An autonomous ship navigation risk propagation hierarchical control method fusing multi-source data

Abstract

The embodiment discloses a kind of autonomous ship seaworthiness risk transmission hierarchical control methods of fusion multi-source data, through establishing " defect-intermediate risk factor-seaworthiness target " three-layer node database, and fusion port state control and inspection data, expert score and machine learning model, construct directed weighted risk transmission network, and introduce the change rate discrimination of integrated toughness index based on node removal, weighted local gravity score and the key structure identification method of SIR transmission simulation combination, can accurately distinguish the object with many connections and big transmission influence, accurately locate the key node, key edge and key transmission path that really determine risk transfer;Finally, the method will mapping result be mapped to the hierarchical control priority of source prevention, transmission blocking, consequence mitigation for different ship types and seaworthiness stages, significantly improve the pertinence, executability and engineering practicability of autonomous ship seaworthiness risk control.

Description

Technical Field

[0001] The present invention relates to the technical fields of marine safety engineering, ship seaworthiness risk assessment, intelligent shipping and digital maritime supervision, and particularly relates to a layered control method for autonomous ship seaworthiness risk propagation integrating multi-source data. Background Art

[0002] With the gradual application of technologies such as autonomous navigation, intelligent perception, remote manipulation, ship-shore communication, shore-based support and cybersecurity, the objects of ship safety assurance have expanded from the single traditional on-board equipment and on-board personnel to a collaborative system of on-board systems, communication links, remote operation centers, and related personnel and organizations. The non-mandatory Code of Safety for Maritime Autonomous Surface Ships adopted by the International Maritime Organization also incorporates navigation, connectivity, remote operation, cybersecurity, remote operation centers and the like into a goal-based safety framework. In this context, the seaworthiness risks of autonomous ships present distinct characteristics of diversified sources, cross-system coupling and propagation along functional chains.

[0003] In the prior art, ship seaworthiness and operation safety evaluation usually adopt methods such as Port State Control inspection statistics, regulatory compliance review, equipment condition monitoring, failure mode and effects analysis, fault tree analysis, hazard and operability analysis, expert evaluation, and Bayesian networks. For autonomous ships, schemes have also been developed, including navigation environment perception and collision risk judgment based on sensor fusion, functional availability determination based on design operational scope or threshold rules, system hazard identification based on system-theoretic process analysis, probabilistic risk reasoning based on Bayesian networks, special assessment for remote operation and cybersecurity, and key factor ranking carried out according to indicators such as node degree, betweenness and closeness. Relevant patents and literatures can identify and evaluate single equipment failures, software failures, collision hazards, communication abnormalities, cybersecurity maturity or specific operation tasks, and some schemes can also update risk values according to real-time status and trigger alarms or control suggestions.

[0004] In general, most of the existing schemes take a certain piece of equipment, a certain function, a certain accident type or a certain operation link as the analysis object, and output a risk level, failure probability, hazard list or key factor ranking. However, the above technologies still have the following deficiencies in the analysis of autonomous ship seaworthiness risk propagation and the generation of control priorities: First, it is difficult to effectively supplement new propagation relationships under data scarcity conditions: traditional ship inspection data is relatively abundant, while actual inspection and long-term operation data for remotely controlled ships and fully autonomous ships are limited. Directly transplanting the relationships of traditional ships is prone to omitting new relationships such as autonomous navigation systems, ship-shore communication, remote operation centers, cybersecurity and human-machine collaboration; relying solely on expert evaluation is susceptible to the influence of sample size and subjective differences.

[0005] Second, multi-source heterogeneous risk factors lack a unified hierarchical expression: defects, equipment, functions, organizations and airworthiness targets formed by different methods usually lack a unified hierarchical expression, making it difficult to continuously explain which systems, functions or organizational interfaces the risk enters from a specific defect or abnormal state, through which it propagates, amplifies and ultimately affects which airworthiness capability; there is also a lack of a unified direction verification and source tracing mechanism between statistical co-occurrence, model prediction and expert judgment.

[0006] Third, a single static ranking is difficult to reflect the actual degree of the propagation impact: using a single centrality or static risk ranking is difficult to directly measure the actual impact of a weakened node, edge, or complete path on the scope of risk propagation and the accessibility of airworthiness targets. It is easy to simply equate "a large number of connections" with "a large propagation impact", and it is impossible to accurately identify the key structures that truly determine the transmission of risk from the source risk layer to the airworthiness target layer.

[0007] Fourth, the analysis results lack a mechanism for conversion into a tiered control strategy: existing analysis results mostly remain at the level of risk identification or ranking, without further being converted into tiered control targets and priorities for source prevention, transmission blocking, and consequence mitigation. This makes it difficult to directly support actual airworthiness risk management activities such as inspection, monitoring, maintenance, redundancy verification, functional isolation, and degraded operation. Summary of the Invention

[0008] This invention discloses a hierarchical control method for the propagation of seaworthiness risks of autonomous ships that integrates multi-source data, in order to overcome the above-mentioned technical problems.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows: A hierarchical control method for the propagation of seaworthiness risks in autonomous vessels that integrates multi-source data includes the following steps: S1: Acquire multi-source airworthiness risk data and preprocess it to establish a three-layer node database of defects-intermediate risk factors-airworthiness targets, including defect-type nodes, intermediate risk factor-type nodes, and airworthiness target-type nodes. S2: Construct a directed weighted risk propagation network for traditional vessels based on port state control and inspection data; S3: Obtain nodes based on autonomous vessels, and based on expert scores and a three-layer node database, obtain newly added directed candidate node pairs to obtain feature vectors of the newly added directed candidate node pairs. S4: Based on the feature vectors of the newly added directed candidate node pairs, a classification model is used to determine whether the newly added directed candidate node pairs have candidate edges; when the newly added directed candidate node pairs have candidate edges, a regression model is used to obtain the weights of the candidate edges; in order to construct a target network based on the target ship type and seaworthiness stage. S5: Based on the target network, construct a random node removal order and a target node removal order for performing node removal; and obtain the node removal order in the order of random node removal and the target node removal order respectively. The overall resilience index after node removal in the round; when the nodes are removed in the random order in the round... The overall resilience index after removing the target nodes is compared to the index after removing the target nodes in the order of removal. If the rate of change of the overall resilience index after wheel node removal exceeds a set threshold, execute S6; where, Remove the round index from the node; S6: Obtain the weighted local gravity score of the nodes in the target network to determine candidate nodes; remove each candidate node from the target network to obtain the comprehensive influence score of the candidate nodes, and then determine the key risk nodes; S7: Using the SIR propagation simulation model, candidate edges and propagation paths in the target network are determined; based on the comprehensive influence score of candidate edges, key edges are determined; then, based on the propagation path hit rate, the average time for defective nodes to propagate to seaworthiness target nodes, and whether the propagation path contains key edges, key propagation paths are determined; and then, based on key risk nodes, key edges, and key propagation paths, the propagation of seaworthiness risks of autonomous vessels is controlled.

[0010] Furthermore, the formula used to obtain the comprehensive resilience index is as follows:

[0011] In the formula: As a comprehensive resilience index; R WCC This represents the percentage of the largest weakly connected component that has been preserved relative to the state before the attack. R E This represents the percentage of the current global efficiency of the directed network that has been preserved relative to its state before the attack. The weight of the current largest weakly connected component relative to the retention rate before the attack; in,

[0012]

[0013] In the formula: For the first The largest weakly connected component of the network after round node removal; The largest weakly connected component of the original network when no nodes were removed; For the first Global efficiency after removing a round node; For the global efficiency of the original network; The number of elements in the set; Remove the round index from the node.

[0014] Furthermore, the formula used to obtain the weighted local gravity score of the node is as follows:

[0015] In the formula: The weighted local gravity score for node i; Let be the weighted connection strength of node i; Let be the weighted connection strength of node j; The radius of local action; This represents the shortest path distance between the two nodes. in,

[0016] In the formula: For nodes Pointing to node Edge weights; All of them are nodes in the network.

[0017] Furthermore, the formula used to obtain the comprehensive influence score of candidate nodes is as follows:

[0018] In the formula: The overall impact score of the candidate nodes; An index for evaluation indicators; This represents the total number of evaluation indicators used for comprehensive evaluation. For the first The weight of each evaluation indicator in the comprehensive calculation; This represents the relative decrease in the evaluation indicator; After removing node i, the first... The decrease in each evaluation indicator relative to the original network;

[0019] In the formula: M 0 represents the original network's metric value. M -i This represents the metric value after removing node i.

[0020] Furthermore, the evaluation metrics for candidate nodes include: maximum weakly connected component, network-wide reachability, path reachability from defective nodes to airworthiness target nodes, and strongest path efficiency.

[0021] Furthermore, the comprehensive influence score of the candidate edges is obtained as follows:

[0022] In the formula: The overall impact score of the candidate edges; ΔA e , ΔP e , ΔF e and ΔO e These represent the decrease in total propagation scale, activation peak, final impact range, and probability of large-scale propagation after removing candidate edge e, respectively. These are the influence weights for total propagation scale, activation peak, final impact range, and probability of widespread propagation, respectively.

