A maritime regional navigation safety management and optimization method

Abstract

The present application provides a kind of offshore area navigation safety control and optimization method, comprising: step 1, the multi-source observation data of offshore navigation target is collected, and pre-processing is carried out, to obtain the space-time observation correlation graph;Step 2, the dynamic credibility of each observation node is evaluated using multi-modal consistency graph neural network;According to the obtained dynamic credibility, an adaptive weighted factor graph model is constructed, and the system state is solved to obtain the optimal fusion track, the corrected environmental parameter and the posterior error covariance matrix;Step 3, based on the optimal fusion track, the corrected environmental parameter and the posterior error covariance matrix, a regional risk field is constructed;Step 4, based on the regional risk field, a multi-objective collaborative optimization model is constructed and solved to obtain the navigation control parameter;Step 5, execute navigation control parameter, and adopt multilevel closed loop feedback to optimize, complete offshore area navigation safety control and optimization.

Description

A method for safety management and optimization of navigation in maritime areas Technical Field

[0001] This invention relates to a method for navigation safety management and optimization, and particularly to a method for navigation safety management and optimization in maritime areas. Background Technology

[0002] With the rapid development of the global shipping industry, maritime navigation safety management is crucial for safeguarding life and property, improving waterway operational efficiency, and protecting the marine environment. Currently, maritime surveillance mainly relies on multi-source observation data, including Automatic Identification System (AIS), shore-based / shipborne radar, meteorological and oceanographic (METOC) sensors (such as wind speed, wind direction, wave, and current sensors), Global Navigation Satellite System (GNSS), and infrared / video surveillance. However, these sensors differ significantly in information type, measurement accuracy, coverage, and update frequency. For example, radar can provide second-level information such as bearing, distance, and relative speed, but is susceptible to sea state, weather, and obstructions; video surveillance is significantly constrained by lighting and clutter; the refresh cycle of AIS data varies from seconds to minutes depending on the navigation status, and its timeliness and spatial resolution are inconsistent with radar and video; while GNSS and VHF Data Exchange System (VDES) can provide positioning and communication support, their data systems are fundamentally different from other active observation methods. These heterogeneities make it difficult to directly align multi-source data in terms of time reference, spatial coordinate system, and noise statistics. In addition, the observation quality of each sensor changes dynamically with the environment. Traditional direct fusion methods are not robust to abnormal data (such as AIS false alarms, radar clutter, etc.), which can easily lead to distortion of fusion results, increased false alarm or missed alarm rates, and thus make it impossible to accurately and reliably perceive the situation of maritime targets and the risks of waterways.

[0003] Current mainstream technical solutions typically first perform spatiotemporal alignment of multi-source data, followed by target-level fusion using statistical filtering methods. While this framework improves surveillance capabilities to some extent, it faces the following bottlenecks in complex sea conditions and high-density traffic scenarios:

[0004] Multi-source fusion lacks the ability to dynamically adapt to the quality of heterogeneous data. Existing methods mostly use fixed weights or simple spatiotemporal correlation rules, lacking quantitative models that can evaluate data consistency and reliability across sensors in real time. This makes it difficult to adjust the fusion strategy according to the dynamic changes in observation quality, resulting in limited ability to suppress abnormal observations and a high overall false alarm rate of the system.

[0005] There is a disconnect between the quantitative assessment of regional navigation risks and the calculation of control parameters. Existing risk assessments are mostly based on static geographical partitions or single historical indicators, lacking the ability to predict the short-term dynamic risks of the entire region. More importantly, when generating control instructions, there is a lack of models and methods to automatically and accurately map the assessment results of dynamic risks into specific navigation control parameters (such as ship speed limits, recommended routes, and prohibited navigation periods), making it difficult to achieve comprehensive optimization among multiple objectives such as ensuring navigation safety, improving navigation efficiency, and meeting environmental protection requirements. Summary of the Invention

[0006] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for the management and optimization of navigation safety in maritime areas, addressing the shortcomings of the existing technology.

[0007] To address the aforementioned technical problems, this invention discloses a method for maritime area navigation safety management and optimization, comprising:

[0008] Step 1: Collect multi-source observation data of maritime navigation targets and preprocess them to obtain a spatiotemporal observation correlation map;

[0009] Step 2: Use a multimodal consensus graph neural network to evaluate the dynamic reliability of each observation node; construct an adaptive weighted factor graph model based on the obtained dynamic reliability, and solve for the system state to obtain the optimal fused trajectory, the corrected environmental parameters, and the posterior error covariance matrix.

[0010] Step 3: Construct the regional risk field based on the optimal fused trajectory, the corrected environmental parameters, and the posterior error covariance matrix;

[0011] Step 4: Based on the regional risk field, construct and solve a multi-objective collaborative optimization model to obtain navigation control parameters;

[0012] Step 5: Execute navigation control parameters and use multi-level closed-loop feedback for optimization to complete the safety management and optimization of navigation in the maritime area.

[0013] Furthermore, the preprocessing described in step 1 includes:

[0014] Step 1-1, spatiotemporal alignment and data standardization, is as follows:

[0015] Coordinated Universal Time (UTC) is adopted as the unified time base. A fixed fusion time grid is selected as the starting time. Second, For time indexing, the discretized time series is defined as follows: ;

[0016] Any sensor on a maritime navigation target ,exist The observation data at each moment is used to compensate for the clock deviation between the sensor and Coordinated Universal Time (UTC), and linear interpolation is used to align the observation data to the most recent fusion time. ;

[0017] All sensor observation data are transformed and unified into the Northeast ENU coordinate system with the preset location as the origin.

[0018] Each observation data point is represented as an observation tuple. The details are as follows:

[0019]

[0020] in, The aligned measurement state vector has dimensions. For its dimensions, This is the corresponding measurement noise covariance matrix. Let be the source quality feature vector of the sensor. This is a sensor type identifier.

