Oil spill emergency response decision system based on kalman filter

Abstract

This invention discloses an oil spill emergency response decision-making system based on Kalman filtering, belonging to the field of marine environmental management technology. It includes: a risk field generation module, which generates a dynamic risk vector field through nonlinear mapping based on the uncertainty of the Kalman filter prediction output and the oil slick prediction boundary. The dynamic risk vector field contains the probability value of oil slick coverage at each location and the corresponding gradient vector, where the gradient vector represents the direction and rate of probability change; and a command cluster generation module, which generates a flexible command cluster for the vessel to be dispatched based on the gradient vector. The flexible command cluster includes a main path and multiple backup path segments at path bifurcation nodes. This system enables deep coupling of prediction and dispatch, dynamically adapts to the oil slick diffusion situation, generates a flexible command cluster and collaborative decision-making joints, optimizes paths in real time and performs closed-loop parameter tuning, improving the accuracy, timeliness, and reliability of collaborative execution in emergency response.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine environmental management, and more particularly, to an oil spill emergency response decision system based on Kalman filtering. BACKGROUND

[0002] When an oil spill accident occurs at sea, the existing technical solution usually treats pollution diffusion prediction and emergency resource scheduling as two relatively independent links for processing. At the prediction level, state estimation algorithms represented by Kalman filtering can dynamically predict the drift trajectory and diffusion range of the oil film, and the output results can indicate the possibility of the oil film affecting sensitive ecological or economic areas in the future. However, at the resource scheduling and decision support level, most systems fail to deeply integrate such high-dynamic prediction information into a unified and operable decision framework. When the Kalman filtering algorithm predicts that the oil film dynamics has changed, such systems lack an integrated decision mechanism that is tightly coupled with the prediction and can perform real-time re-optimization under multi-objective constraints. These constraints usually include the geographical location of scattered resource points, the diverse performance parameters of various emergency equipment such as oil containment boom ships and oil spill agent spraying devices, the demanding emergency arrival time window, and the multiple influences of real-time changing sea conditions such as sea currents and wind speed and direction on navigation and operation efficiency.

[0003] Due to the fragmentation of the prediction and scheduling links, the system cannot automatically and dynamically calculate the optimal or near-optimal dispatching scheme in comprehensive consideration of multiple dimensions such as time, cost, and efficiency based on the latest prediction data. Therefore, although advanced prediction techniques in the existing technology can generate dynamic risk intelligence that evolves over time, subsequent resource allocation and path planning decisions fail to respond to this dynamicity synchronously, still relying on relatively static models or artificial experience judgments. This leads to the possibility that the final generated emergency response scheme may not adapt to the latest situation development, resulting in obvious room for improvement in the timeliness, economy, and risk control effect of resource deployment. SUMMARY

[0004] In view of the problems existing in the prior art, the purpose of the present application is to provide an oil spill emergency response decision system based on Kalman filtering, which can realize deep coupling of prediction and scheduling, dynamically adapt to the oil film diffusion situation, generate flexible instruction clusters and collaborative decision joints, optimize the path in real time and close the loop, and improve the accuracy, timeliness, and collaborative execution reliability of emergency response, effectively avoiding the problems of resource deployment lag and risk prediction deviation.

[0005] To solve the above problems, the present application adopts the following technical solution:

[0006] The oil spill emergency response decision system based on Kalman filtering comprises:

[0007] a risk field generation module, based on the uncertainty of Kalman filter prediction output and the oil film prediction boundary, generates a dynamic risk vector field through nonlinear mapping, the dynamic risk vector field contains the probability value of each position being covered by the oil film and the corresponding gradient vector, the gradient vector represents the direction and rate of change of the probability;

[0008] an instruction cluster generation module, based on the gradient vector, generates a flexible instruction cluster for the ship to be dispatched, the flexible instruction cluster contains a main path and multiple standby path segments at path branching nodes;

[0009] a joint identification module, for identifying the intersection nodes of all flexible instruction clusters, and defining them as cooperative decision joints;

[0010] a sensitivity table module, for identifying whether each cooperative decision joint needs to be re-evaluated due to prediction changes according to whether the change of the gradient vector exceeds a preset threshold when the prediction is updated, and generating a sensitivity table;

[0011] a strategy tree update module, for triggering re-evaluation of the standby path segments after the cooperative decision joint when the ship approaches the joint identified by the sensitivity table as needing re-evaluation, and updating the strategy tree containing the current execution instruction and the standby instruction corresponding to each standby path segment according to the re-evaluation result;

[0012] a parameter adjustment module, for matching the actual navigation trajectory of the ship with the historical standby instructions, and adjusting the parameter value of the nonlinear mapping according to the matching case when the actual trajectory matches a standby instruction and the corresponding predicted scenario is verified.

[0013] Further, the risk field generation module comprises:

[0014] extracting the eigenvalues and eigenvectors of the covariance matrix of the Kalman filter output, and performing probability wave front propagation simulation from the oil film centroid as the starting point, along the direction indicated by the eigenvectors and with the momentum quantified by the eigenvalues, to generate a probability density evolution field;

[0015] based on the sea current vector field data, performing vector synthesis and probability density correction on the probability density evolution field to generate a physically corrected probability field;

[0016] calculating the gradient of the physically corrected probability field, identifying the ridge line in the physically corrected probability field, and performing normal projection reinforcement of the gradient vector towards the ridge line to generate a dynamic risk vector field;

[0017] based on the wind field data, setting a dynamic threshold for the modulus of the dynamic risk vector field, and calculating the dynamic threshold through a nonlinear sensitivity function according to the angle between the wind speed, wind direction and the oil film expansion principal axis.

[0018] Further, the instruction cluster generation module comprises:

[0019] Numerically integrating the dynamic risk vector field from the start position to generate the main path, wherein the integration step length is adjusted according to the modulus of the gradient vector, and a weak gravity item pointing to the task target point is introduced;

[0020] Along the main path, the orthogonal component of the gradient vector perpendicular to the tangent direction of the main path is calculated, and the position where the modulus of the orthogonal component exceeds the threshold value is marked as a branch budding point; a short-range exploration path is generated by performing a finite step field integration in the positive and negative directions of the orthogonal component of each budding point;

[0021] Based on the parameters representing the maneuvering characteristics of the ship, the instantaneous maneuverability envelope of the main path and the short-range exploration path is calculated, and the feasibility of the path is judged and pruned according to the envelope;

[0022] Project all pruned path segments into the space-time coordinate system, detect the space-time intersection between path segments, eliminate conflicts by adjusting the parameters of conflicting path segments, and organize the main path of each ship and its associated backup path segments into a flexible instruction cluster of topological network structure.

[0023] Further, the joint identification module comprises:

[0024] Discretize the path segments of the topological network structure into time-stamped space-time waypoints, generate space-time occupancy voxels based on the space-time waypoints, and fuse all voxels to construct a space-time trajectory body, which represents the occupation frequency in the form of voxel density;

[0025] Perform four-dimensional morphological analysis on the density field of the space-time trajectory body, extract the region with density exceeding the threshold value and space-time connectivity as the collection body, and extract the space-time morphological features of the collection body.

[0026] Further, the joint identification module further comprises:

[0027] Couple the space-time range of the collection body with the dynamic risk vector field, calculate the average modulus and direction consistency of the risk vector field within the range, and combine the proximity relationship with the ridge line in the dynamic risk vector field to assign a decision urgency weight to the collection body;

[0028] Based on the communication link model and resource loading location information, verify the communication connection and resource delivery feasibility of the collection body, and define the space-time coordinates of the collection body whose decision urgency weight is greater than the preset threshold value as the cooperative decision joint.

[0029] Further, the sensitivity table module comprises:

[0030] maintain a historical sequence of dynamic risk vector fields, perform point-by-point vector difference between the current dynamic risk vector field and the historical dynamic risk vector field to obtain an instantaneous difference field, perform asymmetric convolution operation on the instantaneous difference field to synthesize a field difference perturbation field;

[0031] According to the spatiotemporal morphological characteristics of the collaborative decision joint, reconstruct its influence domain, and calculate the spatial integral of the average cosine of the field difference perturbation field vector and the original dynamic risk vector field and the perturbation vector module length in the influence domain, and synthesize them into a perturbation penetration index.

[0032] Further, the sensitivity table module further includes:

[0033] Analyze the scheduling decision network topology composed of flexible instruction clusters and collaborative decision joints, identify the logical dependency relationship between joints, and calculate the topological dependency vulnerability coefficient based on the betweenness centrality of each joint in the network and the number of downstream pending branch paths.