[0023] Furthermore, the multi-source airworthiness risk data includes port state control inspection data for traditional vessels, regulations and technical data based on autonomous vessels, expert scoring data, and the target vessel type, functional configuration, and airworthiness phase configuration.

[0024] Beneficial Effects: This invention provides a hierarchical control method for the propagation of airworthiness risks in autonomous vessels that integrates multi-source data. By establishing a three-layer node database of "defects - intermediate risk factors - airworthiness targets" and integrating port state control inspection data, expert scoring, and machine learning models, it effectively completes the new propagation relationships under the condition of scarce actual operational data of autonomous vessels. This overcomes the problems of risk factor omission and relationship direction distortion caused by directly transplanting traditional vessel relationships. Simultaneously, by constructing a directed weighted risk propagation network, it unifies multi-source heterogeneous risk factors into a hierarchical expression, enabling a continuous and traceable description of the path of risk propagation from specific defects through system, function, or organizational interfaces to the airworthiness target. This solves the problem of the lack of a unified hierarchical expression and method in existing technologies. This addresses the shortcomings of existing verification mechanisms. Furthermore, by introducing a key structure identification method combining the rate of change of comprehensive resilience indicators based on node removal, weighted local gravity scoring, and SIR propagation simulation, it can accurately distinguish objects with a large number of connections and significant propagation impact, avoiding misjudgments caused by single static sorting. This allows for precise location of the key nodes, key connections, and key propagation paths that truly determine risk transmission. Finally, this method maps the identification results to hierarchical control priorities for source prevention, propagation blocking, and consequence mitigation for different ship types and seaworthiness stages. This enables the analysis conclusions to be directly transformed into practical management measures such as inspection, monitoring, maintenance, redundancy verification, and downgraded operation, significantly improving the pertinence, feasibility, and engineering practicality of autonomous ship seaworthiness risk control. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of the autonomous ship airworthiness risk propagation and control method based on multi-source data fusion according to the present invention.

[0027] Figure 2 This is a flowchart illustrating the overall process of constructing and hierarchically controlling the autonomous vessel airworthiness risk propagation network in this embodiment of the invention. Figure 3 This is a structural diagram of the autonomous vessel airworthiness risk propagation assessment and hierarchical control system in an embodiment of the present invention; Figure 4 This is a schematic diagram of the risk nodes and their connections in layers C, G, and B in this embodiment of the invention. Figure 5 This is a flowchart illustrating the machine learning completion process for autonomous vessel risk propagation relationships in this embodiment of the invention. Figure 6 This is a flowchart illustrating the identification of key risk nodes in an embodiment of the present invention; Figure 7 This describes the process for generating key connections, key propagation paths, and hierarchical risk control priorities in embodiments of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] This embodiment introduces a hierarchical control method for the propagation of seaworthiness risks in autonomous vessels that integrates multi-source data, including the following steps: Figure 1 As shown: S1: Acquire multi-source airworthiness risk data and preprocess it to establish a three-layer node database of defects, intermediate risk factors, and airworthiness targets; In this embodiment, the multi-source airworthiness risk data includes port state control inspection data for traditional vessels, regulatory and technical data based on autonomous vessels, expert scoring data, and the target vessel type, functional configuration, and airworthiness stage configuration. The port state control inspection data includes at least inspection number, defect code, defect grouping, and detention-related fields. The inspection data used in this embodiment is sourced from the Tokyo Memorandum of Understanding (TMI) port state control inspection data, primarily including inspection number, defect code, defect grouping, and detention-related information, used to construct a risk propagation network for traditional vessels. The regulatory and technical data based on autonomous vessels is used to supplement risk factors such as autonomous navigation, remote operation, ship-to-shore communication, situational awareness, cybersecurity, redundancy and recovery, and personnel and organizational support for remote operation centers.

[0030] Specifically, the supervision and inspection data undergoes preprocessing, including retaining the target inspection type, deleting missing or invalid defect codes and defect groups, standardizing the defect code format and retained fields, excluding records that cannot be converted, and processing duplicate data based on the inspection number and defect code. Subsequently, defect groups appearing in the same inspection are aggregated into a single transaction to provide a unified data input for subsequent association rule mining, as shown in Table 1.

[0031] Table 1. Multi-source airworthiness risk data and its uses

[0032] Specifically, after preprocessing, port state control (PSC) data with consistent fields, missing and invalid records are obtained, along with knowledge and expert data that can be used to supplement autonomous vessel nodes and relationships.

[0033] Specifically, this embodiment establishes a three-layer node database of defects, intermediate risk factors, and airworthiness targets, namely the C, G, and B three-layer node database, based on the preprocessed multi-source airworthiness risk data.

[0034] In this embodiment, the pre-processed specific defect codes and specific faults or abnormal states in the autonomous vessel are defined as Class C nodes, i.e., defect nodes in the three-layer node database, meaning specific defects, faults, or abnormal sources, such as PSC defect codes, sensor faults, communication anomalies, or network security anomalies, used to indicate the specific location where risks enter and spread through the network; defect groups and intermediate factors at the system, equipment, function, or organizational levels are defined as Class G nodes, i.e., intermediate risk factor nodes in the three-layer node database, including intermediate risk factors at the system, equipment, component, function, or organizational levels, used to indicate the transmission links of risks between systems, functions, or organizations; and the capabilities that need to be achieved or maintained, such as hull structural integrity, emergency facility performance, communication function integrity, situational awareness system integrity, autonomous navigation system integrity, and network security resilience, are defined as Class B nodes, i.e., airworthiness target nodes in the three-layer node database, i.e., the airworthiness targets that the vessel needs to achieve or maintain, used to indicate the airworthiness capabilities that risks may ultimately affect.

[0035] Specifically, each node must store at least the node number, node name, node type, functional category, applicable vessel type, applicable seaworthiness stage, and data source. For remotely controlled vessels and fully autonomous vessels, nodes suitable for the current target network are selected based on whether there are people on board, control method, functional configuration, and seaworthiness stage.

[0036] Through the unified numbering and attribute settings described above, data from different sources are converted into the same set of node objects. The same risk factor, regardless of whether it originates from inspection records, regulatory provisions, technical documents, or expert evaluations, is represented using a unified node number. When different sources describe node names, node categories, or scopes of application inconsistently, the standardized result is determined according to preset data source priorities, regulatory constraints, and manual review rules. The original name, mapping relationship, data source, and update time are retained, thereby avoiding duplicate modeling or relationship mismatches due to name differences and supporting subsequent result traceability.

[0037] The established three-tiered node database solves the problem of "how to represent different data using the same language". The output is a node database with unified numbering, type, applicable ship type, seaworthiness stage, and source information.

[0038] S2: Based on port state control and inspection data, construct a directed weighted risk propagation network for traditional vessels, including the following steps: S21: Calculate the C→G connection edge (defect-intermediate risk factor connection edge) and its weight: Establish the connection edge between the defect node and the intermediate risk factor node, and calculate the weight of the connection edge between the defect node and the intermediate risk factor node. Specifically, a C→G connection is established based on the correspondence between defect codes and defect groups, representing the directed relationship between a specific defect or anomaly source and its corresponding intermediate risk factor. This represents the process by which a specific risk enters the intermediate propagation layer. When the count of each defect code can be obtained from actual data or program records, the edge weight can be calculated based on the relative frequency of occurrence of that defect code within its corresponding defect group. Let the number of occurrences of the c-th node in class C be... n c The set of all C-class nodes that can be mapped to the g-th node of class G is: C g The edge weights from C to G are represented as follows:

[0039] In the formula, w (c,g) This represents the edge weight from the c-th defect node to the g-th intermediate risk factor node; n c The number of times the c-th defect node appears in the port state control inspection record; C g Let c be the set of all defect nodes mapped to the g-th intermediate risk factor node; c is... C g The index for intermediate defect nodes; g is the index for intermediate risk factor nodes; The larger the value, the more frequently the c-th type of defect node appears among the specific defects contained in this type of intermediate risk factor, and the more obvious its relative role as a risk entry point.

[0040] S22: Establish connections between intermediate risk factors; This embodiment utilizes Apriori association rules to establish G→G edges representing directed relationships between different intermediate risk factors, used to indicate coupling, amplification, or transfer between intermediate factors. Defects occurring in the same inspection are grouped as transactions, and Apriori association rules are used to mine statistical associations between G-class nodes. Association rules that meet threshold requirements are used as candidate G→G propagation relationships. Support reflects the frequency of two risk factors occurring simultaneously in all inspections; confidence reflects the conditional probability of the latter occurring simultaneously when the former occurs; lift determines whether the association is significantly higher than the random co-occurrence level under independent conditions. Association rules are used to provide candidate relationships and statistical directions, but are not the sole final basis for technical causality.