[0021] Steps 1-2, consistency gating and observation graph construction, are detailed below:

[0022] For observation data Calculate Mahalanobis distance , means as follows:

[0023]

[0024] in, Based on The target state at a given time is obtained using a pre-defined motion model. Predicted measurement value at time. The new information covariance matrix is ​​calculated as follows:

[0025]

[0026] in, For the observation matrix, for The target state prediction covariance matrix at time 1.

[0027] Set threshold ,like If the observed data is an outlier, it is determined to be outlier and removed; otherwise, it is retained as a candidate observation.

[0028] In length Within the sliding time window, a spatiotemporal observation correlation graph is constructed based on candidate observations. , means as follows:

[0029]

[0030] in, The node set consists of all observation tuples within the sliding time window. That is, the composition of observation nodes; For an edge set, for any two observed nodes and If the Euclidean distance between the observation locations is less than the set spatial neighborhood threshold or the observation data originates from the same navigation target, then an undirected edge is established. Connect the two, and set the attribute of each edge to the time difference between them. Spatial distance difference With relative speed difference .

[0031] Furthermore, the solution for the system state described in step 2 includes:

[0032] Step 2-1, in At time , the set of system state variables to be solved Defined as:

[0033]

[0034] in, Let be the ship's motion state vector. This is a vector of environmental state parameters.

[0035] Step 2-2, set physical constraints, represented as factors in the factor graph, as follows:

[0036] Define motion constraints, or motion factors, as follows:

[0037]

[0038] in, This represents a pre-defined kinematic model. The process noise is represented by the following formula: .

[0039] Define measurement constraints, or measurement factors, as follows:

[0040]

[0041] in, For sensors The observation function maps the system state to the observation space; To measure noise, its covariance matrix is: .

[0042] Steps 2-3 involve solving for the system state, i.e., performing an optimal estimation of the system state. This is equivalent to solving a nonlinear least squares problem, as follows:

[0043]

[0044] in, This indicates the calculation of Mahalanobis distance; The process noise covariance matrix represents the environmental state. For adaptive weights.

[0045] Steps 2-4 involve iteratively solving the nonlinear least squares problem from Step 2-3 using the Gauss-Newton method. When the state update converges, the optimal estimate of the system state is obtained. That is, the fused state set, which contains the optimal fused track. and corrected environmental parameters .

[0046] Using the final Jacobian matrix With weight matrix Constructing an approximate Hessian matrix Its inverse matrix is ​​the posterior error covariance matrix. , means as follows:

[0047] .

[0048] Furthermore, the adaptive weights mentioned in steps 2-3 are specifically calculated using the following methods:

[0049] Step 2-3-1, for the spatiotemporal observation correlation diagram Each node in Construct the initial feature vector of the node , means as follows:

[0050]

[0051] in, and These are the source quality feature vectors taken from the sensor. Signal quality characteristics and physical strength characteristics To determine the new covariance matrix The calculated geometric consistency characteristics are represented as follows:

[0052]

[0053] Step 2-3-2: A graph attention mechanism is used to aggregate neighborhood information, learn the mutual support relationships between nodes, and calculate the node... Dynamic credibility score .

[0054] Step 2-3-3: Establish a non-linear weight mapping function to assign dynamic credibility scores. Transformed into adaptive information adjustment factor , means as follows:

[0055]

[0056] in, To ensure a minimum weighting coefficient, This is the gain index.

[0057] Steps 2-3-4: Adjust the factor based on adaptive information. Constructing an adaptive measurement information matrix Adaptive weights, also known as adaptive weights, are expressed as follows:

[0058]

[0059] in, For sensors exist The inverse of the noise covariance matrix at any given time.

[0060] Furthermore, the calculations described in step 2-3-2 yield the nodes. Dynamic credibility score ,include:

[0061] Step 2-3-2-1, for node Calculate its relationship with neighboring nodes Attention coefficient between The details are as follows:

[0062]

[0063] in, Indicates including nodes The set of neighboring nodes, including For shared feature projection matrices; This is the attention vector; This indicates vector concatenation; For linear units with leakage correction; It is a normalized exponential function.

[0064] Step 2-3-2-2, Update node features , means as follows:

[0065]

[0066] in, It is a non-linear activation function.

[0067] Step 2-3-2-3: Output nodes using a fully connected layer and the Sigmoid function. Dynamic credibility score , means as follows:

[0068]

[0069] in, and These are the weight vector and the bias scalar, respectively.

[0070] Furthermore, the construction of the regional risk field described in step 3 includes:

[0071] Step 3-1, gridded risk source feature extraction, as detailed below:

[0072] Discretize the target sea area into a set of grids. For each grid exist At any given moment, based on the fused state set Extracting multidimensional risk source feature vectors , means as follows:

[0073]

[0074] in, Indicates traffic density. For the intensity of the encounter, The trace represents the situational uncertainty, i.e., the posterior covariance matrix. , Indicates the corrected environmental parameters The environmental stress degree obtained after normalization and weighted fusion.

[0075] Step 3-2 introduces a partial differential equation as the dynamic evolution model of the regional risk field, as follows:

[0076]

[0077] in, Right now The mesh to be solved exist The overall risk intensity at any given moment It is the speed of risk convection. For two-dimensional spatial gradient operators, The risk diffusion coefficient, This is the risk natural decay coefficient. It is a risk source term, derived from the extracted grid feature vector. The linear weighted sum is obtained as follows:

[0078]

[0079] in, , , and These are the normalized weighting coefficients.

[0080] Step 3-3: Discretize and solve the dynamic evolution model to predict the future. The dynamic distribution of regional risk within a given time period is represented as follows:

[0081] .