[0034] Nonlinearly weight and fuse the perturbation penetration index and the topological dependency vulnerability coefficient, introduce a decay factor based on the current pose of the ship and the expected arrival time difference of each joint, generate a re-evaluation priority score for each joint, and divide the sensitivity level and generate a dynamic sensitivity mapping table accordingly.

[0035] Further, the strategy tree updating module includes:

[0036] Based on the current pose, velocity vector of the ship, and the spatiotemporal morphological characteristics of the collaborative decision joint, dynamically define a spatiotemporal cone extending from the current position of the ship to the adjacent region of the spatiotemporal coordinates of the joint as a restricted re-evaluation domain.

[0037] Take the historical evaluation state of the standby path segment in the restricted re-evaluation domain as the reference, combine the cumulative change of the dynamic risk vector field in the domain, and perform incremental field integral correction and spatiotemporal conflict rehearsal.

[0038] If the joint involves multiple ships, perform non-dominated negotiation and merging of the incremental correction scheme set of each ship based on the Pareto frontier search to generate a joint path correction scheme package.

[0039] Mount the joint path correction scheme package as a shadow branch to the corresponding node of the strategy tree, adjust the heading or speed parameter of the current execution instruction to link the shadow branch, and activate the shadow branch after passing through the preset switching point to update the strategy tree.

[0040] Further, the parameter adjustment module includes:

[0041] The similarity of the actual sailing trajectory and the spatio-temporal form feature of the historical preparation instruction is calculated, the feature includes the trajectory curvature change, the sequence and time difference of passing through the geographic marker point, and the action mode of the dynamic risk vector field experienced by the trajectory, when the comprehensive score exceeds the confidence threshold and the uniqueness verification passes, it is determined as a confidence matching case, and the historical preparation instruction is derived from the strategy tree;

[0042] For the confidence matching case, the dynamic risk vector field sequence once predicted in the corresponding spatio-temporal window is extracted, and the apparent risk field is back calculated according to the actual trajectory, and the prediction observation deviation field is generated by comparing the two.

[0043] Further, the parameter adjustment module further comprises:

[0044] The prediction observation deviation field mode is matched with the preset disturbance direction, the non-linear mapping parameter is disturbed in the matching direction to generate a candidate parameter;

[0045] The confidence matching cases and their deviation modes and parameter disturbance records are accumulated to form a case cluster, the results of the historical disturbance on the case cluster are analyzed, and the current parameter vector is iteratively relaxed and updated in the direction of tending to reduce the prediction deviation of the case cluster corresponding type.

[0046] Compared with the prior art, the beneficial effects of the present application are that:

[0047] (1) The present scheme realizes deep coupling of Kalman filter output and dynamic risk vector field, converts oil film prediction uncertainty into quantified risk features, optimizes risk threshold combined with real-time sea conditions such as wind field and sea current, solves the problem of separation of prediction and scheduling in the prior art, makes emergency decision accurately match the oil film dynamic diffusion situation, improves the timeliness and risk control effect of resource deployment, and reduces the pollution risk of sensitive areas.

[0048] (2) The present scheme constructs a topological network type flexible instruction cluster, extracts and generates a plurality of sets of alternative schemes through branch bud points and short-range exploration paths, realizes collaborative planning and conflict resolution of multi-ship paths combined with priority judgment and feasibility verification of collaborative decision joints, avoids response lag caused by single path failure, and improves the flexibility and execution reliability of emergency scheduling.

[0049] (3) The present scheme dynamically evaluates the influence of the decision joint on the prediction change based on the sensitivity table, realizes incremental correction of the path through the limited re-evaluation domain, completes multi-ship scheme negotiation combined with the Pareto frontier search, and updates the strategy tree through shadow branch smooth switching, avoids the efficiency loss of full path recalculation, ensures that the decision adapts to the latest risk situation in real time, and improves the accuracy and efficiency of emergency response.

[0050] (4) This scheme reverses the apparent risk field and generates a prediction bias field by matching the confidence of the actual navigation trajectory with the historical pre-instructions. It optimizes the nonlinear mapping parameters by directional perturbation and combines iterative relaxation updates of case clusters to form a mechanism from prediction to execution, and then to verification and optimization, thereby continuously improving the accuracy of risk prediction and strengthening the system's adaptability to complex sea conditions and oil spill scenarios in the long term. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0052] Figure 1 This is a flowchart of the various modules of the present invention and the data flow between them. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 The oil spill emergency response decision system based on Kalman filtering includes a risk field generation module. Based on the uncertainty of the Kalman filter prediction output and the oil film prediction boundary, a dynamic risk vector field is generated through nonlinear mapping. The dynamic risk vector field contains the probability value of each location being covered by oil film and the corresponding gradient vector. The gradient vector represents the direction and rate of the probability change.

[0055] The risk field generation module is also used to perform the following specific operations:

[0056] S11, extract the eigenvalues ​​and eigenvectors of the covariance matrix output by the Kalman filter, and perform probability wavefront propagation simulation along the direction indicated by the eigenvectors and with the momentum quantized by the eigenvalues, starting from the oil film centroid, to generate a probability density evolution field. The specific operations are as follows:

[0057] First, the eigenvalues ​​and eigenvectors of the covariance matrix in the Kalman filter output are extracted. This covariance matrix directly reflects the uncertainty distribution characteristics of the oil film prediction state. Its eigenvectors can indicate the dominant direction of uncertainty diffusion, while the eigenvalues ​​can quantify the diffusion intensity of uncertainty in that direction. The oil film centroid is used as the starting reference point for the probability wavefront propagation simulation because the oil film centroid can accurately characterize the overall spatial position and movement trend of the oil film, providing a stable initial anchor point for subsequent probability evolution. During the propagation simulation, along the dominant direction indicated by the eigenvectors, the momentum quantified by the eigenvalues ​​is used as the propagation driving force parameter. Through the gradual diffusion of the probability wavefront, a probability density evolution field that can dynamically reflect the probability of oil film coverage at different spatiotemporal locations is generated. This evolution field not only contains the probability values ​​of oil film coverage at each location, but also reflects the evolution law of probability over time.

[0058] S12, Based on ocean current vector field data, perform vector synthesis and probability density correction on the probability density evolution field to generate a physically corrected probability field. The specific operations are as follows:

[0059] After acquiring the probability density evolution field, the risk field generation module performs vector synthesis and probability density correction operations on the evolution field based on real-time collected ocean current vector field data, ultimately generating a physically corrected probability field. Ocean currents, as the core physical factor affecting oil film drift and diffusion, have vector field data containing the direction and velocity of the current, directly determining the actual trajectory and diffusion range of the oil film in the marine environment. During vector synthesis, the probability evolution vectors at each location in the probability density evolution field are coupled with the corresponding ocean current vectors to ensure that the probability evolution trend is consistent with the ocean current flow characteristics, correcting the probability distribution deviation caused by relying solely on mathematical predictions. The probability density correction combines the driving effect of ocean currents on oil film diffusion, adjusting the probability density values ​​at each spatiotemporal location. For example, in areas with higher ocean current velocities, the oil film diffuses faster, and the probability density decay or growth rate at the corresponding location will be adjusted accordingly. In ocean current vortex areas, oil film is prone to aggregation, and the probability density in the corresponding area will be strengthened and corrected. Through the above physical correction operations, the probability field is made more consistent with the physical characteristics of the actual marine environment, improving the accuracy and reliability of oil film coverage probability prediction.

[0060] S13, calculate the gradient of the physical correction probability field, identify the ridges in the physical correction probability field, strengthen the gradient vector by normal projection towards the ridges, and generate a dynamic risk vector field. The specific operations are as follows:

[0061] The risk field generation module performs gradient calculations on the physically modified probability field to obtain gradient vectors representing the changes in probability values ​​at each location. The direction of these gradient vectors characterizes the dominant direction of probability value changes, while the magnitude corresponds to the rate of change of probability values. Subsequently, a field analysis algorithm identifies ridges in the physically modified probability field. As regions with the most significant changes in probability density, ridges are key boundaries or core paths for oil film diffusion. Their location and shape directly reflect the key areas and trends of oil film diffusion. To enhance the indicative role of gradient vectors in risk changes, the risk field generation module performs normal projection enhancement processing on the gradient vectors towards the ridges. That is, through vector projection operations, the gradient vectors at each location are made to better fit the normal direction of the ridges, highlighting the probability change characteristics along the boundary of the core risk region. The gradient vectors enhanced by projection, together with the probability values ​​in the physically modified probability field, constitute a dynamic risk vector field. This vector field can accurately reflect the probability of each location being covered by oil film, as well as the change pattern of probability values ​​along the dominant risk direction. This provides a clear risk guidance basis for the subsequent generation of flexible command clusters, ensuring that command generation can adapt to the risk characteristics of dynamic oil film diffusion.