[0041]

[0042]

[0043]

[0044] In the formula: Support is the frequency with which X and Y co-occur in the same examination. This refers to the preceding term of the association rule, i.e., the set of risk factors that serve as the judgment condition; For the latter term of the association rule, that is, in After an event occurs, examine whether it is a set of risk factors that occur simultaneously. Let X be the probability that Y also occurs given that X has already occurred. Confidence level; The degree is used to indicate that the probability of X and Y occurring together is higher than the probability of them occurring independently; X and Y are both intermediate risk factor nodes.

[0045] Specifically, association rules that meet the minimum support, minimum confidence, and lift requirements are retained, and the confidence is used as the initial weight for G→G edges. In one embodiment, the minimum support is set to 0.01, the minimum confidence is set to 0.10, and the maximum itemset length is set to 2. If the same node pair simultaneously forms two rules, X→Y and Y→X, the confidence, lift, and support are compared sequentially to obtain the statistically dominant direction. Subsequently, the direction is verified based on the allowed network level, the order of device or functional action, regulatory and technical constraints, and expert opinions. Only when the statistical direction is consistent with at least one technical direction and there is no explicit evidence of negation is the direction determined as the G→G propagation direction; node pairs that are still indistinguishable are marked as relationships to be verified and do not directly enter the final target network.

[0046] S23: Establish the connection between the intermediate risk factor node and the airworthiness target node, and calculate the weight of the connection between the intermediate risk factor node and the airworthiness target node. Specifically, the G→B connection is used to represent the impact of intermediate risk factors on the airworthiness target, and to represent the process of intermediate risks converging towards the airworthiness target. First, based on the preset node mapping relationship, the B-type nodes that each G-type node can affect are determined. When the occurrence frequency and detention statistics can be verified, the product of the occurrence frequency and detention rate of each C-type node belonging to the G-type node and related to the airworthiness target B-type node can be accumulated. The following formula is an optional calculation method:

[0047] In the formula: This represents the weight of the edge connecting the g-th intermediate risk factor to the b-th airworthiness target node; Let C represent the set of C nodes that belong to the g-th G-class node and are related to the b-th B-class node; p c This indicates the frequency or percentage of occurrence of the c-th C-type node; dc This represents the percentage of inspection records containing the defect code of the c-th C-class node that are delayed; b is the index of the airworthiness target node.

[0048] Specifically, the larger the weight of the connection, the more frequently the corresponding intermediate risk factors occur, and the stronger their connection with more serious inspection consequences.

[0049] Using the above calculation method, we can simultaneously consider the frequency and severity of the consequences of risk factors, and avoid overestimating rare incidental problems based solely on the retention rate.

[0050] Specifically, the traditional ship directed weighted risk propagation network includes three types of connections: C→G, G→G, and G→B. These three types of connections are determined based on defect attribution, correlations in inspection records, and defect frequency and detention status, respectively, ensuring that each connection has a clear propagation direction and comparable relational weights. Specifically, C→G connects C-type nodes and G-type nodes, indicating that a specific defect, malfunction, or abnormal state enters the corresponding intermediate risk factor; G→G connects two G-type nodes, indicating that the risk continues to propagate, couple, or amplify between different intermediate risk factors; and G→B connects G-type nodes and B-type nodes, indicating that the intermediate risk factor further impacts the seaworthiness objective. The meaning is: a specific problem occurs → continues to propagate through intermediate system, equipment, functional, or organizational factors → ultimately affecting the seaworthiness capability that the ship needs to achieve or maintain. This embodiment addresses the problem of how to establish a basic propagation network using real traditional ship inspection records by establishing a directed weighted risk propagation network for traditional ships.

[0051] S24: Merge the three types of connections C→G, G→G, and G→B to form a directed weighted risk propagation network for traditional ships.

[0052]

[0053]

[0054] In the formula: This is the airworthiness risk propagation network, i.e., the directed weighted risk propagation network of traditional ships; V For a set of nodes, L For a set of directed edges, W Let A be the set of edge weights; A is the adjacency matrix of the traditional ship directed weighted risk propagation network. These are the adjacency matrix elements from node i to node j; This represents the total number of nodes in the network. Let be the weight of the edge connecting node i to node j; It is the starting node of the directed edge; For the directed weighted risk propagation network of traditional ships, it is the terminal node of the directed connection. Specifically, since the weights of the three types of edges C→G, G→G, and G→B come from different sources, in order to enable them to be calculated in the same network, all types of weights are uniformly mapped to the range of 0 to 1. This step is the edge normalization operation, and the calculation is shown below.

[0055]

[0056] In the formula: For normalized nodes Pointing to node The edge weights.

[0057] S3: Obtain intermediate risk factor nodes based on autonomous vessels, and based on expert scores, obtain newly added directed candidate node pairs to obtain feature vectors of the newly added directed candidate node pairs; form autonomous vessel candidate relationships and expert features.

[0058] Specifically, for remotely controlled and fully autonomous vessels, the process begins by supplementing the network with nodes such as autonomous navigation nodes, remote operation center nodes, ship-to-shore communication nodes, network security nodes, situational awareness nodes, redundant nodes, and recovery nodes, based on regulations, standards, relevant research, and the traditional directed weighted risk propagation network of vessels. Then, candidate G-type node pairs are generated within the permissible hierarchy and category range. Relationships already existing in the traditional directed weighted risk propagation network of vessels and still applicable to the target vessel can be retained as known relationships; when new nodes, new functions, or new organizational relationships are involved, the corresponding node pairs are treated as relationships to be predicted, as shown in Table 2.

[0059] Experts can evaluate the strength of the relationship between candidate node pairs, the confidence level of the judgment, and whether there is a direct impact. Since different node pairs may obtain different numbers of valid scores, this embodiment adopts a unified summarization method: for node pairs with only one valid score, that score is used as both the maximum and average score; for node pairs with multiple valid scores, the maximum and arithmetic mean are calculated separately. Blank scores are treated as missing information and are not directly regarded as 0; they are only regarded as unrelated information when the scoring rules explicitly state that 0 represents no relationship and the expert actually fills in 0.

[0060] For each candidate node pair with a starting point u and an ending point v, construct a feature vector. This includes the node category of the starting point of each newly added directed candidate node pair, the node category of the ending point of each pair, hierarchical relationships, and scene features. Categorical features are encoded, numerical features are scaled uniformly, and missing expert scores or confidence levels are marked with a missing label; blank scores are not replaced with a value of 0 to avoid misinterpreting "not evaluated" as "no relation." Candidate node pairs, feature vectors, data sources, and label status are stored in corresponding records, providing a unified input for training sample generation and prediction result verification.

[0061]

[0062]

[0063] In the formula: The maximum value among the valid scores given by experts for the u→v candidate relation; Index number for valid expert ratings; The number of valid expert scores obtained for each candidate node; Let be the score given by the r-th valid expert for the strength of the u→v relationship; u is the starting node of the edge formed by the candidate node pair; v is the ending node of the edge formed by the candidate node pair. This is the average of all valid expert scores for the candidate node pair.

[0064] Table 2. Feature Composition Table of Candidate Node Pairs

[0065] This step identifies which new risk factors have been introduced into the autonomous vessel and which node pairs warrant further evaluation. The output consists of directional candidate node pairs, their expert scores, hierarchical relationships (i.e., the relationships between candidate nodes in the three-layer database), and scenario characteristics.

[0066] S4: Based on the feature vectors of the newly added directed candidate node pairs, use a classification model to determine whether the newly added directed candidate node pairs have candidate edges. When a newly added pair of directed candidate nodes has a candidate edge, a regression model is used to obtain the weight of the candidate edge in order to construct a target network based on the target ship type and seaworthiness stage. This embodiment utilizes machine learning to infer the existence and weights of edges and verify the propagation direction: Directed node pairs with clearly defined relationships are used as training samples, while directed candidate node pairs with unclear relationships are used as prediction samples. First, a classification model determines whether candidate node pairs have edges, then a regression model estimates the relationship strength of the retained edges. The propagation direction is represented by the start and end points of the candidate node pairs and verified using an expert matrix, network hierarchy, and engineering rules. This avoids calculating weights for non-existent relationships and prevents directly interpreting model predictions as strict causal relationships.

[0067] The first task is edge existence classification. Features such as start-point category, end-point category, category combination, expert relationship score, and hierarchical relationship are input into the classification model to obtain the probability that directed candidate node pairs have edges. When this probability reaches a preset threshold and does not violate expert or rule constraints, the candidate edge is retained.

[0068]

[0069] In the formula: For the classification model, candidate edges → The final judgment result given; This is a function that evaluates to 1 if the condition inside the parentheses is true, and 0 otherwise. The existence labels of edges used when training the classification model; The feature vectors of the newly added directed candidate node pairs; This is the threshold for determining the existence of connected edges; The conditional probability output by the classification model.

[0070] The second task is edge weight regression. For edges that have been determined to exist and whose directions have passed verification, the edge strength is estimated based on features such as node type, hierarchical relationship, expert evaluation, and known network relationships.