[0082] Furthermore, the construction and solution of the multi-objective collaborative optimization model described in step 4 includes:

[0083] Step 4-1: Define the control decision variables for each grid as a vector. , means as follows:

[0084]

[0085] in, For grid speed limits, To designate the direction of navigation, This is a temporary no-navigation sign.

[0086] Step 4-2, Construct a multi-objective vector function , means as follows:

[0087]

[0088] in, For a set of global control strategies, For the sub-objective of total risk exposure, To reduce speed and lose sub-targets, To predict emissions or energy consumption sub-targets.

[0089] Step 4-3: Employ an evolutionary algorithm to minimize the multi-objective vector function. Solving for the final strategy yields the solution. As navigation control parameters.

[0090] Furthermore, the optimization using multi-level closed-loop feedback described in step 5 includes:

[0091] Step 5-1: Monitor the sea area status in real time after the command is executed and calculate the compliance rate. With risk improvement .

[0092] Step 5-2, based on the compliance rate With risk improvement Three-layer feedback control is implemented.

[0093] Furthermore, the calculation of the compliance rate described in step 5-1 With risk improvement ,include:

[0094]

[0095]

[0096] in, The scope in which the instruction takes effect. Indicates an indicator function, For grid The average speed of ships inside the vessel To control parameter values, Allowable engineering errors; and These are the steps before and after the instruction is executed. Regional aggregated risk value over time.

[0097] Furthermore, the three-layer feedback control described in step 5-2 employs the following control logic:

[0098] The communication layer feedback control determines that if the terminal broadcast coverage is less than the threshold, the effective time of the navigation control parameter execution command is extended and the command is retransmitted.

[0099] The execution layer feedback control determines that if the terminal broadcast coverage rate is greater than or equal to the threshold, the communication coverage is considered normal, and the execution compliance rate is further assessed. If the value is less than the threshold, it is determined that the instruction execution rate is low, triggering the preset flexible degradation strategy and generating a new downgraded instruction for issuance.

[0100] Effect layer feedback control, if the compliance rate is high If the value is greater than or equal to the threshold, then the degree of risk improvement is further assessed. If the value is negative, the current instruction will be immediately cancelled, and step 1 will be re-executed based on the latest multi-source observation data.

[0101] Beneficial effects:

[0102] 1. This invention improves the accuracy and robustness of multi-source heterogeneous data fusion. By constructing a dynamic evaluation model for observation reliability based on multimodal consistency learning, it achieves real-time evaluation and adjustment of cross-sensor fusion weights, effectively overcoming the limitations of fixed-weight fusion in complex environments. This mechanism can adaptively suppress anomalous observations and clutter interference, significantly reducing the system's false alarm rate, thereby ensuring stable and reliable tracking of non-cooperative targets and weak signals.

[0103] 2. This invention realizes the dynamic quantification of regional navigation risks and the automatic optimization of control strategies. By establishing a gridded short-term risk prediction field, multiple objectives such as safety, efficiency, and environmental protection are modeled in a unified manner. The optimal navigation control parameters are automatically calculated using a multi-objective intelligent optimization algorithm, and scheduling instructions that can be directly broadcast are generated accordingly. This realizes the automation and intelligence of the entire process from "risk perception - strategy optimization - instruction generation". Attached Figure Description

[0104] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0105] Figure 1 is a schematic diagram of the overall process of the present invention.

[0106] Figure 2 is a schematic diagram comparing the target trajectory tracking effect under complex sea conditions in an embodiment of the present invention.

[0107] Figure 3 is a schematic diagram comparing the tracking error convergence curves in an embodiment of the present invention.

[0108] Figure 4 is a schematic diagram comparing the risk field and trajectory distribution in the region before and after control in an embodiment of the present invention.

[0109] Figure 5 is a schematic diagram of the Pareto frontier and strategy optimization in multi-objective optimization in an embodiment of the present invention. Detailed Implementation

[0110] This invention addresses the shortcomings in the consistency and reliability of multi-source heterogeneous data fusion in existing technologies, as well as the disconnect between regional navigation risk quantification assessment and control parameter calculation. It proposes a method for maritime regional navigation safety management and optimization. The method achieves intelligent management and control across the entire chain by constructing a closed loop of "perception-assessment-decision-control," specifically including the following steps: First, multi-source observation data undergoes spatiotemporal alignment and standardization preprocessing. A multimodal consistency graph neural network is used to achieve credibility-driven adaptive fusion of multi-source heterogeneous data, directly improving the accuracy and robustness of perception. Second, through gridded dynamic risk field prediction and multi-objective optimization, optimal control commands are automatically generated and issued. Intelligent optimization from risk quantification to closed-loop management is achieved based on multi-level feedback. As shown in Figure 1, the specific technical solution is as follows:

[0111] 1. Spatiotemporal alignment and consistency preprocessing of multi-source data.

[0112] Due to significant differences in sampling frequency, coordinate reference, and measurement noise characteristics among various sensors used in maritime surveillance scenarios (including shore-based / patrol boat-deployed radar, optoelectronic facilities, and METOC sea state sensors, as well as AIS and GNSS terminals carried by target vessels), direct fusion will introduce substantial errors. This section aims to establish a unified spatiotemporal reference framework and perform standardized preprocessing on the raw observations to provide high-quality input information for subsequent multi-source fusion models based on factor graph optimization.

[0113] (1) Spatiotemporal alignment and data standardization.

[0114] The system uses Coordinated Universal Time (UTC) as the unified time base. A fixed fusion time grid is selected as the starting time. Second, For time indexing, the discretized time series is defined as follows: For any sensor exist The measurement information at each moment is first spatiotemporally aligned: the clock deviation between each sensor and UTC is compensated, and linear interpolation is used to uniformly align the observation data to the most recent fusion moment. All sensor measurements were converted to the Northeast Eon (ENU) coordinate system with the control center as the origin.

