of the feedforward channel weight and the feedback channel weight is 1;

[0142] During the training phase, DQN learns the optimal feedforward and feedback channel weights through a reward function. When validation fails, the feedforward and feedback channel weights are re-optimized. The reward function is the sum of the reward and penalty values, where:

[0143] The reward value for stable track is 2, the reward value for successful collision avoidance is 5, the penalty value for entering the risk zone is -3, and the penalty value for heading oscillation is -2.

[0144] The input factors comprehensively cover track deviation (ship position deviation), risk level (collision hazard level, distance to risk zone), environmental disturbance (flow field intensity), and control status (feedforward / feedback rudder angle), ensuring that the parallel DQN can optimize weights based on the current collision avoidance scenario information. This avoids poor weight adaptability due to missing information (e.g., focusing only on the flow field and ignoring risk, the optimized weights cannot cope with high-risk scenarios), providing a comprehensive basis for accurate weight allocation. The two output heads clearly distinguish the weight allocation of feedforward and feedback, and the weight sum is 1 to ensure reasonable allocation of dual-path control resources (more feedforward means less feedback, and vice versa), avoiding confusion in feedforward-feedback coordination, making the dual-path control logic clear, the correction range controllable, and improving command stability.

[0145] By quantifying rewards and penalties, learning priorities are clearly defined: successful collision avoidance (the core objective) receives the highest reward (+5), while entering the risk zone (serious safety hazard) incurs the heaviest penalty (-3), balancing track stability (+2) and oscillation avoidance (-2). This guides the model to optimize a weight combination that prioritizes safety while also considering stability (e.g., in high-risk scenarios, feedback weights are increased to quickly correct deviations, while in low-risk scenarios, feedforward weights are increased to stabilize the track), ensuring that weight optimization aligns with the core requirements of collision avoidance. A closed loop is formed: weight optimization → command verification → re-optimization if standards are not met, ensuring that the final output feedforward-feedback weights are adaptable to actual navigation scenarios (e.g., if TCPA is found to be insufficient during verification, re-optimization increases the feedback weights to a greater extent of correction), avoiding residual safety risks in collision avoidance commands due to fixed weights, and further improving control robustness.

[0146] An electronic device includes: a processor; and a memory storing computer program instructions that, when executed by the processor, cause the processor to perform the dynamic assessment and collision avoidance decision-making method for the navigation collision risk of a trailing suction hopper dredger as described above.

[0147] A computer-readable storage medium for storing a program that, when executed by a processor, implements the dynamic assessment and collision avoidance decision-making method for navigation collision risk of a trailing suction hopper dredger as described above.

[0148] 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 embodied 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.

[0149] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, 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.

[0150] 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.

[0151] 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.

[0152] 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.

[0153] 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 dynamic assessment and collision avoidance decision-making method for collision risks of trailing suction hopper dredgers, characterized in that, Includes the following steps: Standardize the basic data of the trailing suction hopper dredger's navigation and operation in all dimensions to obtain a standardized dataset. The basic data of the trailing suction hopper dredger's navigation and operation in all dimensions includes the ship's own data, environmental data and personnel operation data. Ship handling characteristics and environmental characteristics parameters are extracted from standardized datasets, and potential collision risk zones are determined based on these parameters. Determine the collision hazard level and corresponding response level signals within the potential collision risk zone: The channel density is obtained by the ratio of the number of vessels in the potential collision risk zone to the area of ​​the potential collision risk zone; the speed difference is obtained by the difference between the standardized speed of the vessel and the standardized speed of the target vessel. A triangular membership function is applied to the superposition values ​​of channel density, speed difference, and potential energy field to assign membership degrees to different fuzzy levels of channel density, speed difference, and potential energy field, resulting in fuzzy feature vectors. The assignment principle is that the median value of the fuzzy level is the peak membership degree of 1, which linearly decreases to 0 towards the two boundaries. The fuzzy feature vectors are compared according to a preset collision risk level rule base to determine the activated valid rules, and the final membership degree is determined based on the activated valid rules. The collision risk level rule base adopts an IF-THEN logical structure, where IF represents the input condition, with the same format as the fuzzy feature vector, and THEN represents the output result, which is the fuzzy risk level. The process of determining the final membership degree is as follows: The matching degree between the fuzzy feature vector and each input condition in the collision risk level rule base is calculated, and the IF-THEN rules with a matching degree greater than 0 are identified and recorded as the activated valid rules. For the valid activated rules, the final membership degree of each fuzzy risk level is calculated using the maximum membership degree synthesis method, that is, the maximum matching degree of the fuzzy risk level corresponding to all activated rules is taken; The final membership degree is defuzzified to obtain the collision hazard level. Based on the preset response level correspondence rule, the collision hazard level and the corresponding response level signal are obtained: For each fuzzy risk level, assign a clear numerical range of collision hazard degree and determine the midpoint value of the range; Determine the sum of the products of the final membership degree of each fuzzy risk level and the midpoint value of the corresponding fuzzy risk level interval, and then determine the sum of the final membership degrees of each fuzzy risk level. The ratio of the sum of products to the sum of final membership degrees is used as the collision hazard level; Based on the trained LSTM model, personalized operation strategies are determined according to different navigation scenarios, collision risk levels and corresponding response level signal combinations, and a personalized operation strategy library is established. By employing feedforward-feedback dual-path control and parallel DQN dynamic optimization of weights, potential collision risk zones and collision hazard levels are identified. Based on flow field prediction, advance compensation is provided and rudder angles are corrected in real time, resulting in the final collision avoidance command. Determining the feedforward compensation rudder angle based on flow field prediction of potential collision risk zones: The flow velocity is obtained from the flow field prediction, and the flow field compensation coefficient and correction coefficient corresponding to the flow velocity are obtained from the database. Calculate the ratio of the current velocity magnitude to the reference current velocity, and calculate the angle between the current direction and the ship's course; Multiply the flow field compensation coefficient, correction coefficient, ratio, and included angle to obtain the feedforward compensation rudder angle; The process of determining the speed adjustment instruction is as follows: Obtain the relative distance between the current vessel and the target vessel, and the shortest distance from the current vessel to the boundary of the risk zone; When the shortest distance is less than the set shortest threshold, calculate the absolute value of the difference between the target ship's speed and the ship's speed, and use it as the relative approximation speed. The corresponding deceleration ratio is obtained from the database based on the shortest distance. The deceleration ratio is multiplied by the ship's speed to obtain the speed adjustment command. If the verification fails, the deceleration ratio is corrected. The deviation between the current ship position and the safe course is multiplied by the feedback correction gain to obtain the feedback correction rudder angle; Based on parallel DQN optimization of feedforward and feedback weights, the product of feedforward weight and feedforward compensated rudder angle and the product of feedback weight and feedback corrected rudder angle are calculated: The state inputs for the parallel DQN are constructed, including the current ship position deviation, collision risk level, risk zone distance, flow field intensity, feedforward compensation rudder angle, and feedback correction rudder angle. Parallel DQN sets up two output heads, including feedforward channel weights and feedback channel weights; The constraint is that the sum of the feedforward channel weight and the feedback channel weight is 1; During the training phase, DQN learns the optimal feedforward and feedback channel weights through a reward function. When validation fails, the feedforward and feedback channel weights are re-optimized. The reward function is the sum of the reward and penalty values, where: The reward for stable track is 2, the reward for successful collision avoidance is 5, the penalty for entering the risk zone is -3, and the penalty for heading oscillation is -2. Obtain the rudder angle adjustment range adapted to the current situation from the personalized operation strategy library, add the rudder angle adjustment range to the two products, and obtain the final rudder angle command; Confirm speed adjustment instructions; Verify the speed adjustment command and the final rudder angle command. If the nearest encounter distance is greater than the distance threshold and the nearest encounter time is greater than the time threshold, the verification is successful; otherwise, adjust the speed adjustment command and the final rudder angle command. The output of the successfully verified speed adjustment command and final rudder angle command is the final collision avoidance command.