[0071] In the formula: Candidate edges predicted by the regression model → The weights; This represents the regression model after training; θ represents the set of parameters determined after training the regression model. In this embodiment, the classification and regression models can employ CatBoost, or an automated machine learning method can be used for model selection. CatBoost is suitable for inputs that simultaneously include categorical features such as node categories and category combinations, as well as numerical features such as expert ratings and node attributes. Furthermore, it can reduce the risk of target information leakage through ordered category statistics and ordered boosting mechanisms. CatBoost is merely a preferred implementation and does not constitute a limitation on specific model brands, number of trees, tree depth, learning rate, or other hyperparameters.

[0072] In a preferred implementation, node pairs explicitly supported by regulations, technical knowledge, traditional known ship relationships, or valid expert opinions are marked as confirmed positive samples; node pairs with clear negative evidence are marked as confirmed negative samples; node pairs lacking only evaluation or evidence remain as unlabeled predicted samples and are not automatically set as negative samples due to "unevaluated". The edge existence classification label is 0 or 1, and the edge weight regression objective is the relative strength of known relationships after scaling.

[0073] The training and validation processes preferably preserve the sample source, labeling criteria, feature list, missing data markers, data version, model version, random seed, and threshold version. When multiple samples come from the same inspection, the same ship, or the same group of experts, they can be grouped into training and validation sets according to their data source to avoid samples from the same source entering both sets simultaneously. Edge existence determination threshold. It should be determined under verification data or pre-set engineering constraints.

[0074] Specifically, after the machine learning prediction is completed, the output relationships need to be verified according to rules. This verification includes at least: excluding relationships that violate C→G, G→G, or G→B level constraints; deleting inapplicable nodes and edges according to ship type and seaworthiness stage; excluding relationships explicitly rejected by experts or explicitly not supported by regulations and technical rules; manually reviewing relationships whose model probabilities are close to the judgment threshold or whose different criteria conflict; and limiting the final edge weights to a preset range. The verified edges represent potential and interpretable risk coupling channels, not the frequency of accidents already observed in autonomous vessels.

[0075] This embodiment addresses the problem of "how to help determine whether candidate relationships should be retained and how strong the relationships are when actual records are insufficient" by employing classification and regression models. The output is a list of candidate edges jointly verified by classification, regression, experts, and rules; this output represents modeled relationships, not actual accident probabilities.

[0076] The system acquires scenario conditions such as target vessel type, presence of personnel on board, functional configuration, and seaworthiness stage. It then selects applicable nodes and relationships from the node and edge databases to form a corresponding directed weighted risk propagation network. Each target network stores vessel type, functional configuration, seaworthiness stage, node source, edge source, and parameter configuration to differentiate between networks and verify the calculation basis. This embodiment does not limit the number of networks to be built simultaneously; a single network can be generated for a target vessel and a seaworthiness stage, or multiple networks can be generated for comparison.

[0077] In a complete application, seven risk propagation networks can be generated, including a traditional initial seaworthiness network, a manned remote-controlled ship initial and ongoing seaworthiness network, an unmanned remote-controlled ship initial and ongoing seaworthiness network, and a fully autonomous ship initial and ongoing seaworthiness network. These seven networks are used to illustrate the applicability of the present invention to different ship types and seaworthiness stages, and do not constitute a limitation on the scope of protection of this embodiment.

[0078] This step addresses the question of "how to generate networks suitable for specific ship types and stages from the same set of candidate data". The output is one or more target networks with clearly defined scope and source criteria.

[0079] S5: Based on the target network, construct a random node removal order and a target node removal order for performing node removal; and obtain the node removal order in the order of random node removal and the target node removal order respectively. The overall resilience index after node removal in the round; when the nodes are removed in the random order in the round... The overall resilience index after removing the target nodes is compared to the index after removing the target nodes in the order of removal. When the rate of change of the overall resilience index after the removal of a round node exceeds a set threshold, execute S6 to perform target network structure and attack response diagnosis.

[0080] This step addresses the question of whether risk propagation is maintained evenly across the entire network or concentrated on a small number of nodes. The output includes overall network structure metrics and resilience changes under random and targeted removal conditions.

[0081] Specifically, the system calculates metrics such as the number of nodes, number of edges, network density, average degree, average weight, total weight, clustering coefficient, and directed reachability of the target risk propagation network to characterize its overall structure. Then, random node removal and target node removal are performed. Random node removal removes nodes in a random order to simulate non-selective node failure; target node removal prioritizes nodes with significant structural impact based on weighted node strength and other metrics to verify whether the network relies on a few critical locations for propagation. This "attack" is a computational perturbation operation in complex network analysis, only altering a copy of the network being analyzed and not sending attack commands to actual ship equipment, communication links, or control systems.

[0082] Preferably, after each round of node removal, the retention ratio of the maximum weakly connected component and the retention ratio of the global efficiency are calculated separately and combined into a comprehensive resilience index. The formula used to obtain the comprehensive resilience index is as follows:

[0083] In the formula: As a comprehensive resilience index; R WCC This represents the percentage of the largest weakly connected component that has been preserved relative to the state before the attack. R E This represents the percentage of the current global efficiency of the directed network that has been preserved relative to its state before the attack. This represents the weight of the current largest weakly connected component relative to the retention rate before the attack. .

[0084] When the overall resilience decreases significantly faster under targeted attacks than under random attacks, it indicates that the network's propagation capability is mainly maintained by a few structural units, requiring further identification of key risk nodes and critical connections. Otherwise, the network construction parameters should be reassessed. In this case, insensitivity to resilience may indicate a problem with the network construction itself, requiring backtracking and checking: whether the edge weights are too uniform; if all edge weights are similar, it indicates a lack of hierarchical structure, requiring checks on the rationality of weight calculations and node classification; whether the C / G / B layer partitioning is accurate; whether there are a large number of redundant nodes; and whether candidate edges are too dense, etc.

[0085] Specifically, the maximum weakly connected component retention ratio and the global efficiency retention ratio can be calculated using the following formulas:

[0086]

[0087] In the formula: For the first The largest weakly connected component of the network after round node removal; The largest weakly connected component of the original network when no nodes were removed; For the first Global efficiency after removing a round node; For the global efficiency of the original network; The number of elements in the set; Remove the round index from the node.

[0088] Specifically, for directed networks, global efficiency is optimized by calculating the directed shortest path according to the following formula. If there is no directed reachable path between node i and node j, then the contribution of this item is recorded as 0.

[0089]

[0090] In the formula: For the global efficiency of directed networks; This represents the total number of nodes in the network. This represents the shortest path distance between the two nodes.

[0091] Specifically, when WCC 0 | or E When the value is 0, the corresponding retention ratio is recorded as 0 or removed from the current comprehensive resilience calculation, and the remaining weights are renormalized. The weight values ​​should be saved along with the airworthiness phase and assessment purpose.

[0092] S6: Obtain the weighted local gravity score of the nodes in the target network to determine candidate nodes; remove each candidate node from the target network to obtain the relative decrease in the evaluation index before and after the removal of the candidate node, so as to obtain the comprehensive impact score of the candidate node; to identify key risk nodes.

[0093] This step addresses the question of "which nodes, once controlled, are most likely to weaken the propagation from layer C to layer B". The output is a candidate list of C, G, and B nodes and a ranking based on their combined impact.

[0094] First, a weighted local gravity model is used to screen candidate nodes. This model considers the node's own weighted connectivity strength, the importance of neighboring nodes, and the network distance between them. The stronger the connection and the closer the node is to its surrounding important nodes, the greater its structural influence within the local area. Nodes whose weighted local gravity exceeds a set threshold are selected as candidate nodes for subsequent calculations.

[0095] Preferably, the formula used to obtain the weighted local gravity score of a node is as follows:

[0096] In the formula: The weighted local gravity score for node i; Let be the weighted connection strength of node i; Let be the weighted connection strength of node j; The radius of local action; This represents the shortest path distance between the two nodes. Specifically, the weighted local gravity model is used to narrow down the scope of subsequent calculations and is not directly used as the basis for the final ranking of critical risk nodes.

[0097] In one implementation, the weighted connection strength of node i is calculated as the sum of the weights of the incoming and outgoing edges, i.e.:

[0098] In the formula: For nodes Pointing to node Edge weights; Determined based on the directed shortest path in the target network. Alternatively, two sets of candidate nodes can be formed using inflow and outflow strengths respectively, but the criteria should be consistent within the same sorting process.

[0099] In this embodiment, the evaluation metrics for candidate nodes include: maximum weakly connected component, network-wide reachability, path reachability from defective nodes to airworthiness target nodes, and strongest path efficiency.