[0115] After the alignment described above, each observation is standardized into a structured observation tuple. :

[0116]

[0117] In the formula, The aligned measurement state vector has the following dimensions. It can be configured according to the integration requirements (e.g.) =5, including position ,speed and azimuth ); The corresponding measurement noise covariance matrix is ​​used to linearly increase the values ​​of the diagonal elements of the covariance matrix during time interpolation based on the time interval between the interpolation point and the original sampling point, so as to more accurately reflect the total uncertainty after data alignment. This is the sensor source quality feature vector, used for subsequent confidence assessment (including radar echo intensity, AIS signal-to-noise ratio, confidence level of visual detection, etc.). This is a sensor type identifier.

[0118] (2) Consistency gating and observation graph construction.

[0119] To initially remove significantly outlier observations and construct the input graph structure for the graph neural network, this step performs consistency gating and builds a spatiotemporal observation correlation graph. First, chi-square threshold detection is performed for standardized observations. Calculate the Mahalanobis distance of its measured innovation. for:

[0120]

[0121] In the formula, Based on The target state at time t is calculated using a uniformly accelerated motion model. Predicted values ​​for time measurements. The new information covariance matrix; This is the observation matrix, used to map the state vector to the observation vector; for The target state prediction covariance matrix at time step 1 reflects the uncertainty of the state prediction. A threshold is set. (Chi-square distribution) (at the 0.99 quantile), if If the observation is found to be an outlier, it is removed; otherwise, it is retained and added to the candidate observation pool for further processing.

[0122] Subsequently, at a length of Within the sliding time window, a spatiotemporal observation correlation graph is constructed based on candidate observations that have passed the above gating. As input to the subsequent consistency learning model:

[0123]

[0124] In the formula, The node set consists of all standardized observation tuples that pass the test within the sliding window. The graph is composed of observations, each of which is a node. This is an edge set established based on the principles of spatiotemporal proximity and identity consistency. For any two observation nodes... and If the two observation locations are spatially adjacent (the Euclidean distance between them is less than the set spatial neighborhood threshold, which is set to 0.5km-3km, and the specific value can be determined based on the navigation density of the controlled sea area) or have related identities (the two observations originate from the same physical target, for example, both carry the same Maritime Mobile Service Identifier (MMSI)), then an undirected edge is established. Connect them. The attribute features of each edge are set as the time difference, spatial distance difference, and relative velocity difference between the two, which are used to characterize the correlation strength between the observations.

[0125] 2. Credibility-driven adaptive factor graph fusion.

[0126] This section aims to address the issues of fixed weights and sensitivity to outliers in multi-source data fusion. Its core idea is to utilize the spatiotemporal observation correlation graph constructed in Section 1. The dynamic credibility of each observation node is learned by using a multimodal consensus graph neural network (MC-GNN) and transformed into adaptive weights in the factor graph optimization framework, thereby adaptively suppressing the interference of false reports and anomalous observations during the fusion process.

[0127] (1) Framework for system state modeling and factor graph fusion.

[0128] Use a length of That is, including A sliding time window at discrete moments. At time , the set of system state variables to be estimated Defined as:

[0129]

[0130] In the formula, The vector representing the ship's motion state; This is a vector of environmental state parameters (such as local ocean current velocity, wave height, etc.) related to the observation, used to correct the influence of the environment on the measurement in the observation model.

[0131] A factor graph is a probabilistic graphical model that decomposes a globally complex joint probability distribution into the product of multiple local functions, thereby transforming the multi-source data fusion problem into a nonlinear least-squares optimization problem. In this invention, the factor graph serves as the core fusion framework, used to uniformly represent ship kinematics, environmental evolution characteristics, and multi-source sensor observation models. It abstracts the three key aspects—"how the ship moves (kinematics)," "how the environment changes (environmental evolution)," and "what the sensors measure (observation)"—into three "factor nodes" in the graphical model, thus obtaining the factor graph model of this application. Specifically, the system follows two physical constraints, corresponding to two types of factor nodes in the factor graph:

[0132] The motion factor is expressed as:

[0133]

[0134] In the formula, This represents a pre-defined kinematic model (such as a uniform motion model). The process noise is represented by the following formula: .

[0135] The measurement factor is expressed as:

[0136]

[0137] In the formula, For sensors The observation function maps the state to the observation space; To measure noise, its covariance matrix is: .

[0138] (2) Credibility assessment based on MC-GNN.

[0139] The graph output from Section 1 Input a multimodal consensus graph neural network to dynamically evaluate the credibility of each observation node. For each node in the graph... (corresponding to a certain standardized observation) First, construct the initial feature vector of the node. ;

[0140]

[0141] In the formula, This is a geometric consistency feature, derived from the chi-square test information in Section 1. The smaller the value, the better it matches the motion prediction. Signal quality characteristics, directly taken from The signal-to-noise ratio and detection confidence level are among the parameters; a higher value indicates that the original signal is more reliable. Physical strength characteristics, taken from The echo intensity or signal amplitude is used to distinguish strong targets from weak clutter.

[0142] For connection nodes and edge Its eigenvector is It represents the time difference, spatial distance difference, and velocity difference between two observations, and is used to quantify the spatiotemporal correlation strength.

