Effective maritime surveillance hinges on the reliable detection of abnormal ship behavior (DASB). However, existing data-driven models often fail in complex environmental conditions like windy condition, struggling to differentiate normal physical responses from intentional anomalous maneuvers. This inability to disentangle environmental effects substantially limits detection reliability. This paper proposes a novel explainable physics-informed DASB framework by integrating gradient-enhanced physics-informed neural networks (gPINNs) and a scene-aware adaptive threshold (SAAT). First, we introduce gPINNs incorporating a Physical Parameterization Network for implicit hydrodynamic coefficient modeling and utilize gradient-enhanced physical residuals to capture both state consistency and trend plausibility. Additionally, we develop a 4-dimensional explainable anomaly scoring mechanism and the SAAT strategy for robust decision-making in dynamic scenes. Validation on real-world datasets confirms the proposed gPINNs-SAAT method’s superior performance in both behavior prediction and anomaly detection accuracy against benchmark models. The findings provide a robust and explainable decision support framework for intelligent maritime surveillance, effectively disentangling environmental effects from true behavioral anomalies.

Autonomous collision avoidance is critical for ensuring the safety and efficiency of maritime navigation. However, existing approaches often struggle to achieve realistic manoeuvrability, robust generalisation, and compliance with the Convention on the International Regulations for Preventing Collisions at Sea (COLREGs). To address these challenges, this study proposes a Reinforcement Learning (RL)-based collision avoidance framework, integrating three key innovations. Firstly, a discrete action space is designed to accurately capture the rudder control characteristics commonly used in real maritime operations. This is integrated with a Manoeuvring Modelling Group (MMG) model, ensuring that the generated trajectories are dynamically feasible and operationally realistic. Secondly, a multi-dimensional reward function is developed, incorporating collision risk, distance to target, navigational efficiency, operational comfort, and compliance with COLREGs. This is further supported by a line-of-sight (LOS) tracking mechanism, which stabilises heading corrections based on dynamic path requirements, significantly improving the agent’s course-keeping ability. Finally, the framework includes a robust generalisation strategy, using polygonal obstacle modelling to represent complex, irregular hazards more accurately. This is combined with real-world bathymetric data and multi-ship encounters for rigorous validation, ensuring the system can operate effectively in uncertain, multi-agent, and non-cooperative environments. The proposed model is trained using the Phasic Policy Gradient (PPG) algorithm within an Actor-Critic (AC) architecture, enabling robust policy learning under uncertainty. Simulation results demonstrate that the framework effectively reduces collision risk, maintains stable trajectories, and adheres to COLREGs, making it a practical and scalable solution for next-generation autonomous ship navigation.

Reliable visual perception is fundamental to autonomous navigation and efficient maritime management. However, in complex marine environments, optical sensors face significant challenges. Environmental factors such as fog and low-light conditions inevitably lead to significant signal attenuation and noise. This degradation adversely impacts downstream perception algorithms, creating a critical bottleneck that ultimately jeopardizes navigation safety. To mitigate the detrimental effects of image degradation, this study aims to capture the shared characteristics of multi-type degradations, enabling a unified parameter set to enhance images across diverse adverse conditions. Specifically, we decompose the objective of multi-scenario enhancement into two synergistic sub-tasks: texture reconstruction and color restoration, implemented via a dual-branch encoder–decoder architecture (TCE). Within this framework, an edge-information enhancement module — leveraging Large Kernel Differential Convolutional (LKD) — and a Color-Domain Reconstruction Branch (CRB) based on image decomposition are integrated to drive the restoration process. Furthermore, edge and reflectance images are incorporated as supervisory signals during training to assist the network in suppressing content-irrelevant interference. Experimental results on maritime datasets demonstrate that our approach effectively recovers fine-grained structural details concealed by fog or low-light, achieving superior visual quality and computational efficiency. Notably, the model maintains robust performance across different degradation types using a single set of weights. Object detection benchmarks further validate the practical utility of our method in bolstering downstream visual perception tasks under foggy and low-light maritime environments. The source code is available at https://github.com/MAPGroup/TCE.