[0100] Specifically, to measure the ability of risk to propagate along a complete path, the path... P The strength of the reachability, i.e., the path reachability, is represented by the product of the normalized weights of the edges along the path:

[0101] In the formula: Let P be the path strength and path reachability of the propagation path. This is a propagation path formed by connecting directed edges sequentially. For directed edges on the path → The edge weights used for path calculation after normalization; Based on this, the strongest paths from each Category C risk entry point to each Category B airworthiness target are calculated, and their average value is taken as the efficiency of the strongest path from C to B:

[0102] In the formula: The optimal path efficiency from layer C to layer B; It is the set of all C-type source risk nodes in the original network; It is the set of all Class B airworthiness target nodes in the original network; For set C A C-class node in 0; For set B A B-class node in 0; This represents a candidate directed propagation path from node c to node b. Candidate paths Path strength; Specifically, for each candidate node i, a single node removal is performed, and the changes in the maximum weakly connected component, overall network reachability, C-to-B reachability, and the efficiency of the strongest path from C to B before and after the removal are calculated. Any evaluation metric... M The relative decrease is expressed as:

[0103] In the formula: The relative decrease in evaluation metric M after removing candidate node i; M 0 represents the original network's metric value. M -i This represents the metric value after removing node i.

[0104] The decreases in each indicator are normalized and weighted to obtain the overall impact score of each node:

[0105] In the formula: The comprehensive influence score of candidate node i; An index for evaluation indicators; This represents the total number of evaluation indicators used for comprehensive evaluation. For the first The weight of each evaluation indicator in the comprehensive calculation; This represents the relative decrease in the evaluation indicator; After removing node i, the first... The decrease in each evaluation indicator relative to the original network; These are the candidate nodes for the current evaluation; Specifically, a higher overall impact score for a node indicates a more significant decrease in network connectivity or propagation capability from layer C to layer B after the node is removed. The final output simultaneously shows the overall ranking of all nodes, as well as the classification rankings of C-class, G-class, and B-class nodes, thereby determining the control order of risk entry points, intermediate propagation locations, and target convergence locations. When the original evaluation indicators... When the value is 0, the corresponding relative decrease is set to 0 or the indicator is removed from the current weighted calculation, and the weights of the remaining indicators are renormalized; when there is no reachable path between C-type nodes and B-type nodes, the corresponding strongest path strength is recorded as 0, in order to avoid the impact of division by zero or no path on the stability of the calculation.

[0106] S7: Using the SIR propagation simulation model, candidate edges and propagation paths in the target network are determined; based on the comprehensive influence score of candidate edges, key edges are determined; then, based on the propagation path hit rate, the average time for defective nodes to propagate to seaworthiness target nodes, and whether the propagation path contains key edges, key propagation paths are determined; and then, based on key risk nodes, key edges, and key propagation paths, the propagation of seaworthiness risks of autonomous vessels is controlled.

[0107] This step addresses the question of "along which interfaces and along which complete chains does the risk primarily propagate?" The output consists of key G→G and G→B edges, as well as representative paths from C-type risk entry points to B-type airworthiness targets.

[0108] Specifically, a SIR propagation simulation is performed on the target directed weighted network to determine the state of nodes in the target network. The inactive state S indicates that a risk factor has not yet been affected by upstream risks but may still be activated; the activated state I indicates that the risk factor has been triggered and can continue to propagate downstream; and the controlled state R indicates that the risk factor has been controlled through inspection, maintenance, isolation, or other measures and no longer propagates outwards. These states correspond to the susceptible, infected, and recovered states in the classic SIR model, respectively.

[0109] Specifically, for a node j in an inactive state, the probability that it will be activated at propagation step t due to the combined effect of all activated upstream nodes is expressed as:

[0110] In the formula: Let be the probability that node j is activated by an already activated upstream node in the t-th propagation step; Let J be the set of upstream neighbors of node j; This refers to the overall propagation intensity parameter; Let be the weight of the edge connecting node i to node j; Specifically, activated nodes are classified according to their recovery or control probability. γ The process transitions to a controlled state. To reduce fluctuations caused by a single random simulation, Monte Carlo simulations are repeated multiple times for each set of propagation parameters and initial disturbance conditions.

[0111] Wherein, the propagation intensity parameter λ , restore or control probability γ The initial set of activated nodes and the upper limit of propagation steps can be configured according to the ship type, seaworthiness stage, historical inspection results, or verification scenario. For the same candidate connection, the propagation results before and after intervention are compared under the same parameter combination, initial disturbance scale, and number of simulations. By fixing the random seed set or increasing the number of repetitions, random fluctuations are reduced, making the comprehensive impact scores of different candidate connections comparable.

[0112] Candidate edges The "weakening" can be achieved through Implement, where 0≤ η ≤1; η =0 is equivalent to removing the edge, 0< η <1 indicates a proportional reduction in transmission intensity. This indicates no intervention. When comparing different candidate edges, the same method is used. η The system consists of a random seed set, an initial set of activated nodes, and a combination of propagation parameters. The weight of the link after intervention (weakening) is the actual propagation strength of the link after the intervention measures are applied; The weakening coefficient (intervention intensity) has a value range of 0 ≤ η ≤ 1, representing the scaling ratio of the original weights. η = 0 means the edge is completely removed; 0 < η < 1 means the propagation intensity is reduced proportionally; η = 1 means no intervention. This represents the original edge weight, i.e., the original propagation strength of the edge when no intervention is applied; Under the same propagation strength parameters, the propagation results were calculated for the original target network and after removing or weakening candidate edges one by one. Candidate edges mainly included G→G and G→B edges. The comprehensive influence score of candidate edges was calculated by comparing the total propagation scale, activation peak, final impact range, and the decrease in the probability of large-scale propagation.

[0113] In the formula: The overall impact score of the candidate edges; ΔA e , ΔP e , ΔF e and ΔO e These represent the decrease in total propagation scale, activation peak, final impact range, and probability of large-scale propagation after removing candidate edge e, respectively. These are the influence weights for total propagation scale, activation peak, final impact range, and probability of widespread propagation, respectively.

[0114] Specifically, the higher the overall impact score of a candidate edge, the more significant its role in maintaining risk propagation. When the overall impact score of a candidate edge exceeds a set threshold, it is identified as a key edge.

[0115] Specifically, a widespread propagation event can be defined as the cumulative number of activated nodes or its proportion to the total number of nodes in the target network reaching a preset threshold at the end of a single simulation; the probability of widespread propagation is the ratio of the number of simulations that meet this condition to the total number of simulations. The threshold is configured according to ship type, seaworthiness stage, or verification scenario, and is not fixed to a specific value when experimental records are not obtained. Weights of each evaluation indicator are as follows. ω A , ω P , ω F and ω O It is a non-negative number, and the denominator is not 0.

[0116] In each propagation simulation, the node activation time, upstream source, and path hit count are recorded. Following the chronological order of propagation, the C-class risk entry point, the sequentially activated G-class intermediate risk factors, and the finally reached B-class airworthiness target are connected to form a complete C→G→…→B propagation path. Path hit rate is expressed as:

[0117] In the formula: The path hit rate of propagation path P; Indicates the propagation path P The number of times recorded throughout the entire simulation; This indicates the total number of simulations.

[0118] By combining path hit rate, average time to reach layer B, and whether the path contains critical edges, representative critical propagation paths are determined.

[0119] Specifically, when the same node is affected by multiple activated upstream nodes in the same propagation step, all the earliest upstream sources can be retained, or a primary upstream source can be determined based on the transition probability contribution of the current step. The path hit rate, average arrival time, and critical propagation path determination should use the same source tracing strategy, and this strategy should be written into the simulation metadata along with the random seed.

[0120] Furthermore, key risk nodes, key connections, and key propagation paths are transformed into structured, hierarchical risk control results. This addresses the question of "how to translate calculation results into an executable control sequence." The output consists of three categories of control objects: source prevention, propagation blocking, and consequence mitigation, along with their priorities. In this embodiment, hierarchical risk control is a process of classifying control objects into source prevention, propagation blocking, and consequence mitigation based on the position and overall impact of key objects in the propagation chain, and transforming network analysis results into priorities for inspection, monitoring, maintenance, and security control.

[0121] Specifically, if the controlled object is a Class C node, the starting point of a critical propagation path, or a high-frequency, high-severity risk entry point, it is mapped as a source prevention object; if the controlled object is a Class G node, a critical G→G link, or an intermediate propagation interface in a critical propagation path, it is mapped as a propagation blocking object; if the controlled object is a Class B node, a critical G→B link, or the end point of a critical propagation path, it is mapped as a consequence mitigation object. When the same object simultaneously meets multiple level conditions, the main control level is determined based on its node or link comprehensive impact score, path hit rate, airworthiness target severity, and preset safety priority rules, while other associated control levels are retained, as shown in Table 3.

[0122] Table 3 Output Fields of Hierarchical Risk Control Results

[0123] Specifically, risk control targets are ranked according to their comprehensive impact score (ranked by the comprehensive impact score of candidate nodes or candidate edges), and can be divided into Level 1, Level 2, and Level 3 based on preset thresholds or quantiles. Level 1 targets are prioritized for inclusion in design review, experimental verification, continuous monitoring, or maintenance lists; Level 2 targets are included in periodic inspections and status assessments; and Level 3 targets are stored as general risk information.