[0143] A graph attention mechanism is used to aggregate neighborhood information and learn the mutual support relationships between nodes. For nodes... Its relationship with neighboring nodes ( The attention coefficients among the neighboring nodes (including node i) The calculation is as follows:

[0144]

[0145] In the formula, For shared feature projection matrices; This is the attention vector; This represents vector concatenation. These parameters are obtained through offline training using historical data. It is a linear unit with leakage correction, used to solve the gradient vanishing problem when the neuron output is negative; The normalized exponential function is used to map the attention coefficients to... Subsequently, node features Updated to:

[0146]

[0147] In the formula, The activation function is non-linear. Finally, a fully connected layer and a sigmoid function are used to output the dynamic confidence score of the observation node. :

[0148]

[0149] In the formula, and These are the weight vector and bias scalar of the readout layer, respectively, which are fixed parameters learned during the network training phase by minimizing the sample classification error. This indicates that the observation is highly consistent with surrounding multi-source information and has a high degree of credibility; This indicates that it may be abnormal or interference.

[0150] (3) Construction and optimization of adaptive factor graph.

[0151] This step transforms the data from "deep learning features" to "mathematically optimized weights." Specifically, it utilizes the dynamic credibility score output from the previous step. Transformed into adaptive information adjustment factor :

[0152]

[0153] In the formula, This serves as a safety net weighting coefficient to ensure numerical stability. The gain exponent is used to amplify the suppression effect on low-confidence observations. Therefore, an adaptive measurement information matrix is ​​constructed. :

[0154]

[0155] when At lower levels, , It approaches the zero matrix, thereby automatically weakening the constraint of the observation during optimization.

[0156] Based on the aforementioned adaptive weights, the maximum a posteriori probability estimate of the system state is obtained within a sliding window, and the aforementioned adaptive information matrix is ​​used. Substituting into the factor graph model, specifically, this matrix The observation residual term (i.e., the third term in the following equation) is applied to the nonlinear least squares objective function, thereby automatically weakening the constraint of the observation in the optimization solution. Based on this, the optimal estimate of the system state is obtained. This can be transformed into solving the following nonlinear least squares problem:

[0157]

[0158] In the formula, It is a set of merged states. This indicates the calculation of Mahalanobis distance; Let be the process noise covariance matrix representing the environmental state. The first term... For motion constraints, the corresponding motion factor nodes in the factor graph; the second term For environmental state smoothing constraints based on random walk modeling, corresponding to environmental evolution factor nodes, used to constrain the temporal continuity of environmental parameters such as flow velocity and wave height; the third term For adaptive weights The observation constraints correspond to the measurement factor nodes. The Gauss-Newton method is used to iteratively solve the objective function. When the state update converges, the final Jacobian matrix is ​​used. Constructing an approximate Hessian matrix with the weight matrix Its inverse matrix is ​​mathematically equivalent to the posterior error covariance matrix of the system state. (Based on Cramero's lower bound theory):

[0159]

[0160] After optimization and convergence, the output is optimal. This includes the optimal fused trajectory. Corrected environmental parameters and the posterior error covariance matrix The trace of the covariance matrix This directly reflects the uncertainty of current track tracking. This value will serve as a key risk source input and will be passed to the regional risk assessment module in Section 3.

[0161] 3. Regional navigation risk modeling and multi-objective collaborative optimization.

[0162] Based on the high-precision fused track, environmental parameters, and posterior uncertainty output in Section 2, this section constructs a dynamic and predictable regional risk field. On this basis, a multi-objective optimization model is used to automatically weigh multiple objectives such as navigation safety, navigation efficiency, and environmental protection, and calculate the optimal navigation control parameters.

[0163] (1) Extraction of gridded risk source features.

[0164] Discretize the target sea area into a set of regular grids. For each grid exist At time, based on the fused state set output in Section 2 Extract a multidimensional risk source feature vector :

[0165]

[0166] In the formula, Traffic density is represented by the ratio of the number of ships in the grid to the grid area, which is directly obtained from the spatial distribution statistics of the merged tracks. Encounter intensity is used to quantify the urgency of collision risk. Based on the fused velocity vectors of ships within the grid, the nearest encounter time and nearest encounter distance between each pair of ships are calculated and mapped together into a scalar index. The situational uncertainty is represented by the trace of the posterior covariance matrix in Section 2, which contains ships within the grid. This innovative approach transforms the quality of preceding data fusion into a risk increment for the current level of risk assessment, automatically increasing the conservative risk estimate for the region when tracking uncertainty is high. This represents the degree of environmental stress, derived from the environmental state parameters output in Section 2. The result is obtained after normalization and weighted fusion.

[0167] (2) Prediction of dynamic risk field based on the “convection-diffusion” mechanism.

[0168] To describe the propagation, diffusion, and dissipation of risk in space and time, the following partial differential equation, inspired by the "convection-diffusion" mechanism, is introduced as a dynamic evolution model of the risk field:

[0169]

[0170] In the formula, The mesh to be solved exist The overall risk intensity at any given moment. It is the risk convection velocity, whose direction is consistent with the average heading of ships within the grid, and whose magnitude is consistent with the grid speed limit control variable to be optimized. Positive correlation ( ). It is a two-dimensional spatial gradient operator. The risk diffusion coefficient has a range of values ​​of [value range missing]. km 2 / s represents the spread effect of risk to surrounding areas. This is the risk natural decay coefficient, with a value range of [value range missing]. . It is a risk source term, derived from the extracted grid feature vector. Linear weighting yields:

[0171]

[0172] In the formula, The normalized weighting coefficients can be set according to maritime regulatory preferences. To eliminate dimensional differences, the above feature components are mapped to a minimum value using the Min-Max method before being substituted into the weighting formula. Within the interval. By discretizing and solving the evolution equation of the risk field in the above region, future rolling predictions can be made. Dynamic distribution of regional risks within a time period This provides forward-looking input for subsequent optimization decisions.

[0173] (3) Construction of a multi-objective collaborative optimization model.