2. The method for dynamic assessment and collision avoidance decision-making of navigation collision risk for trailing suction hopper dredgers according to claim 1, characterized in that, The process of determining potential collision risk zones based on ship maneuvering characteristics and environmental characteristics parameters is as follows: Obtain feature coefficient data; Based on the characteristic coefficient data, the potential energy value of each trailing suction hopper dredger is determined by the ship maneuvering characteristics and environmental characteristic parameters. The potential energy value of the ship is superimposed with the potential energy value of the target ship to obtain the potential energy field superposition value. If the potential energy field superposition value is greater than the set potential energy threshold, the area where the ship and the target ship are located is marked as a risk sub-region. The potential collision risk zone is obtained by superimposing all the risk sub-regions.

3. The method for dynamic assessment and collision avoidance decision-making of navigation collision risk for trailing suction hopper dredgers according to claim 2, characterized in that: The ship maneuvering characteristics include standardized values ​​of mud hopper loading and ship speed; the environmental characteristic parameters include standardized values ​​of the encounter distance between the target ship and the ship and the distance between the ship and the target ship; and the characteristic coefficient data include the maneuvering response delay coefficient corresponding to mud hopper loading, the detection error correction coefficient corresponding to visibility, and the COLREGs rule weight. The method for obtaining the potential energy value is as follows: The first potential energy factor is obtained by multiplying the product of the standardized value of the mud hopper loading and the standardized value of the ship speed by the ratio of the standardized value of the distance between the ship and the target ship, and then multiplying the ratio of the product of the product of the mud hopper loading and the standardized value of the ship speed to the ratio of the product ... The second potential energy factor is obtained by multiplying the detection error correction coefficient corresponding to visibility with the COLREGs rule weight and the reciprocal of the normalized value of the encounter distance between the target ship and the ship. The potential energy value is obtained by adding the first potential energy factor and the second potential energy factor.

4. The method for dynamic assessment and collision avoidance decision-making of navigation collision risk for trailing suction hopper dredgers according to claim 1, characterized in that, The process of building a personalized operation strategy library is as follows: Using historical operational data as input samples and collision hazard level and corresponding response level signals as scenario risk labels, a scenario feature-operation action training dataset is constructed. The scenario features include mud hopper loading, sea state level, and collision hazard response level, while the operation actions include rudder angle adjustment range, rudder angle adjustment frequency, and braking trigger delay. The network structure is set as an input layer, a hidden layer and an output layer. The input layer has a feature dimension of 6, the hidden layer has 2 layers with 64 neurons per layer, and the output layer has an output dimension of 3, which correspond to the rudder angle adjustment range, rudder angle adjustment frequency and braking trigger delay, respectively. The loss function is the minimum sum of squared errors between the model's output actions and the captain's actual actions, and the model is iterated until the loss value is ≤0.

05. Based on the trained LSTM model, personalized operation strategies are generated by combining collision risk and corresponding response level signals for different navigation scenarios, thus constructing a personalized driving strategy library.

Images