[0124] The system output includes at least the object number, name, object type, applicable vessel type, seaworthiness stage, functional or interface correspondence, scenario conditions, control level, priority level, comprehensive impact score, recommended measures, evidence source, data version, model version, parameter set, and output time. It can also output structured records to at least one of the following modules: seaworthiness inspection management module, status monitoring module, maintenance management module, or risk warning module. Without an authorized execution interface and independent safety verification mechanism, the output serves only as a technical decision support record and does not directly constitute an automatic takeover, isolation, or downgrade instruction for vessel equipment.

[0125] Example 1: Construction of a Traditional Initial Seaworthiness Risk Propagation Network for Ships This embodiment uses traditional port state control inspection data to illustrate the construction and hierarchical control method for the autonomous ship seaworthiness risk propagation network. The application scenario, data time range, number of networks, node numbers, and parameter values ​​are only for illustrating the implementation of this invention and do not constitute a limitation on the data source, target ship type, seaworthiness stage, network size, or model parameters. This embodiment uses Tokyo Memorandum port state control inspection data to construct a traditional ship initial seaworthiness risk propagation network. The original data covers January 2021 to June 2025, containing a total of 514,305 inspection defect records. After retaining the initial inspection records, 341,915 data points are obtained. After further deleting missing, invalid, or non-standardized key fields, 284,298 valid records usable for network construction are obtained.

[0126] The retained data fields include inspection number, defect code, defect group, and detention-related fields. Defect codes are used to generate C-class nodes, defect groups are used to generate G-class nodes, and preset airworthiness targets are used to generate B-class nodes. This embodiment uses the following processing logic: constructing C→G edges based on defect attribution relationships, constructing G→G edges using association rules, and constructing G→B edges based on occurrence frequency and detention-related statistics. Defect groups under the same inspection number are aggregated into transactions, and Apriori association rules with a minimum support of 0.01, a minimum confidence of 0.10, and a maximum itemset length of 2 are used.

[0127] The resulting traditional ship risk propagation network comprises 615 nodes and 781 directed weighted edges, as shown in Table 4. This network serves two purposes: firstly, it analyzes risk propagation relationships in traditional ships; secondly, it provides empirical reference for node setup, known relationships, and training samples in networks for remotely controlled and fully autonomous ships. Subsequent networks are not simply replicas of the traditional network; rather, they supplement nodes and propagation relationships based on newly added functions, ship-shore relationships, and organizational structures of autonomous ships.

[0128] Table 4 Data Processing Results of Traditional Ship Network Construction

[0129] Example 2: Constructing a risk propagation network for remotely controlled and fully autonomous vessels: Specifically, remotely controlled and fully autonomous vessels currently lack sufficient port state control (PSC) inspection data. To address this, this embodiment supplements the data with 259 risk nodes from regulations, technical guidelines, and related research. These include 16 Category B airworthiness target nodes, 37 Category G intermediate risk factor nodes, and 206 Category C specific defects, malfunctions, or anomaly sources nodes. These nodes are categorized into 12 functional groups, covering hull structure, mechanical propulsion, electrical systems, automatic control, ship-to-shore communication, remote operations centers, cybersecurity, situational awareness, cargo handling, navigation and collision avoidance, redundancy recovery, and other factors.

[0130] Experts evaluate the relationships between node categories. For node pairs with only one valid expert rating, the maximum and average ratings are both taken from that rating. For node pairs with multiple expert ratings, the maximum and arithmetic mean are calculated separately for each expert, while retaining both expert confidence and direct influence judgments. Node pairs with clearly defined relationships are used as training samples, while newly added nodes with unclear relationships are used as prediction targets.

[0131] First, a classification model is used to predict whether there are edges between directed candidate node pairs. Then, a regression model is used to estimate the weights of the retained edges. The start and end points of the candidate node pairs are used to represent the propagation direction. The final direction needs to be verified according to the expert relationship matrix, node level, regulations and technical constraints, ship type, and seaworthiness stage. The model can use CatBoost or automated machine learning methods.

[0132] By configuring the target vessel type and seaworthiness stage, initial and ongoing seaworthiness networks are established for manned remotely controlled vessels, unmanned remotely controlled vessels, and fully autonomous vessels. Together with the traditional vessel network, this forms a total of seven application networks.

[0133] The six autonomous vessel networks described above were constructed under conditions of actual operation and scarce PSC data. The expert relationship matrix was used to constrain the engineering rationality of candidate relationships, and machine learning was used to supplement edge existence and weights. Therefore, these networks are suitable for mechanism analysis, control position selection, and subsequent verification design, but should not be directly used as a database of real-world accident probabilities or equipment failure rates for autonomous vessels.

[0134] Example 3: Identification of Key Risk Nodes Random and targeted attacks were conducted on seven application networks. Under random attacks, network performance degraded relatively slowly, while under targeted attacks, network performance degraded rapidly after a small number of high-impact nodes were removed. This indicates that the risk propagation capability is not evenly distributed across all nodes, but rather relies more heavily on a few key structures.

[0135] Based on the above diagnosis, a weighted local gravity model was first used to screen candidate nodes, then candidate nodes were removed one by one, and the decrease in the maximum weakly connected component, overall network reachability, C-to-B reachability, and strongest path efficiency was calculated. The results show that the impact of G-layer and B-layer nodes on propagation continuity and target layer convergence is generally higher than that of a single C-layer node.

[0136] In traditional vessel networks, nodes such as B07, B08, B02, G01, G04, G13, G03, and G12 rank highly, indicating that traditional risks are mainly concentrated in fire safety, watertightness and weathertightness, cargo and pollution prevention, navigation safety, and deficiencies in traditional equipment and operations. In remotely controlled and fully autonomous vessel networks, B27 and G45 repeatedly rank highly, demonstrating that the completeness of autonomous navigation systems and the information foundation such as navigation sensors, positioning, and electronic charts constitute relatively stable points of convergence and intermediate propagation of risks.

[0137] C14402 consistently ranks high in the C-layer rankings of the seven application networks, indicating that some traditional defects remain relatively common entry points for risk. However, from the perspective of overall propagation impact, such source risks usually need to enter the G-layer and be transmitted or amplified by intermediate factors before having a more significant impact on the airworthiness objectives at the B-layer.

[0138] Example 4: Identification of critical edges and critical propagation paths: For each application network, multiple sets of propagation strength parameters, recovery parameters, and initial perturbation conditions at layer C are set. Each parameter combination is repeatedly run 500 Monte Carlo simulations, with a maximum of 100 propagation steps per run. Subsequently, candidate G→G and G→B edges are removed sequentially, and the changes in propagation scale, activation peak, final impact range, and large-scale propagation probability are compared.

[0139] In the traditional ship network, G04→B02 is the critical link with the highest overall impact ranking. G→B links such as G02→B07, G01→B07, and G01→B03 are also among the top, indicating that the risks of traditional ships often directly affect seaworthiness targets such as hull structure, emergency facilities, and fire protection structure through intermediate defect groups.

[0140] Among the six autonomous vessel application networks, G20→G08, G09→G07, G09→G08, and G45→B27 consistently rank in the top four. The first three G→G connections form a recurring intermediate propagation framework, while G45→B27 represents a critical exit point where risks converge from navigation sensors, positioning, and electronic chart configurations to the functional integrity of the autonomous navigation system.

[0141] Based on the activation time and upstream source of each node, the complete propagation path can be further reconstructed. The representative path in traditional ships is usually relatively short, mainly consisting of "specific defects - traditional intermediate factors - seaworthiness targets"; the representative path of remotely controlled ships begins to include factors such as remote operator qualifications, certification documents, visual perception, and situational awareness; the representative path of fully autonomous ships further emphasizes factors such as the qualifications of remote operation center monitoring personnel, network anomaly monitoring, audit logs, and network security resilience.

[0142] The results of this embodiment further show that the average comprehensive impact score of the highest-ranking critical link G04→B02 for traditional vessels is 0.002006; among the six autonomous vessel networks, the average comprehensive impact score of G20→G08 ranges from 0.007147 to 0.008638, and the score range of G45→B27 ranges from 0.001626 to 0.001680. The average hit rate of the leading representative path is mainly between 0.0060 and 0.0100, with an average arrival time of approximately 3.33 to 11.69 propagation steps. This indicates that risk propagation has probabilistic and multi-path characteristics, and the critical propagation path is used to represent the typical channels of priority concern, rather than a fixed accident chain that is bound to occur every time.

[0143] Example 5: Generating tiered risk control priorities: The system generates three types of control outcomes—source prevention, propagation blocking, and consequence mitigation—based on the role of key risk nodes, key links, and key propagation paths in the propagation chain. For traditional ships, source prevention focuses on structural condition, fire safety, life-saving equipment, pollution prevention, dangerous goods, and certificate defects; propagation blocking focuses on intermediate nodes such as G01, G04, G03, and G09 and their key links; and consequence mitigation focuses on seaworthiness targets such as B07, B08, and B02.