[0062] S14, Based on wind field data, a dynamic threshold is set for the magnitude of the dynamic risk vector field. The dynamic threshold is calculated using a nonlinear sensitivity function based on the angle between wind speed, wind direction, and the principal axis of oil film expansion. The specific operation is as follows:

[0063] Based on real-time acquired wind field data, the risk field generation module sets a dynamic threshold for the magnitude of the dynamic risk vector field. This threshold is used to define the range of vector magnitudes for different risk levels. As an important environmental factor affecting oil film expansion, the wind speed directly affects the expansion rate of the oil film. The angle between the wind direction and the main axis of oil film expansion determines the driving efficiency of the wind field on oil film expansion. When the angle is small, the wind field has a stronger driving effect on oil film expansion, and the oil film risk spreads faster. When the angle is large, the wind field is more likely to block or deflect the oil film expansion, and the risk diffusion trend will be weakened accordingly.

[0064] The risk field generation module calculates the dynamic threshold using a nonlinear sensitivity function. This function adapts to the current technical scenario using an existing nonlinear mathematical model and can select the appropriate function type based on sea state characteristics: when it is necessary to enhance the driving response of wind speed to the threshold, an exponential function is used, with wind speed as the independent variable. By setting a reasonable base, such as 1.2 to 1.5, the threshold increases exponentially with increasing wind speed, quickly adapting to scenarios where risk spreads rapidly under high wind speeds; when it is necessary to weaken the threshold fluctuations in the low wind speed range, a power function, such as a second or third power function, is used. By adjusting the power coefficient, such as 1.5 to 2.5, the threshold increases slowly at low wind speeds and gradually increases at high wind speeds, balancing the threshold adaptation accuracy across different wind speed ranges; when the wind speed reaches a preset critical value, the oil film diffusion tends to... When saturation occurs, a saturation function of the Sigmoid type is selected. By setting a saturation range, such as an upper threshold of 0.8 to 0.9, the threshold increase tends to be gradual after the wind speed exceeds the critical value, avoiding over-judgment of risk areas due to excessively high thresholds at high wind speeds. When it is necessary to adapt to the differences in driving efficiency between different angle ranges, a piecewise nonlinear function is selected. The angle is divided into a low angle range (0° to 30°), a medium angle range (30° to 60°), and a high angle range (60° to 90°). A larger function slope is set in the low angle range, such as 0.05 to 0.08, so that the threshold increases rapidly as the angle decreases. A smaller function slope is set in the high angle range, such as 0.01 to 0.03, so that the threshold decreases rapidly as the angle increases, accurately matching the driving efficiency of the wind field under different angles.

[0065] The input parameters of the aforementioned nonlinear sensitivity function are uniformly wind speed, the angle between wind direction and the main axis of oil film expansion, and the output parameter is a dynamic threshold. By adjusting the core parameters of the function, such as the base of the exponential function, the interval threshold of the piecewise function, and the saturation coefficient of the sigmoid function, the threshold change is made to closely match the actual impact of sea state on oil film diffusion: when the wind speed increases and the angle between the wind direction and the main axis of oil film expansion is small, the function outputs a higher dynamic threshold to adapt to the scenario of rapid risk diffusion and avoid the proliferation of high-risk areas due to low threshold; when the wind speed is low or the angle is large, the function outputs a correspondingly lower dynamic threshold to avoid over-judging the risk range and ensure that medium-risk areas are not missed. The setting of the dynamic threshold gives the magnitude of the risk vector field a clear risk classification meaning, which can accurately match the oil film risk change characteristics under real-time wind field conditions.

[0066] In some embodiments of the present invention, an instruction cluster generation module is also included, which generates a flexible instruction cluster for the ship to be scheduled based on the gradient vector. The flexible instruction cluster includes a main path and multiple backup path segments at path bifurcation nodes.

[0067] The instruction cluster generation module is also used to perform the following specific operations:

[0068] S21, starting from the initial position, the main path is generated by numerical integration of the dynamic risk vector field. The integration step size is adjusted according to the magnitude of the gradient vector, and a weak gravity term pointing towards the task target point is introduced. The specific operation is as follows:

[0069] The instruction cluster generation module uses the ship scheduling starting position, such as an emergency port or resource deployment point, as the integration starting point. It performs numerical integration on the dynamic risk vector field to generate the main navigation path for the ship. The logic of numerical integration is to advance the path generation along the dominant direction of risk change indicated by the gradient vector, so that the main path can accurately match the diffusion trend of oil film risk, prioritizing the avoidance of high-probability coverage areas or pointing to the core treatment area. The integration step size adopts a dynamic adjustment mechanism, and its value is negatively correlated with the magnitude of the gradient vector: when the magnitude of the gradient vector is large, it indicates that the rate of change of the probability of oil film coverage in that area is faster and the risk is more dynamic. At this point, the integral step size is reduced to improve the accuracy of the path in capturing subtle changes in risk. When the gradient vector magnitude is small, the regional risk changes are gradual, and the magnitude can be appropriately increased to improve the path generation efficiency. To prevent the main path from deviating from the core objective of the emergency mission due to simply following the risk vector field, the instruction cluster generation module introduces a weak gravitational term pointing to the mission objective points such as the oil spill centroid and the protection boundary of the sensitive ecological area. This gravitational term fine-tunes the path direction through a weak vector pull, ensuring that the main path always points towards the key points of mission handling without interfering with the core logic of risk orientation, thus achieving a dynamic balance between risk avoidance and mission objectives.

[0070] S22, along the main path, calculate the orthogonal components of the gradient vector at each point that are perpendicular to the tangent direction of the main path. Mark the points where the magnitude of the orthogonal components exceeds a threshold as branch budding points. Perform finite step field integration from each budding point along the positive and negative directions of its orthogonal components to generate a short-range exploration path. The specific operations are as follows:

[0071] After generating the main path, the instruction cluster generation module calculates the orthogonal component of the gradient vector in the direction perpendicular to the tangent of the main path at each spatial node along the main path. This orthogonal component represents the lateral risk change characteristics of the main path, and its magnitude directly reflects the fluctuation intensity of lateral risk. The instruction cluster generation module presets a threshold λ for the determination of the orthogonal component magnitude. When the magnitude of the orthogonal component of a node exceeds λ, it indicates that there is a significant lateral risk change at that location, such as a potential oil film diffusion branch or a sudden change in risk gradient, and it is marked as a branch budding point. For each branch budding point, the instruction cluster generation module performs a finite-step dynamic risk vector field numerical integration along the positive and negative directions of the orthogonal component to generate a short-range exploration path: the positive integration corresponds to the direction of increasing lateral risk, and the negative integration corresponds to the direction of decreasing lateral risk. The setting of the finite step size can control the length of the exploration path, avoid excessive extension leading to waste of computational resources, and ensure that the path covers the key lateral risk areas. The core function of the short-range exploration path is to provide the ship with alternative navigation schemes outside the main path to cope with the path adjustment needs that may be brought about by subsequent changes in risk prediction.

[0072] S23, Based on parameters characterizing ship maneuverability, calculate instantaneous maneuverability envelopes for the main path and short-range exploration path, and perform feasibility assessment and pruning of the paths according to the envelopes. The specific operations are as follows:

[0073] The instruction cluster generation module extracts the maneuverability parameters of the vessel to be scheduled, including maximum turning angle, minimum turning radius, maximum speed, and acceleration or deceleration performance thresholds. These parameters directly determine the vessel's operational boundaries under different navigation conditions. Based on these parameters, the instantaneous maneuverability envelope is calculated node-by-node for the generated main path and each short-range exploration path. This envelope represents the boundary of the range of maneuvers the vessel can perform at the corresponding node based on its current speed and heading. For example, the maneuverability envelope of a certain node limits the maximum allowable value of the vessel's turning angle and the adjustable range of its speed. Subsequently, the instantaneous maneuverability envelope is used as the criterion for... All paths undergo feasibility verification and pruning: If the turning angle of a path segment exceeds the ship's maximum turning angle, the turning radius is less than the minimum turning radius, or the speed requirement exceeds the ship's performance threshold, then the path segment is determined to be an infeasible path that the ship cannot execute, and it is removed; For nodes in the path that locally exceed the maneuver envelope, if the path can be brought back to the feasible range by fine-tuning the speed and heading of adjacent nodes, then local correction is performed without overall removal. Through maneuverability verification and pruning, it is ensured that the retained path segments all conform to the ship's actual navigation capabilities, avoiding the generation of theoretically feasible but practically unexecutable invalid paths.