[0174] Define the control decision variables for each grid as a vector. , For grid speed limits, To designate the direction of navigation, This is a temporary no-navigation sign. The optimization problem aims to minimize a multi-objective vector function comprised of safety, efficiency, and environmental protection. :

[0175]

[0176] In the formula, This is a set of global control strategies. The sub-objectives are defined as follows:

[0177] Safety objective: Minimize total risk exposure :

[0178]

[0179] That is, minimizing the spatiotemporal integral of the risk intensity of all grids within the prediction period.

[0180] Efficiency objective: Minimize speed loss :

[0181]

[0182] In the formula, For the free-flow speed (historical statistical mean or 85th percentile) of the grid under uncontrolled conditions, minimize the deviation between the speed limit and the free-flow speed.

[0183] Environmental goal: Minimize estimated emissions / energy consumption :

[0184]

[0185] Based on the approximate cubic relationship between ship fuel consumption and speed, the total energy consumption of the controlled area is minimized.

[0186] (4) Automatic optimization of strategies based on evolutionary algorithms.

[0187] The aforementioned objectives conflict with each other, constituting a complex multi-objective constrained optimization problem. This invention employs a Non-Dominated Sorting Genetic Algorithm II (NSGA-II) with an elitist strategy to solve this problem.

[0188] Coding and Constraints: Global Control Strategy Encoded as chromosomes. International Maritime Collision Avoidance Regulations (COLREGs) and channel physical boundaries are used as hard constraints. Specifically, this is achieved by optimizing the grid to specify the navigation direction. Traffic flow is streamlined to align directions, thereby reducing the likelihood of intersections and collisions.

[0189] Pareto Front Search: The algorithm iteratively evolves the population through operations such as selection, crossover, and mutation, eventually converging to a set of Pareto optimal solutions. Any solution in this set cannot further improve any objective without compromising other objectives.

[0190] Inflection point decision: Automatically select the inflection point from the Pareto frontier as the final strategy. This point represents the optimal balance between achieving maximum safety gains at a relatively low cost in terms of efficiency or environmental impact. This refers to the set of optimal navigation control parameters automatically generated by the system.

[0191] 4. Control command generation and closed-loop feedback control.

[0192] This section aims to present the set of optimal control parameters calculated in Section 3. This is transformed into executable and monitorable navigation control commands that conform to international standards. Through a multi-level feedback control mechanism, dynamic evaluation and adaptive adjustment of control strategies are achieved, ultimately forming a complete control closed loop of "decision-execution-evaluation-optimization".

[0193] (1) Instruction encoding and generation based on S-124 standard.

[0194] The discretized optimal control parameters are mapped to structured instruction objects. For the first... one instruction object Its encoded tuple is defined as:

[0195]

[0196] In the formula, The spatial extent (mesh geometry set) in which the command takes effect; For control actions (such as speed limits, one-way traffic, temporary navigation restrictions); The time window in which the instruction takes effect; For control parameter values ​​(such as a speed limit of 12 knots); For version identifier; Instruction priority (1 is the highest, 3 is the lowest).

[0197] Based on the International Hydrographic Organization (IHO) S-124 (Navigation Services and Warning Areas) data model standard, the above tuples are serialized into machine-readable GML / XML format messages. Before distribution, the system automatically performs command compliance verification to ensure that the generated commands do not conflict with established static waterway rules, thereby guaranteeing the legality and executability of the commands.

[0198] (2) Adaptive instruction distribution based on communication status.

[0199] To ensure reliable delivery of instructions, an adaptive distribution strategy is adopted that coordinates VDES and the BeiDou short message system. The terminal broadcast coverage rate for the current batch of instructions is defined. for:

[0200]

[0201] In the formula, For the target number of ships, It is an indicator function (received the first) The confirmation receipt for a single ship is 1).

[0202] Distribution strategy is based on instruction priority Dynamic adjustment:

[0203] High priority instructions ( ): It adopts a dual-channel concurrent mode of VDES and Beidou to maximize delivery probability and timeliness.

[0204] Low and medium priority instructions ): The default is to use VDES single-channel broadcast. If calculated in real time... Below the preset switching threshold (If the value is 0.85 in this invention), the BeiDou link will be automatically activated for supplementary broadcasting to ensure the coverage of the command.

[0205] (3) Multi-level closed-loop feedback and strategy optimization.

[0206] Utilizing shore-based AIS, radar, and other systems to monitor the sea area status in real time after command execution, and calculate the compliance rate. With risk improvement Two key evaluation metrics:

[0207]

[0208]

[0209] In the formula, the speed limit order is used as an example. For grid The average speed of the vessel (if it is a heading-specified command, then the indicator function) The judgment criteria in the text have been adjusted accordingly to heading deviation. ); Allowable engineering errors; and These are before and after the instruction is executed (after) The regional aggregated risk value (time) is calculated using the same method as in Section 3.

[0210] Based on these indicators, the following three-layer feedback control logic is implemented:

[0211] Communication layer feedback: If detected The system has determined that there is insufficient communication coverage. The system will maintain the instruction version number. The command remains unchanged; instead, the effective window is extended and a resend is triggered in an attempt to cover unresponsive vessels.

[0212] Execution layer feedback: If communication coverage is normal ( ), but compliance rate (This embodiment takes) If the command execution rate is low (possibly due to rough seas, ship maneuvering restrictions, etc.), a flexible strategy downgrade is triggered, for example, downgrading "No Navigation" to "Speed ​​Limit" and "Designated Course" to "Recommended Course". A new downgraded command is generated, and the version number is updated. It will be issued later.

[0213] Effect layer feedback: If the instruction is executed ( (Higher), but actual risk improvement If the strategy is deemed ineffective or fails to meet expectations, the current command is immediately revoked, and the latest measured maritime situation (tracks, environment, etc.) is fed back to the optimization model in Section 3 as new initial conditions, triggering a new round of optimization calculations and generating a version number of [version number missing]. Update strategy.