[0144] For remotely controlled vessels with a crew, the control focus is on coordination among onboard personnel, remote support, and autonomous systems, which may include control transfer, remote support procedures, emergency power, software updates, and ship-shore joint exercises. For remotely controlled vessels without a crew, the control focus is on communication redundancy, remote watchkeeping, remote takeover, degraded operation, and the capabilities of personnel in the remote operations center. For fully autonomous vessels, the control focus is on the reliability of autonomous systems, cybersecurity resilience, situational awareness reliability, and the monitoring capabilities of the remote operations center.

[0145] From the perspective of the airworthiness phase, the initial airworthiness phase mainly outputs the priorities of design review, functional testing, redundancy verification, and verification experiments; the continuous airworthiness phase mainly outputs the priorities of dynamic monitoring, degradation control, remote intervention, local isolation, degraded operation, return-to-port maintenance, and operating condition restrictions. The system outputs the above-mentioned hierarchical control results, along with the object number, calculation basis, applicable ship type, airworthiness phase, and recommended measures, to the corresponding inspection, monitoring, maintenance, or risk warning modules, as shown in Table 5. It should be noted that this embodiment completes the propagation structure analysis based on existing data, knowledge, expert evaluation, and model inference; the continuous calibration network using real autonomous ship operation records is an extended application after obtaining the corresponding data and is not considered a result verified in this embodiment.

[0146] Table 5. Layered Risk Control Mapping Table

[0147] Specifically, Figure 2 This document illustrates the overall execution flow of the autonomous vessel seaworthiness risk propagation network construction and hierarchical control method in this embodiment. The process takes traditional vessel port state control inspection data, autonomous vessel regulations and technical documents, expert scoring data, and vessel type and seaworthiness stage configuration as inputs. It sequentially performs multi-source data preprocessing, constructs Class C source risk nodes, Class G intermediate risk factor nodes, and Class B seaworthiness target nodes, establishes a directed weighted risk propagation network for traditional vessels, generates candidate propagation relationships and expert characteristics for autonomous vessels, uses machine learning to predict the existence of candidate edges and estimate edge weights, and verifies the propagation direction by combining expert matrices, hierarchical relationships, and engineering rules. It generates a target network according to the target vessel type and seaworthiness stage, identifies key nodes, key edges, and key propagation paths, and finally outputs risk control priorities at three levels: source prevention, propagation blocking, and consequence mitigation. For traditional vessels, propagation relationships are mainly supported by actual inspection data; for remotely controlled vessels and fully autonomous vessels, newly added relationships represent potential coupling channels supported by knowledge, experts, and models, and do not represent the observed accident probability or inspection frequency.

[0148] Figure 3 This document illustrates the architecture of the autonomous vessel seaworthiness risk propagation assessment and hierarchical control system of this embodiment. The system includes an input data layer, a processing and calculation layer, and a result output layer. The input data layer acquires port state control inspection records, regulations and technical knowledge, expert scores, vessel type, functional configuration, and seaworthiness stage. The processing and calculation layer performs data preprocessing, node and edge database construction, traditional vessel network construction, autonomous vessel relationship inference, target network generation, network structure diagnosis, key node identification, propagation simulation, key edge and propagation path identification, and control priority calculation. The result output layer outputs the target risk propagation network, key node list, key edge list, key propagation path, hierarchical control objects, control priority levels, and recommended measures. These functional modules can be centrally deployed on a single device or collaboratively implemented by shipborne computing equipment, a shore-based remote operation center server, a maritime regulatory analysis server, and an independent risk assessment platform.

[0149] Figure 4This embodiment illustrates the hierarchical relationships between risk nodes in layers C, G, and B, and the three types of directed edges. Layer C represents risk entry points such as specific defects, faults, or abnormal states; layer G represents intermediate risk factors at the system, equipment, functional, or organizational levels; and layer B represents the seaworthiness objectives that the ship needs to achieve or maintain. C→G edges represent the propagation relationship of specific risks into corresponding intermediate risk factors; G→G edges represent the coupling, amplification, or transfer relationship of risks between different intermediate factors; and G→B edges represent the convergence relationship of intermediate risk factors acting on seaworthiness objectives. Through this three-layer structure, this invention transforms "where the risk originates—through which factors it propagates—and ultimately affects which seaworthiness capability" into explicit network objects that can be numbered, weighted, calculated, and traced, providing a unified data foundation for subsequent key structure identification and hierarchical control mapping.

[0150] Figure 5 This embodiment demonstrates the machine learning-assisted construction process for understanding the risk propagation relationships between remotely controlled and fully autonomous vessels. The system first generates directed candidate node pairs based on node category combinations, permissible hierarchical directions, regulatory and technical constraints, known relationships between traditional vessels, and expert evaluations, extracting features such as node category, hierarchical relationships, and expert scores. Subsequently, a classification model is used to predict whether candidate node pairs will form connections, and a regression model is used to estimate the weights of the retained connections. The starting and ending points of candidate node pairs already contain directional information; the final direction still needs to be verified using an expert relationship matrix, network hierarchy, vessel type, seaworthiness stage, and engineering rationality. The machine learning output is not directly equivalent to the actual causal relationship of an accident, but rather serves to supplement potential propagation channels with engineering rationality and interpretability under conditions of data scarcity.

[0151] Figure 6 The document illustrates the identification process for key risk nodes in this embodiment. The system first performs analytical network disturbance diagnosis through random node removal and target node removal to determine whether the network propagation capability is concentrated on a few structural units; the node removal is only for computational analysis and does not represent an attack on the actual ship system. Subsequently, a weighted local gravity model is used to screen candidate nodes with strong local influence, and these candidate nodes are removed one by one. The decrease in the maximum weakly connected component, overall network reachability, C-layer to B-layer reachability, and the efficiency of the strongest path from C-layer to B-layer are compared before and after removal. A comprehensive impact score for each node is calculated based on the decrease in each indicator, and the classification rankings of C-, G-, and B-class nodes are output respectively. The more significant the decrease in propagation capability caused by node removal, the more suitable the node is as a priority control target for source prevention, propagation blocking, or target capability protection.

[0152] Figure 7This embodiment illustrates the generation process of critical links, critical propagation paths, and hierarchical risk control priorities. The system repeatedly executes risk propagation simulations under the same network, propagation parameters, and initial disturbance conditions, removing or weakening candidate G→G and G→B links respectively. The total propagation scale, activation peak, final impact range, and large-scale propagation probability are compared before and after intervention. Critical links are determined based on the reduction in each factor. The system further records node activation time, upstream source, and path hit count, embedding the critical links into the complete propagation path from a Class C risk entry point through Class G intermediate factors to a Class B airworthiness target. Finally, based on the critical object's position in the propagation chain and its comprehensive impact score, Class G nodes and G→G links are mapped to propagation blocking objects, and Class B nodes and G→B links are mapped to consequence mitigation objects. The system then outputs the control priority level, applicable ship type, airworthiness stage, and recommended control measures.

[0153] The method in this embodiment can be executed by shipborne computing equipment, shore-based remote operation center server, maritime regulatory analysis server, or independent risk assessment platform, following the sequence of multi-source data acquisition and preprocessing—hierarchical node database construction—traditional ship network construction—autonomous ship candidate relationship generation—machine learning-assisted relationship inference and verification—target network generation—structural diagnosis—key node identification—key connection and propagation path identification—hierarchical control output. It can also be completed collaboratively by multiple functional modules that communicate with each other.

[0154] The multi-source data fusion-based autonomous ship seaworthiness risk propagation and control method proposed in this embodiment has the following beneficial effects: (1) This embodiment abandons the isolated approach of processing inspection data, regulatory knowledge or expert opinions separately in the prior art, and innovatively constructs a three-layer directed weighted network model of “C-G-B”. The model defines specific defects, faults or abnormal states as C-type source risk nodes, system, equipment, function and organizational factors as G-type intermediate risk nodes, and the seaworthiness capability that the ship needs to achieve or maintain as B-type seaworthiness target nodes. The model fully depicts the entire process of risk from “entry-propagation-convergence” through three types of directed edges of C→G, G→G and G→B.

[0155] In the model construction process, this embodiment integrates multi-source heterogeneous information such as traditional port state control inspection data, regulations and technical documents, expert scores, and configurations based on ship type and seaworthiness stage. A classification model is used to determine the existence of candidate edges, a regression model is used to estimate edge weights, and the propagation direction is verified by combining network hierarchy constraints, engineering action sequence, regulatory requirements, and expert opinions. This mechanism ensures that the model completion results do not directly replace engineering judgments, but rather form verifiable potential propagation relationships while preserving the data sources and judgment criteria. Therefore, even with a severe shortage of actual inspection and long-term operational data for remotely controlled and fully autonomous vessels, this embodiment can still construct an airworthiness risk propagation network with clear hierarchy, propagation direction, and relative intensity. This allows risk information from different sources to be correlated, calculated, and traced within a unified structure. It can not only identify specific defects or abnormal states but also further explain through which systems, functions, or organizational factors the risk is transmitted to the airworthiness target. This effectively avoids the defects of directly applying traditional vessel relationships, such as omission of risk factors, distortion of relationship direction, and insufficient model applicability. It solves the problem of incomplete and uninterpretable risk modeling of autonomous vessels under conditions of data scarcity in existing technologies, and significantly improves the completeness, interpretability, and engineering credibility of the airworthiness risk propagation network.