[0074] S24: Project all pruned path segments onto the spatiotemporal coordinate system, detect the spatiotemporal intersection between path segments, resolve conflicts by adjusting the parameters of conflicting path segments, and organize the main path of each ship and its associated backup path segments into a flexible instruction cluster with a topological network structure. The specific operations are as follows:

[0075] The command cluster generation module synchronously projects the pruned main path and all short-range exploration path segments into a spatiotemporal coordinate system. This coordinate system includes spatial dimensions, such as latitude and longitude and altitude, as well as temporal dimensions, such as sailing time, which can accurately characterize the spatiotemporal occupancy features of the path segments. Through a spatiotemporal intersection detection algorithm, it traverses the spatiotemporal coordinate intervals of all path segments to identify path segments with spatiotemporal overlap, i.e., conflict scenarios where multiple ships or different alternative paths of the same ship occupy the same spatial area at the same time. For the detected spatiotemporal conflicts, the command cluster generation module resolves the conflicts by adjusting the parameters of the conflicting path segments: if it is a conflict between different alternative paths of the same ship, the start time or speed of one of the path segments can be fine-tuned. The parameters are staggered to avoid overlapping time and space occupancy intervals. If there is a path conflict between different ships, the spatial trajectory of the low-priority ship's path can be adjusted based on the ship's mission priority and path risk adaptability, such as by slight offsets or by delaying departures, to ensure that each path segment does not interfere with each other in the time and space dimension. After the conflict is resolved, the instruction cluster generation module takes the main path of each ship as the core link and its associated short-range exploration path segment as the backup link. At the path bifurcation node, that is, at the branch budding point, a connection is established with the main path to construct a flexible instruction cluster with a topological network structure. This topological structure can clearly show the relationship between the main path and the backup path and the switching node, which is convenient for subsequent dynamic adjustment of navigation instructions based on risk prediction.

[0076] In some embodiments of the present invention, a joint recognition module is also included, which is used to identify the intersection nodes of all flexible instruction clusters and define them as collaborative decision joints;

[0077] The joint recognition module is also used to perform the following specific operations:

[0078] S31, the path segments of the topological network structure are discretized into spatiotemporal waypoints with timestamps. Spatiotemporal occupancy voxels are generated based on the spatiotemporal waypoints, and all voxels are fused to construct a spatiotemporal trajectory volume. The spatiotemporal trajectory volume uses voxel density to represent the occupancy frequency. The specific operations are as follows:

[0079] The joint recognition module first discretizes the path segments of the topological network structure of each flexible command cluster. Each path segment is decomposed into a series of time-stamped spatiotemporal waypoints according to preset spatial and time intervals. The timestamps precisely correspond to the estimated time of the ship's arrival at the waypoint, while the spatial coordinates specify the latitude, longitude, and altitude information of the waypoint. Through discretization, continuous paths can be transformed into a set of discrete nodes that can be quantified and analyzed. Based on the generated spatiotemporal waypoints, the joint recognition module constructs spatiotemporal occupancy voxels. Each voxel corresponds to a fixed spatiotemporal grid unit, and its spatiotemporal range is determined by discrete... The interval parameters are defined, for example, the side length of the spatial grid is set to 50 meters and the time grid interval is set to 5 minutes. When a spatiotemporal waypoint falls into the spatiotemporal range of a certain voxel, it is marked that the voxel is occupied. Then, the joint recognition module merges the spatiotemporally occupied voxels corresponding to all path segments to construct a spatiotemporal trajectory volume. The trajectory volume uses voxel density as the core characterization parameter. The value of voxel density is equal to the total number of spatiotemporal waypoints falling into the voxel, which represents the frequency of the spatiotemporal region being occupied by different path segments. The higher the frequency, the more likely the region is to become the intersection area of ​​multiple flexible command clusters.

[0080] S32, Perform four-dimensional morphological analysis on the density field of the spatiotemporal trajectory volume, extract regions with density exceeding a threshold and spatiotemporal connectivity as aggregates, and extract the spatiotemporal morphological features of the aggregates. The specific operations are as follows:

[0081] The joint recognition module performs four-dimensional morphological analysis on the constructed spatiotemporal trajectory volume. This four-dimensional analysis includes three spatial dimensions (latitude, longitude, and altitude) and one temporal dimension (flight time). Morphological analysis accurately identifies the clustering characteristics of density distribution within the trajectory volume. First, a voxel density threshold is set. This threshold is calculated by statistically analyzing the mean and standard deviation of the density of all voxels in the spatiotemporal trajectory volume. For example, the threshold is set to the mean density plus 1.2 times the standard deviation. This filters out high-occupancy voxels whose density exceeds the threshold. Then, spatiotemporal connectivity is checked on these high-occupancy voxels to determine if the spatial distance between adjacent voxels is less than a preset threshold. The system extracts continuous voxel regions that meet spatiotemporal connectivity by considering factors such as the radius (e.g., 100 meters) and the time interval (e.g., 10 minutes). These regions are essentially spatiotemporal convergence areas of multiple flexible instruction cluster path segments. After extracting the confluence, the joint recognition module further extracts its spatiotemporal morphological features, including the spatiotemporal range of the confluence (i.e., the latitude and longitude span in space and the time interval of navigation); the central spatiotemporal coordinates (i.e., the geometric center of the region and the average occupancy time); the density peak position (i.e., the spatiotemporal point with the highest voxel density); and the density gradient distribution (i.e., the density decay rate from the peak to the edge).

[0082] S33, the spatiotemporal range of the aggregate is coupled with the dynamic risk vector field, the average magnitude and directional consistency of the risk vector field within the range are calculated, and the aggregate is assigned a decision urgency weight based on its proximity to the ridge line in the dynamic risk vector field. The specific operation is as follows:

[0083] To quantify the emergency decision-making priorities of a cluster, the joint recognition module spatiotemporally couples the spatiotemporal range of the cluster with the dynamic risk vector field. This means precisely mapping the spatiotemporal coordinates of the cluster onto the dynamic risk vector field, limiting the scope of risk feature analysis to the spatiotemporal region covered by the cluster. Within this coupling range, the joint recognition module calculates two risk feature parameters: first, the average magnitude of the risk vector field, obtained by calculating the arithmetic mean of the risk vector magnitudes corresponding to all voxels within the region. A larger average magnitude indicates a faster rate of change in the oil film coverage probability of the cluster area, and stronger risk dynamism; second, the directional consistency of the risk vectors, calculated by determining the average angle between all risk vectors within the region. The cosine value indicates that the closer the directional consistency is to 1, the more concentrated the risk diffusion direction is within the region, and the clearer the risk evolution trend. At the same time, the joint recognition module calculates the shortest spatiotemporal distance between the spatiotemporal coordinates of the aggregation center and the ridge line of the dynamic risk vector field. The closer the distance, the closer the aggregation is to the core risk area of ​​oil film diffusion, and the higher the degree of risk impact. Based on the above parameters, the joint recognition module assigns a decision urgency weight to each aggregation using a weighted summation method. The weight coefficient of average modulus is set to 0.5, the weight coefficient of directional consistency is set to 0.3, and the weight coefficient of ridge line proximity is set to 0.2. The higher the weight value, the more the intersection area corresponding to the aggregation needs to be given priority for emergency decision consideration.