[0214] Example:

[0215] To verify the effectiveness of this invention, a high-fidelity simulation verification environment was constructed based on a historical dataset of a busy port waterway. The region grid size was set to 100 m × 100 m, and the system fusion and operation cycle was 1 second.

[0216] (1) Data fusion performance verification.

[0217] Under simulated severe sea conditions (radar clutter density increased by 30%), the reliability-driven adaptive fusion method proposed in this invention is compared with the traditional Kalman filtering method based on fixed weights. The results are shown in Figures 2 and 3.

[0218] The track tracking effect is shown in Figure 2. Under strong clutter interference, the track generated by the traditional method (shown as the blue dashed line in Figure 2) exhibits obvious jumps and divergence. However, the method of this invention (shown as the red solid line in Figure 2) effectively suppresses outlier noise by dynamically weighting the observation reliability through MC-GNN. The generated fused track is smooth and closely follows the real track (shown as the black solid line in Figure 2).

[0219] The error convergence is shown in Figure 3. Compared to the large fluctuations in tracking error of traditional methods (shown by the blue solid line in Figure 3), the root mean square error of the position of the method of this invention (shown by the red solid line in Figure 3) is significantly reduced and maintained at a low level. Statistical results show that, under the same detection rate, the overall false alarm rate of the system is reduced by about 18%, which fully verifies the robustness of the method in complex environments.

[0220] (2) Verification of the effectiveness of the control strategy.

[0221] In a simulated complex multi-ship convergence scenario, the collaborative optimization control strategy generated in Section 3 of this invention was applied, and the experimental results are shown in Figures 4 and 5:

[0222] Figure 4 shows the evolution and avoidance of the risk field, with a heatmap illustrating the distribution of the high-risk core area. Without control measures (as shown by the dashed line in Figure 4), vessel trajectories directly pass through the high-risk area. After implementing control measures (as shown by the solid line in Figure 4), vessels automatically executed detour and speed limit commands before entering the high-risk area, successfully avoiding the peak collision risk. Monitoring data shows that the regional comprehensive risk index decreased by more than 42% compared to before control measures were implemented.

[0223] Figure 5 illustrates the multi-objective trade-off decision. The Pareto front curve calculated by the NSGA-II algorithm visually demonstrates the inverse relationship between the "air traffic efficiency loss rate" and the "regional comprehensive risk index." The final strategy automatically selected by the system shows that by accepting an air traffic efficiency loss rate of approximately 8.2%, the regional comprehensive risk index can be reduced to around 0.18, successfully controlling the efficiency loss within the expected target of 10%, thus achieving the optimal balance between navigation safety and efficiency.

[0224] The experimental data and visualization results above demonstrate that the method proposed in this invention can significantly improve the accuracy and robustness of maritime situational awareness while further achieving an effective trade-off among multiple objectives. Compared to the problem of disconnect between risk assessment and control in existing technologies, this invention achieves a leap from precise perception to intelligent closed-loop optimal control, significantly improving the intelligence and refinement of maritime area navigation safety management.

[0225] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a method for controlling and optimizing navigation safety in maritime areas, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0226] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MCU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0227] This invention provides a concept and method for the management and optimization of navigation safety in maritime areas. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for safety management and optimization of navigation in maritime areas, characterized in that, include: Step 1: Collect multi-source observation data of maritime navigation targets and preprocess them to obtain a spatiotemporal observation correlation map; Step 2: Use a multimodal consensus graph neural network to evaluate the dynamic reliability of each observation node; construct an adaptive weighted factor graph model based on the obtained dynamic reliability, and solve for the system state to obtain the optimal fused trajectory, corrected environmental parameters, and posterior error covariance matrix; Step 3: Construct a regional risk field based on the optimal fused trajectory, corrected environmental parameters, and posterior error covariance matrix; Step 4: Construct and solve a multi-objective collaborative optimization model based on the regional risk field to obtain navigation control parameters; Step 5: Execute the navigation control parameters and use multi-level closed-loop feedback for optimization to complete the maritime area navigation safety management and optimization; wherein, the solution for the system state in Step 2 includes: Step 2-1, in At time , the set of system state variables to be solved Defined as: ;in, Let be the ship's motion state vector. For the environmental state parameter vector; Step 2-2, set physical constraints, represented as factors in the factor diagram, as follows: Set motion constraints, i.e., motion factors, as follows: ;in, This represents a pre-defined kinematic model. The process noise is represented by the following formula: Set measurement constraints, or measurement factors, as follows: ;in, For sensors The observation function maps the system state to the observation space; To measure noise, its covariance matrix is: Steps 2-3 involve solving for the system state, i.e., performing an optimal estimation of the system state. This is equivalent to solving a nonlinear least squares problem, as follows: ;in, This indicates the calculation of Mahalanobis distance; The process noise covariance matrix represents the environmental state. For adaptive weights; in steps 2-4, the Gauss-Newton method is used to iteratively solve the nonlinear least squares problem in step 2-3. When the state update converges, the optimal estimate of the system state is obtained. That is, the fused state set, which contains the optimal fused track. and corrected environmental parameters Using the final Jacobian matrix With weight matrix Constructing an approximate Hessian matrix Its inverse matrix is ​​the posterior error covariance matrix. The construction of the regional risk field in step 3 includes: step 3-1, gridded risk source feature extraction, specifically as follows: discretizing the target sea area into a grid set. For each grid exist At any given moment, based on the fused state set Extracting multidimensional risk source feature vectors , means as follows: ;in, Indicates traffic density. For the intensity of the encounter, The trace represents the situational uncertainty, i.e., the posterior covariance matrix. ， Indicates the corrected environmental parameters The environmental stress degree is obtained after normalization and weighted fusion; Step 3-2, a partial differential equation is introduced as the dynamic evolution model of the regional risk field, as follows: ;in, Right now The mesh to be solved exist The overall risk intensity at any given moment It is the speed of risk convection. For two-dimensional spatial gradient operators, The risk diffusion coefficient, This is the risk natural decay coefficient. It is a risk source item; Step 3-3, discretize and solve the dynamic evolution model to predict the future. The dynamic distribution of regional risks within a given time period; the construction and solution of the multi-objective collaborative optimization model described in step 4 includes: step 4-1, defining the control decision variables of each grid as a vector. , means as follows: ;in, For grid speed limits, To designate the direction of navigation, Temporary no-fly zone sign; Step 4-2, construct multi-objective vector function , means as follows: ;in, For a set of global control strategies, For the sub-objective of total risk exposure, To reduce speed and lose sub-targets, To predict emissions or energy consumption sub-targets; step 4-3, an evolutionary algorithm is used to minimize the multi-objective vector function. Solving for the final strategy yields the solution. As navigation control parameters.