[0156] (2) In this embodiment, random node removal and target node removal operations are used to determine whether the network relies heavily on a few structural units, thereby determining whether a refined critical structure analysis is needed. Subsequently, a weighted local gravity model is used to initially screen candidate critical nodes and narrow down the analysis scope. Based on this, candidate nodes are removed one by one, and multiple network performance indicators such as the maximum weakly connected component, network-wide reachability, C-to-B layer reachability, and the strongest path efficiency from C-to-B layer are recalculated. The comprehensive impact score of each node is formed based on the decrease in each indicator, thereby accurately identifying critical risk nodes.

[0157] For candidate critical links G→G and G→B, this embodiment performs SIR dynamic propagation simulation under the same propagation parameters and initial perturbation conditions. By removing or proportionally weakening individual links, it compares the changes in total propagation scale, activation peak, final impact range, and large-scale propagation probability. Simultaneously, it records node activation time, upstream source, and path hit count, thereby identifying critical links and fully reconstructing the propagation path C→G→…→B. Compared to existing methods that rely solely on static indicators such as degree and betweenness, this embodiment directly compares the network connectivity, inter-layer reachability, propagation scale, and the degree of impact on the target layer before and after node or link intervention. This effectively distinguishes between objects with "a large number of connections" and those with "significant actual propagation impact," avoiding misjudgment of critical objects. It provides a quantitative basis for prioritizing the allocation of limited inspection, monitoring, and maintenance resources to high-impact locations, solving the problem of inaccurate identification of critical structures caused by static ranking using a single centrality indicator in existing technologies. This significantly improves the accuracy and consistency of identifying critical risk nodes, critical propagation interfaces, and complete propagation chains.

[0158] (3) This embodiment abandons the traditional approach of fixed networks, uniform thresholds, or one-time risk ranking. Instead, it sets ship type, whether there are people on board, functional configuration, and seaworthiness stage as configurable conditions. Applicable objects are selected from a unified node and edge database, and corresponding target risk propagation networks are generated as needed. Based on this, according to the specific position of key objects in the propagation chain, C-class nodes and path starting points are mapped as source prevention objects, G-class nodes and G→G edges are mapped as propagation blocking objects, and B-class nodes and G→B edges are mapped as consequence mitigation objects.

[0159] Furthermore, this embodiment combines quantitative indicators such as node comprehensive impact score, edge intervention impact, path hit rate, applicable ship type, and seaworthiness stage to output control priority levels, recommended measures, and data basis. Thus, this embodiment can determine the priority control objects of risk entry points, intermediate propagation locations, and seaworthiness targets, providing clear results on "what to control first, at which propagation level to control, and applicable ships and stages." When ship functional configuration, operating conditions, or seaworthiness stage changes, the system can regenerate the network and control ranking, enabling risk control results to be dynamically adjusted according to target objects and assessment conditions. This effectively reduces the problem of analysis conclusions remaining at the level of risk lists or general rankings and being difficult to implement, providing direct and operable technical support for practical engineering applications such as seaworthiness inspections, status monitoring, maintenance arrangements, redundancy verification, functional isolation, degraded operation, and consequence protection. The control priority generation method proposed in this embodiment, which combines on-demand target network generation with a hierarchical mapping of "source prevention—propagation blocking—consequence mitigation," solves the problems of difficulty in implementing risk analysis conclusions and lack of specificity in control measures in existing technologies, significantly improving the pertinence, enforceability, and traceability of autonomous ship safety management measures.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hierarchical control method for the propagation of seaworthiness risks of autonomous ships that integrates multi-source data, characterized in that, Includes the following steps: S1: Acquire multi-source airworthiness risk data and preprocess it to establish a three-layer node database of defects-intermediate risk factors-airworthiness targets, including defect-type nodes, intermediate risk factor-type nodes, and airworthiness target-type nodes. S2: Construct a directed weighted risk propagation network for traditional vessels based on port state control and inspection data; S3: Obtain nodes based on autonomous vessels, and based on expert scores and a three-layer node database, obtain newly added directed candidate node pairs to obtain feature vectors of the newly added directed candidate node pairs. S4: Based on the feature vectors of the newly added directed candidate node pairs, a classification model is used to determine whether the newly added directed candidate node pairs have candidate edges; when the newly added directed candidate node pairs have candidate edges, a regression model is used to obtain the weights of the candidate edges. To construct a target network based on target vessel type and seaworthiness stage; S5: Based on the target network, construct a random node removal order and a target node removal order for performing node removal; And obtain the order of removal of random nodes and the order of removal of target nodes respectively. The overall resilience index after node removal in the round; when the nodes are removed in the random order in the round... The overall resilience index after removing the target nodes is compared to the index after removing the target nodes in the order of removal. If the rate of change of the overall resilience index after wheel node removal exceeds a set threshold, execute S6; where, Remove the round index from the node; S6: Obtain the weighted local gravity score of the nodes in the target network to determine candidate nodes; remove each candidate node from the target network to obtain the comprehensive influence score of the candidate nodes, and then determine the key risk nodes; S7: Using the SIR propagation simulation model, candidate edges and propagation paths in the target network are determined; based on the comprehensive influence score of candidate edges, key edges are determined; then, based on the propagation path hit rate, the average time for defective nodes to propagate to seaworthiness target nodes, and whether the propagation path contains key edges, key propagation paths are determined; and then, based on key risk nodes, key edges, and key propagation paths, the propagation of seaworthiness risks of autonomous vessels is controlled.

2. The method for hierarchical control of seaworthiness risk propagation in autonomous vessels by integrating multi-source data as described in claim 1, characterized in that, The formula used to obtain the comprehensive resilience index is as follows: In the formula: As a comprehensive resilience index; R WCC This represents the percentage of the current largest weakly connected component that has been preserved relative to the state before the attack. R E This represents the percentage of the current global efficiency of the directed network that has been preserved relative to its state before the attack. The weight of the current largest weakly connected component relative to the retention rate before the attack; in, In the formula: For the first The largest weakly connected component of the network after round node removal; The largest weakly connected component of the original network when no nodes were removed; For the first Global efficiency after removing a round node; For the global efficiency of the original network; The number of elements in the set; Remove the round index from the node.

3. The method for hierarchical control of seaworthiness risk propagation in autonomous vessels by integrating multi-source data as described in claim 1, characterized in that, The formula used to obtain the weighted local gravity score of a node is as follows: In the formula: The weighted local gravity score for node i; Let be the weighted connection strength of node i; Let be the weighted connection strength of node j; The radius of local action; This represents the shortest path distance between the two nodes. in, In the formula: For nodes Pointing to node Edge weights; All of them are nodes in the network.

4. The method for hierarchical control of seaworthiness risk propagation in autonomous vessels by integrating multi-source data as described in claim 1, characterized in that, The formula used to obtain the overall influence score of candidate nodes is as follows: In the formula: The overall impact score of the candidate nodes; For the index of evaluation indicators; This represents the total number of evaluation indicators used for comprehensive evaluation. For the first The weight of each evaluation indicator in the comprehensive calculation; This represents the relative decrease in the evaluation indicator; To remove a node The later The decrease in each evaluation indicator relative to the original network; In the formula: M 0 represents the original network's metric value. M -i This represents the metric value after removing node i.

5. The method for hierarchical control of seaworthiness risk propagation in autonomous vessels by integrating multi-source data as described in claim 1, characterized in that, Evaluation metrics for candidate nodes include: maximum weakly connected component, network-wide reachability, path reachability from defective nodes to airworthiness target nodes, and strongest path efficiency.

6. The method for hierarchical control of seaworthiness risk propagation in autonomous vessels by integrating multi-source data as described in claim 1, characterized in that, The comprehensive influence score of the candidate edges is obtained as follows: In the formula: The overall impact score of the candidate edges; ΔA e , ΔP e , ΔF e and ΔO e These represent the decrease in total propagation scale, activation peak, final impact range, and probability of large-scale propagation after removing candidate edge e, respectively. These are the influence weights for total propagation scale, activation peak, final impact range, and probability of widespread propagation, respectively.

7. The method for hierarchical control of seaworthiness risk propagation in autonomous vessels by integrating multi-source data as described in claim 1, characterized in that, The multi-source airworthiness risk data includes port state control inspection data for traditional vessels, regulations and technical data based on autonomous vessels, expert scoring data, and the target vessel type, functional configuration, and airworthiness phase configuration.

Images