[0084] S34, based on the communication link model and resource loading location information, verifies the feasibility of communication connections and resource delivery for the aggregate. The spatiotemporal coordinates of aggregates that pass verification and whose decision urgency weight is greater than a preset threshold are defined as collaborative decision-making joints. The specific operations are as follows:

[0085] The joint recognition module, based on a pre-set communication link model and real-time acquired resource loading location information, performs dual verification of the feasibility of communication connections and resource delivery for the extracted aggregates. This ensures that subsequent collaborative decision-making joints can meet the actual execution requirements of emergency response. The pre-set communication link model is constructed and trained through a process of feature modeling, data training, and verification optimization to adapt to the complex communication environment of marine oil spill emergency scenarios. The specific construction and training logic is as follows:

[0086] In the model building phase, the spatiotemporal distance between the confluence and the communication base station or emergency command center, real-time sea state parameters such as wind speed, wave height, and sea fog concentration, communication equipment parameters such as transmission power, antenna gain, and operating frequency band, and geographical obstruction conditions such as the distribution of islands and reefs are used as input features. Communication signal strength, transmission delay, and anti-interference capability are used as output indicators. A nonlinear mapping model is constructed using the Gradient Boosting Tree (GBRT) algorithm. This model fits the complex correlation between input features and communication performance indicators through ensemble learning of multiple decision trees, avoiding prediction biases of a single model in a dynamic maritime environment. In the model training phase, measured communication data under different sea states, distances, and equipment configurations are first collected, covering indicators such as signal strength and transmission delay. Simultaneously, extreme sea states, such as high winds and waves, are generated through simulation. Supplementary datasets based on sea fog conditions were used to avoid missing scenarios in the actual measurements. The datasets were preprocessed to remove outliers, such as extreme delays caused by sudden interference, and features were normalized and encoded, for example, sea fog concentration was divided into 5 levels and quantized. The preprocessed datasets were then divided into training and testing sets in a 7:3 ratio. The model was trained using the training set data, and hyperparameters were optimized using a grid search method, such as setting the learning rate to 0.05 to 0.1, tree depth to 3 to 5, and the number of decision trees to 100 to 150, minimizing the mean squared error between predicted and measured values. The model accuracy was verified using the testing set data, requiring signal strength prediction error to be no more than 3 dBm and transmission delay prediction error to be no more than 50 milliseconds. If the accuracy requirements were not met, additional actual measurement data for the corresponding scenario was added, and the model was retrained until the accuracy met the standards.

[0087] In the communication connection feasibility verification, the communication link model trained above is used as input. The spatiotemporal coordinates of the aggregation area, real-time sea state data, communication equipment configuration parameters, and geographical obstruction information are used to calculate the communication signal strength, transmission delay, and anti-interference capability of the area. This determines whether the emergency communication requirements are met. For example, the communication signal strength must be no less than 85 dBm, the transmission delay no more than 500 milliseconds, and the anti-interference capability score must be no less than 0.7 (out of 1). This ensures the smooth transmission of decision-making instructions and real-time data feedback. In the resource delivery feasibility verification, the resource loading location, such as emergency equipment storage ports, material reserve points, and the aggregation area, is considered. The spatiotemporal coordinates are used to calculate the shortest travel time and path accessibility for emergency resources to reach the convergence area, and to determine whether resource delivery can be completed within the emergency time window. For example, the shortest travel time is required to be no more than 2 hours and the path must be free of navigation obstacles, such as shoals or reefs. For convergences that pass the dual feasibility verification and whose decision urgency weight is greater than the preset threshold of 0.6, the joint recognition module extracts its central spatiotemporal coordinates as the core coordinates of the collaborative decision joint, and formally defines the spatiotemporal region corresponding to the convergence as the collaborative decision joint. Such joints are not only the intersection nodes of multiple flexible command clusters, but also key areas with emergency decision priority and execution feasibility.

[0088] In some embodiments of the present invention, a sensitivity table module is also included, which is used to identify whether each collaborative decision joint needs to be re-evaluated due to the change in the prediction when the prediction is updated, based on whether the change in the gradient vector exceeds a preset threshold, and generate a sensitivity table.

[0089] The sensitivity meter module is also used to perform the following specific operations:

[0090] S41, maintain the historical sequence of the dynamic risk vector field, perform point-by-point vector difference between the current dynamic risk vector field and the historical dynamic risk vector field to obtain the instantaneous difference field, and perform an asymmetric convolution operation on the instantaneous difference field to synthesize the field difference perturbation field. The specific operations are as follows:

[0091] The sensitivity table module continuously maintains the historical sequence of the dynamic risk vector field. This sequence stores multiple frames of historical vector field data within a preset time period in timestamp order, such as one frame every 5 minutes within the last hour, ensuring that the risk distribution status at different times can be traced, providing a benchmark for subsequent change calculations. When the Kalman filter outputs a new prediction result and updates the current dynamic risk vector field, the sensitivity table module performs a point-by-point vector difference operation between the current vector field and the latest frame of the historical vector field in the historical sequence. That is, for the same spatiotemporal coordinate point in the two vector fields, the gradient vector of the current vector field is subtracted from the corresponding gradient vector of the historical vector field to obtain the vector change at each spatiotemporal point. The changes collectively constitute the instantaneous difference field, which directly characterizes the instantaneous dynamic changes of the risk vector field. To mitigate the interference of instantaneous noise, such as local vector mutations caused by sensor errors, on change assessment, and to highlight the core areas of risk changes, the sensitivity table module performs an asymmetric convolution operation on the instantaneous difference field, selecting a preset asymmetric convolution kernel, such as a 3×3 spatial kernel and a 1×1 temporal kernel. The weight of the core area is set to 0.6, and the weight of the edge area is set to 0.1 to 0.2. Through convolution operations, the difference field is smoothed and its features are enhanced, ultimately synthesizing a differential perturbation field. This perturbation field can accurately reflect the effective changing trend of the risk vector field.

[0092] S42, reconstruct the influence domain of the collaborative decision-making joint based on its spatiotemporal morphological characteristics. Within the influence domain, calculate the average cosine of the angle between the field difference perturbation field vector and the original dynamic risk vector field, as well as the spatial integral of the perturbation vector magnitude, and synthesize them into a perturbation penetration index. The specific operation is as follows:

[0093] The sensitivity table module reconstructs the influence domain of each joint based on its spatiotemporal morphological characteristics, including spatiotemporal range, center coordinates, and density peak location. Using the joint's center spatiotemporal coordinates as the core, and combining the span of its spatiotemporal range and density gradient distribution, a spatiotemporal region slightly larger than the joint itself is defined as the influence domain. This ensures coverage of the joint and the surrounding path areas affected by its decisions. Within the reconstructed influence domain, the sensitivity table module performs two calculations: First, it calculates the average cosine of the angle between the field difference perturbation field vector and the original dynamic risk vector field (i.e., the historical vector field). This is obtained by calculating the arithmetic mean of the cosine of the angle between the two vectors at all spatiotemporal points within the influence domain. This value is closer to... The value of 1 indicates that the higher the directional consistency between the disturbance vector and the original risk vector, the more significant the impact of the disturbance on the continuity of the risk evolution trend. Secondly, the spatial integral of the disturbance vector magnitude is calculated, that is, the disturbance vector magnitudes of all spatiotemporal points in the influence domain are accumulated and summed. The larger the integral result, the higher the intensity of the disturbance in the region and the more obvious the change in risk distribution. Subsequently, the sensitivity table module weights and fuses the average cosine value of the included angle with the spatial integral of the disturbance vector magnitude. The weight coefficients are set to 0.4 and 0.6, respectively, to obtain the disturbance penetration index, which comprehensively characterizes the degree of disturbance impact on the joint. The index ranges from 0 to 1. The higher the index, the more significant the impact of the disturbance on the collaborative decision-making joint.

[0094] S43. Analyze the topology of the scheduling decision network composed of flexible instruction sets and cooperative decision joints, identify the logical dependencies between joints, and calculate the topology dependency vulnerability coefficient based on the betweenness centrality of each joint in the network and the number of downstream unresolved branch paths. The specific operations are as follows:

[0095] The sensitivity table module first constructs a scheduling decision network topology composed of all flexible instruction clusters and collaborative decision joints. Collaborative decision joints serve as nodes in the topology, and path segments of flexible instruction clusters act as edges connecting these nodes, forming a complete decision-related network. Based on this topology, the sensitivity table module uses graph analysis algorithms to identify logical dependencies between joints, clarifying the preceding related joints (upstream joints influencing its decisions) and the subsequent related joints (downstream joints affected by its decisions), thus constructing a dependency graph between joints. Subsequently, two key topological parameters are calculated: the betweenness centrality of a joint, calculated by statistically analyzing the frequency of its occurrence on all shortest paths in the network and combining this with the total number of paths in the network. The higher the number centrality, the stronger the pivotal role of the joint in the scheduling decision network, and the greater its impact on the overall decision-making chain. The second factor is the number of downstream undecided branch paths, which is the total number of backup path segments associated with the joint that have not yet been determined for execution. The larger the number, the higher the decision uncertainty of the joint, and the wider the scope of its impact on subsequent path scheduling. The sensitivity table module normalizes the betweenness centrality and the number of downstream undecided branch paths, and then calculates the topology dependency vulnerability coefficient by weighted summation. The weight of betweenness centrality is set to 0.5, the weight of the number of paths is set to 0.5, and the coefficient ranges from 0 to 1. The higher the coefficient, the more vulnerable the joint is in the topology network, and the more obvious the chain reaction on the overall decision after being affected by disturbances.