2. The method for maritime area navigation safety management and optimization according to claim 1, characterized in that, The preprocessing described in step 1 includes: Step 1-1, spatiotemporal alignment and data standardization, specifically as follows: Coordinated Universal Time (UTC) is used as the unified time base, with... A fixed fusion time grid is selected as the starting time. Second, For time indexing, the discretized time series is defined as follows: ; Any sensor on a maritime navigation target ,exist The observation data at each moment is used to compensate for the clock deviation between the sensor and Coordinated Universal Time (UTC), and linear interpolation is used to align the observation data to the most recent fusion time. All sensor observation data are transformed and unified into the Northeast ENU coordinate system with a preset origin; each observation data is represented as an observation tuple. The details are as follows: ;in, The aligned measurement state vector has dimensions. For its dimensions, This is the corresponding measurement noise covariance matrix. Let be the source quality feature vector of the sensor. Sensor type identification; Steps 1-2, consistency gating and observation map construction, are as follows: For observation data Calculate Mahalanobis distance Set threshold ,like If the observed value is an outlier, it is removed; otherwise, it is retained as a candidate observation. Within the sliding time window, a spatiotemporal observation correlation graph is constructed based on candidate observations. , means as follows: ;in, The node set consists of all observation tuples within the sliding time window. That is, the composition of observation nodes; For an edge set, for any two observed nodes and If the Euclidean distance between the observation locations is less than the set spatial neighborhood threshold or the observation data originates from the same navigation target, then an undirected edge is established. Connect the two, and set the attribute of each edge to the time difference between them. Spatial distance difference With relative speed difference 。 3. The method for maritime area navigation safety management and optimization according to claim 2, characterized in that, The adaptive weights mentioned in steps 2-3 are specifically calculated using the following method: Step 2-3-1, for the spatiotemporal observation correlation graph... Each node in Construct the initial feature vector of the node Step 2-3-2: A graph attention mechanism is used to aggregate neighborhood information, learn the mutual support relationships between nodes, and calculate the node... Dynamic credibility score Step 2-3-3: Establish a non-linear weight mapping function to assign dynamic credibility scores. Transformed into adaptive information adjustment factor Steps 2-3-4: Adjust the factor based on adaptive information. Constructing an adaptive measurement information matrix That is, adaptive weights.

4. The method for maritime area navigation safety management and optimization according to claim 3, characterized in that, The calculations described in step 2-3-2 yielded the nodes. Dynamic credibility score This includes: Step 2-3-2-1, for node Calculate its relationship with neighboring nodes Attention coefficient between The details are as follows: ;in, Indicates including nodes The set of neighboring nodes, including For shared feature projection matrices; This is the attention vector; This indicates vector concatenation; For linear units with leakage correction; The normalized exponential function is used; step 2-3-2-2, update the node features. , means as follows: ;in, The activation function is non-linear; in step 2-3-2-3, the output node is obtained through a fully connected layer and the Sigmoid function. Dynamic credibility score 。 5. The method for maritime area navigation safety management and optimization according to claim 4, characterized in that, Step 5 describes the use of multi-level closed-loop feedback for optimization, including: Step 5-1, real-time monitoring of the sea area status after command execution and calculation of the execution compliance rate. With risk improvement Step 5-2, based on the compliance rate With risk improvement Three-layer feedback control is implemented.

6. The method for maritime area navigation safety management and optimization according to claim 5, characterized in that, The calculation of the compliance rate described in step 5-1 With risk improvement ,include: ； ;in, The scope in which the instruction takes effect. Indicates an indicator function, For grid The average speed of ships inside the vessel To control parameter values, Allowable engineering errors; and These are the steps before and after the instruction is executed. Regional aggregated risk value over time.

7. The method for maritime area navigation safety management and optimization according to claim 6, characterized in that, The three-layer feedback control described in step 5-2 employs the following control logic: Communication layer feedback control: if the terminal broadcast coverage is less than a threshold, it is determined that the communication coverage is insufficient, the effective time of the navigation control parameter execution command is extended, and it is retransmitted; Execution layer feedback control: if the terminal broadcast coverage is greater than or equal to a threshold, it is determined that the communication coverage is normal, and the execution compliance rate is further assessed. If the execution rate is less than the threshold, it is determined to be low, triggering a preset flexible degradation strategy to generate and issue a new, downgraded instruction; the effect layer feedback control determines the execution compliance rate. If the value is greater than or equal to the threshold, then the degree of risk improvement is further assessed. If the value is negative, the current instruction will be immediately cancelled, and step 1 will be re-executed based on the latest multi-source observation data.

Images