[0096] S44, the disturbance penetration index and the topology dependency vulnerability coefficient are nonlinearly weighted and fused, and an attenuation factor based on the ship's current pose and the expected arrival time difference of each joint is introduced to generate a reassessment priority score for each joint. Based on this, sensitivity levels are divided and a dynamic sensitivity mapping table is generated. The specific operations are as follows:

[0097] The sensitivity table module employs a nonlinear weighted fusion algorithm to integrate the disturbance penetration index and the topological dependence vulnerability coefficient. The Sigmoid function is selected as the fusion function. By adjusting function parameters, such as setting the slope coefficient to 2.5 and the offset to 0.5, the fusion result accurately reflects the synergistic effect of the two indicators. This highlights the impact of high disturbance penetration while also considering the risk of high topological vulnerability, avoiding priority bias caused by single-indicator evaluation. Simultaneously, an attenuation factor is introduced. This factor is calculated based on the ship's current position and attitude, such as latitude, longitude, heading, and speed, and the estimated arrival time difference between each collaborative decision-making joint. The smaller the time difference, the smaller the attenuation factor, ranging from 0.3 to 1, indicating the ship's distance from the expected arrival time. The closer a decision is to a joint, the higher its timeliness requirement and the less likely its priority is to be attenuated. The greater the time difference, the larger the attenuation factor and the lower the priority, thus adapting to the timeliness requirements of emergency response. The fusion result is multiplied by the attenuation factor to generate a reassessment priority score for each collaborative decision-making joint, with a value range of 0 to 1. Then, sensitivity levels are divided according to the score: a score above 0.7 is a high sensitivity level, 0.3 to 0.7 is a medium sensitivity level, and a score below 0.3 is a low sensitivity level. A dynamic sensitivity mapping table is generated based on the sensitivity level to clearly indicate whether each collaborative decision-making joint needs to be reassessed due to changes in prediction. High and medium sensitivity levels need to be reassessed, while low sensitivity levels can be temporarily disregarded.

[0098] In some embodiments of the present invention, a strategy tree update module is also included, which is used to trigger a re-evaluation of each alternative path segment after the joint when the ship approaches a collaborative decision joint that is identified by the sensitivity table as needing to be re-evaluated, and update the strategy tree containing the currently executed instructions and the preparatory instructions corresponding to each alternative path segment according to the re-evaluation results.

[0099] The strategy tree update module is also used to perform the following specific operations:

[0100] S51, based on the ship's current pose, velocity vector, and spatiotemporal morphological characteristics of the cooperative decision-making joints, dynamically defines a spatiotemporal cone extending from the ship's current position to the vicinity of the joint's spatiotemporal coordinates as a constrained reassessment domain. The specific operation is as follows:

[0101] The strategy tree update module first obtains the ship's current pose, such as latitude, longitude, heading, and altitude; velocity vectors, such as speed and direction of travel; and spatiotemporal morphological characteristics of the collaborative decision-making joints, such as spatiotemporal range, center coordinates, and estimated arrival time. Based on these, it dynamically defines the restricted reassessment domain. This domain adopts a spatiotemporal cone structure, with the apex of the cone being the intersection of the ship's current position and current time. The generatrix of the cone extends along the ship's current heading to the vicinity of the spatiotemporal coordinates of the collaborative decision-making joint. The spatial angle and time span of the cone are determined through dynamic calculation, with the spatial angle based on the ship's speed. The fluctuation range of the vector and the spatial span of the joint are set. For example, when the ship's speed fluctuates greatly, the angle is appropriately expanded to 15° to 20° to ensure that the possible navigation deviation range of the ship is covered. The time span is set to 1.2 to 1.5 times the expected arrival time difference from the ship's current position to the joint, taking into account both navigation timeliness and assessment redundancy. The construction of the spatiotemporal cone can accurately limit the spatiotemporal range of reassessment, avoid invalid assessment of paths in irrelevant areas, improve the efficiency of reassessment, and ensure that the assessment range covers the joint and the key navigation section before the ship arrives at the joint.

[0102] S52, based on the historical assessment status of the backup path segment within the restricted reassessment domain, and combined with the cumulative change of the dynamic risk vector field within this domain, incremental field integral correction and spatiotemporal conflict rehearsal are performed. The specific operations are as follows:

[0103] The strategy tree update module uses the historical assessment status of all alternative path segments within the restricted reassessment domain as a benchmark. This historical status includes core information such as the path's feasibility score, risk suitability, spatiotemporal trajectory parameters, and conflict records with other paths. This serves as a reference for incremental correction. It combines the cumulative change of the dynamic risk vector field within the assessment domain—that is, the total change in the magnitude and direction of the risk vector field from the historical assessment time to the current time—by calculating the cumulative sum of risk vector changes at each spatiotemporal point within the domain. Incremental field integral correction is then applied to the alternative path segments. This incremental correction only performs local integral adjustments on path segments in areas of risk change, rather than recalculating the entire path. For example, for areas where the risk vector magnitude increases by more than 30%, the integral step size is reduced to 0.6 to 0.8 times the original step size, and the risk suitability of the path is recalculated to ensure that the path matches the latest risk distribution. Simultaneously, spatiotemporal conflict rehearsals are conducted. By simulating the navigation trajectory of ships along the corrected path and combining it with the real-time navigation plans of other ships, spatiotemporal intersections between path segments are detected, potential navigation conflicts are predicted, and key information such as conflict location and time is recorded.

[0104] S53, if the joint involves multiple ships, then the incremental correction scheme sets of each ship are subjected to non-dominated negotiation and merging based on Pareto front search to generate a joint path correction scheme package. The specific operation is as follows:

[0105] When a collaborative decision-making joint involves multiple vessels, meaning that the flexible command families of multiple vessels converge at this joint and all require path reassessment, the strategy tree update module first collects the incremental correction scheme set for each vessel within the constrained reassessment domain. Each scheme set contains multiple alternative path correction schemes for that vessel, along with evaluation indicators such as risk suitability, sailing time, and task completion rate for each scheme. Subsequently, an algorithm based on Pareto front search is used to perform non-dominant screening on all schemes, selecting non-dominant schemes—that is, optimal solutions for which no other scheme is superior in all evaluation indicators—and eliminating inferior schemes to narrow down the candidate range. Based on the selected set of non-dominated solutions, non-dominated negotiation is conducted among multiple vessels. During the negotiation process, the solutions are adjusted and merged based on the task priorities of each vessel, such as emergency response priority, risk prevention priority, path conflict situation, and resource suitability. For example, when there is a spatiotemporal conflict between the paths of two vessels, the original modified solution of the higher-priority vessel is retained first, and the path parameters of the lower-priority vessel are finely adjusted to resolve the conflict. Finally, the optimal solution after negotiation is integrated to generate a joint path modification solution package. This solution package not only ensures the optimality of the paths of each vessel, but also achieves the coordination and conflict-free operation of multiple vessels, which is suitable for the needs of joint emergency response of multiple vessels.

[0106] S54, the joint path correction scheme package is attached as a shadow branch to the corresponding node of the policy tree, the heading or speed parameters of the currently executed command are adjusted to connect the shadow branch, and the shadow branch is activated after passing the preset switching point to update the policy tree. The specific operations are as follows:

[0107] The strategy tree update module uses the generated joint path correction scheme package as a shadow branch, attaching it to the node in the strategy tree corresponding to the collaborative decision joint. The topology of the shadow branch is consistent with the main branch of the strategy tree, containing the corrected alternative path segment and corresponding preparatory instructions, but it is not activated immediately to avoid directly replacing the currently executed instructions and causing navigation interruption or confusion. Subsequently, based on the path parameters of the shadow branch, the heading or speed parameters of the ship's currently executed instructions are adjusted. For example, when there is a heading deviation between the shadow branch path and the current execution path, the current heading is gradually adjusted to the direction connecting with the shadow branch, with the adjustment rate controlled between 30% and 50% of the ship's maximum turning angle. To ensure a smooth transition in navigation attitude; when the speed is mismatched, the speed is finely adjusted to match the navigation rhythm of the shadow branch, so that the ship can smoothly connect to the subsequent shadow branch path. The preset switching point is set based on the spatiotemporal distance between the ship and the collaborative decision-making joint. For example, when the spatial distance between the ship and the joint is less than 5 nautical miles and the estimated arrival time difference is less than 10 minutes, the switching point activation mechanism is triggered. Once the ship passes the preset switching point, the strategy tree update module immediately activates the shadow branch, replacing the original path branch and preparatory instructions, and completes the dynamic update of the strategy tree, so that the strategy tree always adapts to the latest risk prediction and path correction results, ensuring the real-time and accuracy of emergency dispatch decisions.

[0108] In some embodiments of the present invention, a parameter adjustment module is also included, which is used to match the actual navigation trajectory of the ship with historical pre-instructions. When the actual trajectory matches a pre-instruction and the corresponding predicted scenario is confirmed, the parameter value of the nonlinear mapping is adjusted according to the matching case.

[0109] The parameter adjustment module is also used to perform the following specific operations:

[0110] S61, calculate the spatiotemporal morphological similarity between the actual flight trajectory and the historical preliminary instructions. These features include changes in trajectory curvature, the order and time difference of passing geographic markers, and the dynamic risk vector field's influence pattern experienced by the trajectory. When the comprehensive score exceeds the confidence threshold and the uniqueness verification passes, it is determined to be a confidence-matched case. The historical preliminary instructions originate from the policy tree. The specific operations are as follows:

[0111] The parameter adjustment module first extracts historical preparatory instructions stored in the strategy tree. These instructions contain core information such as spatiotemporal trajectory planning and risk adaptation parameters for the corresponding alternative path segments, serving as a reference benchmark for matching the actual navigation trajectory. Simultaneously, the module collects the ship's actual navigation trajectory data, extracting the trajectory's spatiotemporal morphological characteristics. Specifically, this includes trajectory curvature changes, the order and time difference of passing through geographic markers, and the dynamic risk vector field's action mode experienced by the trajectory. Trajectory curvature changes are calculated by constructing a curvature sequence from the curvature values ​​of each trajectory node and comparing it with the curvature sequences corresponding to historical preparatory instructions. The order of passing through geographic markers must be perfectly matched. The time difference is calculated by taking the absolute value of the difference between the actual passing through each marker and the preset time of the instruction, and then averaging the results. The time difference index and the risk vector field action mode are used to extract the magnitude and direction of the risk vector at each moment in the area traversed by the trajectory, forming an action mode sequence. This sequence is then matched with the risk action mode preset in the instruction. The parameter adjustment module uses a weighted summation method to calculate the comprehensive similarity score of the above features. The weight of trajectory curvature change is set to 0.3, the weight of geographic marker time difference is set to 0.4, and the weight of risk action mode is set to 0.3. The score range is 0 to 1, and the confidence threshold is set to 0.7. When the comprehensive score exceeds this threshold, a uniqueness verification is performed. If no other historical pre-instruction has a similarity score close to the current actual trajectory (i.e., the difference is less than 0.05), it can be determined as a confidence matching case, ensuring the accuracy and uniqueness of the matching results.

[0112] S62, for confidence matching cases, extract the previously predicted dynamic risk vector field sequence within the corresponding spatiotemporal window, and infer the apparent risk field based on the actual trajectory. Compare the two to generate a prediction observation bias field. The specific operations are as follows:

[0113] For the confidence-matched cases after the judgment, the parameter adjustment module first defines the corresponding spatiotemporal window, using the start and end times of the actual navigation trajectory as the time boundary and the spatial range traversed by the trajectory as the spatial boundary, ensuring that the window can completely cover the navigation time period and area corresponding to the case. Within this spatiotemporal window, the dynamic risk vector field sequence generated by historical prediction is extracted. This sequence contains the risk vector magnitude and direction information of each time and spatial point within the window, completely recording the predicted risk distribution status at that time. At the same time, the apparent risk field is inferred from the actual navigation trajectory of the ship: combining the actual oil film coverage, real-time sea state data, and measured risk information of the area traversed by the trajectory during the ship's navigation, the actual risk distribution of each spatiotemporal point within the spatiotemporal window is calculated, and the apparent risk field is constructed. The apparent risk field is compared point by point with the historical predicted dynamic risk vector field sequence, and the risk vector magnitude deviation and direction deviation of each spatiotemporal point are calculated. These are integrated to form a prediction observation deviation field, which can accurately characterize the difference between historical predicted risk and actual risk.

[0114] S63, Match the predicted observation bias field pattern with a preset perturbation direction, and perform directional perturbation on the nonlinear mapping parameters along the matching direction to generate candidate parameters. The specific operation is as follows:

[0115] The parameter adjustment module presets multiple sets of perturbation directions, each corresponding to a typical prediction bias pattern. For example, for a bias pattern of underestimating the risk magnitude, a perturbation direction that increases the risk mapping weight is preset; for a pattern of risk direction prediction bias, a perturbation direction that adjusts the gradient vector projection coefficient is preset, forming a perturbation direction library. This library performs pattern recognition on the predicted observation bias field, extracting core features such as the dominant bias type, bias distribution range, and bias magnitude. These features are then matched with preset patterns in the perturbation direction library to determine the optimal perturbation direction. For example, when the bias field exhibits local... When the risk magnitude of a region is overestimated by more than 30%, a perturbation direction that reduces the weight of the corresponding nonlinear mapping in that region is matched. The parameter values ​​of the nonlinear mapping are perturbed in a directional manner along the matched perturbation direction. The perturbation step size is dynamically set according to the magnitude of the deviation. For example, when the deviation value exceeds 50% of the threshold, the perturbation step size is set to 10% to 15% of the current parameter value; when the deviation value is between 20% and 50% of the threshold, the perturbation step size is set to 5% to 10%, ensuring that the perturbation amplitude matches the degree of deviation. Multiple sets of candidate parameters are generated through directional perturbation, and each set of candidate parameters corresponds to a perturbation scheme.

[0116] S64, accumulate confidence matching cases and their deviation patterns and parameter perturbation records to form case clusters, analyze the results of historical perturbations on case clusters, and iteratively relax and update the current parameter vector in the direction that makes the prediction deviation of the corresponding type of the case cluster tend to decrease. The specific operation is as follows:

[0117] The parameter adjustment module continuously accumulates confidence matching cases, integrating and recording the deviation pattern, corresponding directional perturbation scheme, parameter values ​​after perturbation, and subsequent changes in prediction deviation for each case to form a case cluster. Cases are then categorized and archived according to deviation pattern type; for example, cases with different patterns such as risk underestimation and directional prediction deviation are grouped into different clusters for targeted analysis. The module analyzes the effects of historical perturbation records within each case cluster, statistically analyzing the reduction in subsequent prediction deviation for different perturbation schemes under the same deviation pattern. The optimal perturbation scheme that minimizes prediction deviation is then selected, clarifying the optimal direction and magnitude of parameter adjustment under this pattern. Based on the optimal perturbation scheme, an iterative relaxation update algorithm is used to adjust the parameter vector of the current nonlinear mapping. Each iteration only makes small adjustments to the parameters, with the adjustment range being 30% to 50% of the optimal perturbation range. This avoids parameter mutations that could lead to a decrease in prediction stability. During the iteration process, the prediction effect after parameter adjustment is continuously verified. If the prediction deviation decreases to a preset threshold after adjustment, such as when the deviation value is less than 5%, the iteration stops. If the threshold is not reached, the above analysis and adjustment process is repeated until the parameter adjustment stabilizes the prediction deviation within a reasonable range. This achieves dynamic optimization of the nonlinear mapping parameters and improves the accuracy of subsequent risk prediction and path planning.

[0118] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. An oil spill emergency response decision system based on Kalman filtering, characterized in that, include: The risk field generation module generates a dynamic risk vector field based on the uncertainty of the Kalman filter prediction output and the oil film prediction boundary through nonlinear mapping. The dynamic risk vector field contains the probability value of each location being covered by oil film and the corresponding gradient vector. The gradient vector represents the direction and rate of the probability change. The instruction cluster generation module generates a flexible instruction cluster for the ship to be scheduled based on the gradient vector. The flexible instruction cluster includes a main path and multiple backup path segments at path bifurcation nodes. The method by which the instruction cluster generation module generates the flexible instruction cluster is as follows: starting from the starting position, the main path is generated by numerical integration of the dynamic risk vector field; branch budding points are identified along the main path and short-range exploration paths are generated as backup path segments; the feasibility of the main path and backup path segments is verified and pruned based on parameters characterizing the ship's maneuverability; after spatiotemporal conflict detection and resolution of the pruned path segments, the main path and its associated backup path segments are organized into a flexible instruction cluster with a topological network structure. The joint recognition module is used to identify the intersection nodes of all flexible instruction clusters and define them as collaborative decision-making joints; The sensitivity table module is used to identify whether each collaborative decision-making joint needs to be re-evaluated due to the change in the prediction when the prediction is updated, based on whether the change in the gradient vector exceeds a preset threshold, and generate a sensitivity table. The strategy tree update module is used to trigger a re-evaluation of each alternative path segment after the collaborative decision joint when the ship approaches a collaborative decision joint that is identified by the sensitivity table as needing re-evaluation, and update the strategy tree containing the currently executed instructions and the preparatory instructions corresponding to each alternative path segment according to the re-evaluation results. The parameter adjustment module is used to match the ship's actual navigation trajectory with historical pre-instructions. When the actual trajectory matches a pre-instruction and the corresponding predicted scenario is confirmed, the parameter values ​​of the nonlinear mapping are adjusted according to this matching case.

2. The oil spill emergency response decision system based on Kalman filtering according to claim 1, characterized in that, The risk field generation module includes: The eigenvalues ​​and eigenvectors of the covariance matrix output by the Kalman filter are extracted. Starting from the centroid of the oil film, a probability wavefront propagation simulation is performed along the direction indicated by the eigenvectors and with the momentum quantized by the eigenvalues ​​to generate a probability density evolution field. Based on ocean current vector field data, the probability density evolution field is vector synthesized and probability density corrected to generate a physically corrected probability field. Calculate the gradient of the physical correction probability field, identify the ridges in the physical correction probability field, strengthen the gradient vector by normal projection toward the ridges, and generate a dynamic risk vector field. Based on wind field data, a dynamic threshold is set for the modulus of the dynamic risk vector field. The dynamic threshold is calculated using a nonlinear sensitivity function based on the angle between the wind speed, wind direction and the main axis of oil film expansion.

3. The oil spill emergency response decision system based on Kalman filtering according to claim 2, characterized in that, The instruction cluster generation module includes: Starting from the initial position, the main path is generated by numerical integration of the dynamic risk vector field, where the integration step size is adjusted according to the magnitude of the gradient vector, and a weak gravitational term pointing to the task target point is introduced. Along the main path, calculate the orthogonal components of the gradient vector at each point that are perpendicular to the tangent direction of the main path. Mark the points where the magnitude of the orthogonal components exceeds the threshold as branch budding points. Perform finite step field integration from each budding point along the positive and negative directions of its orthogonal components to generate a short-range exploration path. Based on parameters characterizing ship maneuverability, instantaneous maneuverability envelopes are calculated for the main path and short-range exploration path, and the feasibility of the path is judged and pruned according to the envelopes. All pruned path segments are projected onto a spatiotemporal coordinate system, the spatiotemporal intersection between path segments is detected, conflicts are resolved by adjusting the parameters of conflicting path segments, and the main path of each ship and its associated backup path segments are organized into a flexible instruction cluster of topological network structure.

4. The oil spill emergency response decision system based on Kalman filtering according to claim 3, characterized in that, The joint recognition module includes: The path segments of the topological network structure are discretized into spatiotemporal waypoints with timestamps. Spatiotemporal occupancy voxels are generated based on the spatiotemporal waypoints, and all voxels are fused to construct a spatiotemporal trajectory volume. The spatiotemporal trajectory volume uses voxel density to represent the occupancy frequency. A four-dimensional morphological analysis is performed on the density field of the spatiotemporal trajectory volume to extract regions with density exceeding a threshold and spatiotemporal connectivity as aggregates, and the spatiotemporal morphological features of the aggregates are extracted.

5. The oil spill emergency response decision system based on Kalman filtering according to claim 4, characterized in that, The joint recognition module also includes: The spatiotemporal range of the aggregate is coupled with the dynamic risk vector field. The average magnitude and directional consistency of the risk vector field within the range are calculated. Combined with the proximity relationship with the ridge line in the dynamic risk vector field, the aggregate is assigned a decision urgency weight. Based on the communication link model and resource loading location information, the feasibility of communication connection and resource delivery of the aggregate is verified. The spatiotemporal coordinates of the aggregate that pass the verification and whose decision urgency weight is greater than the preset threshold are defined as collaborative decision joints.

6. The oil spill emergency response decision system based on Kalman filtering according to claim 5, characterized in that, The sensitivity meter module includes: Maintain the historical sequence of the dynamic risk vector field, perform point-by-point vector difference between the current dynamic risk vector field and the historical dynamic risk vector field to obtain the instantaneous difference field, and perform an asymmetric convolution operation on the instantaneous difference field to synthesize the field difference perturbation field; The influence domain of the collaborative decision-making joint is reconstructed based on its spatiotemporal morphological characteristics. Within the influence domain, the cosine of the average angle between the field difference disturbance field vector and the original dynamic risk vector field, as well as the spatial integral of the disturbance vector magnitude, are calculated and combined into a disturbance penetration index.

7. The oil spill emergency response decision system based on Kalman filtering according to claim 6, characterized in that, The sensitivity meter module also includes: The topology of the scheduling decision network, which consists of flexible instruction clusters and collaborative decision joints, is analyzed. The logical dependencies between collaborative decision joints are identified. Based on the betweenness centrality of each collaborative decision joint in the network and the number of downstream unresolved branch paths, the topology dependency vulnerability coefficient is calculated. The disturbance penetration index and the topology dependence vulnerability coefficient are nonlinearly weighted and fused, and an attenuation factor based on the ship's current pose and the expected arrival time difference of each collaborative decision joint is introduced to generate a re-evaluation priority score for each collaborative decision joint. Based on this, sensitivity levels are divided and a dynamic sensitivity mapping table is generated.

8. The oil spill emergency response decision system based on Kalman filtering according to claim 7, characterized in that, The policy tree update module includes: Based on the ship's current pose, velocity vector, and spatiotemporal morphological characteristics of the collaborative decision-making joint, a spatiotemporal cone extending from the ship's current position to the vicinity of the spatiotemporal coordinates of the collaborative decision-making joint is dynamically defined as a constrained re-evaluation domain. Based on the historical assessment status of the backup path segment within the restricted reassessment domain, and combined with the cumulative change of the dynamic risk vector field within the domain, incremental field integral correction and spatiotemporal conflict rehearsal are performed. If the collaborative decision-making joint involves multiple ships, the incremental correction scheme sets of each ship are non-dominated and merged based on Pareto front search to generate a joint path correction scheme package. The joint path correction scheme package is attached as a shadow branch to the corresponding node of the policy tree. The heading or speed parameters of the currently executed command are adjusted to connect the shadow branch. After passing a preset switching point, the shadow branch is activated to update the policy tree.

9. The oil spill emergency response decision system based on Kalman filtering according to claim 8, characterized in that, The parameter adjustment module includes: The similarity of the spatiotemporal morphological features between the actual flight trajectory and the historical pre-instructions is calculated. The features include changes in trajectory curvature, the order and time difference of passing through geographic markers, and the action mode of the dynamic risk vector field experienced by the trajectory. When the comprehensive score exceeds the confidence threshold and the uniqueness verification is passed, it is determined to be a confidence matching case. The historical preparatory instructions are derived from the policy tree; for confidence matching cases, the sequence of dynamic risk vector fields that were predicted within the corresponding spatiotemporal window is extracted, and the apparent risk field is inferred from the actual trajectory. The two are then compared to generate a prediction observation bias field.

10. The oil spill emergency response decision system based on Kalman filtering according to claim 9, characterized in that, The parameter adjustment module further includes: The predicted observation bias field pattern is matched with a preset perturbation direction, and the nonlinear mapping parameters are perturbed in a directional manner along the matching direction to generate candidate parameters. Accumulate confidence matching cases and their deviation patterns and parameter perturbation records to form case clusters. Analyze the results of historical perturbations on the case clusters, and iteratively relax and update the current parameter vector in the direction that makes the prediction deviation of the corresponding type of the case cluster tend to decrease.

Images