To investigate the fire spread patterns of multiple vehicles within PCTC (Pure Car and Truck Carrier) under the influence of jet fire. This study proposed an equivalent material and constructed a homogenized BEV fire model that can change in real time according to the heating conditions, avoiding the fire from developing at a fixed heat release rate. At the same time, a battery pack jet fire model suitable for large-scale simulation was constructed. Subsequently, these two models were combined to explore the fire spread patterns and hazard areas under different jet fire directions. The results show that the ceiling jet is the main cause of ignition of surrounding vehicles, and thermal radiation is the main way of vehicle ignition. Reducing the vehicle spacing and the outbreak of jet fire significantly accelerates the development of the fire, with the maximum increase in the fire heat release rate reaching 54.6%. Among them, the jet fire to the right is the most dangerous, which can advance the ignition time of the adjacent vehicle by 70.2 s. Based on the critical heat flux criterion for area classification, the expansion of the hazardous region within 0.2 m is relatively slow. At a distance of 1.8 m, the left and front sides of the burning vehicle remain in a low-risk state for 274.2 s and 377.4 s, respectively. This research can provide model support and theoretical basis for large-scale vehicle fire simulation and PCTC fire prevention.

Port approach channels concentrate vessel conflicts, waiting, and speed adjustments that can increase fuel use and CO2 emissions, yet real-time wind is rarely represented as both an emission driver and a scheduling constraint. This study develops a wind-constrained vessel traffic organization framework and a Pareto-based multi-objective proximal policy optimization algorithm (Pareto MO-PPO). A wind-aware propulsion model links vessel speed, transit time, relative wind, and CO2 emissions, while wind-dependent engine-load limits restrict the feasible speed range. A preference-conditioned actor-critic network learns policies across the efficiency and emission trade-off, and an external archive retains non-dominated policies. The framework was evaluated in a Caofeidian Port scenario. Relative to first-come, first-served (FCFS) scheduling, the balanced policy reduced emissions by 11.41% while increasing system time by 7.49%. Detailed results show how wind-aware Pareto policy learning can provide port operators with explicit operating choices rather than a single fixed weight solution.

Reliable autonomous traffic organization is critical for smart ports operating under dynamic ocean conditions. This study proposes a comprehensive framework for sea port scheduling based on Multi-Source Information Fusion (MSIF). To address the challenge of fusing heterogeneous data streams-including vessel trajectories, resource availability, and environmental constraints-a Mixed-Integer Programming (MIP) formulation is first established. Subsequently, a high-fidelity digital twin integrating Cellular Automata (CA) and Multi-Agent Systems (MAS) is developed to simulate heterogeneous cooperative behaviours and stochastic uncertainties. Crucially, a novel Simulation-based Multi-Objective Genetic Algorithm (SMOGA) serves as the autonomous decision-making engine, fusing simulation feedback with evolutionary search to optimize vessel sequencing and resource allocation dynamically. Validated with real-world data from Tianjin Port, the framework outperforms both traditional rules and modern swarm intelligence algorithms, increasing throughput by 16.5% and reducing turnaround time by 11.7%. This research demonstrates how fusing multi-scale information into an intelligent evolutionary framework enables robust, autonomous decision-making in congested maritime environments.

This study employs ERA5 and NCEP-FNL wind fields to drive the WAVEWATCH-III model for simulating wind wave fields during the rapid intensification of Typhoon Hato and their impact on vessel drift after anchor dragging. Results indicate that each wind field demonstrates different advantages in reproducing actual sea states, with the ERA5-WW3 scheme showing the closest agreement between computed drift velocities and observed values. The high resolution wave model effectively captures wave evolution characteristics during typhoon rapid intensification, providing reliable data for ship anchor dragging analysis that meets practical navigation accuracy requirements.

To systematically investigate the complex causal mechanisms of maritime accidents, this study proposes an automated analytical framework that integrates Natural Language Processing (NLP) with complex network theory. The framework is designed to transform unstructured accident investigation reports into a quantifiable causal network that reflects systemic risk. Drawing on 564 official reports, this study constructs a standardised dataset of causal factors through a two-stage process combining automated preprocessing and manual coding. NLP techniques are then employed to extract causal relationships from the texts, enabling the construction of a weighted, directed complex network from discrete factors. To ensure the reliability of the framework, the extracted causal logic is verified by a domain expert panel, and the identified risk propagation patterns are validated against representative empirical cases. Topological analysis reveals that the causal network exhibits the “small-world” and “scale-free” properties characteristic of complex systems, indicating a high potential for efficient risk propagation mediated by a few key hubs. A multi-dimensional centrality assessment identifies static risk sources of high influence, including “Inadequate Supervision”, “Vessel Stability/Stowage Issues”, and “Adverse Weather/Sea State”. Furthermore, a risk pathway identification algorithm is applied to extract five typical risk propagation patterns. These pathways dynamically illustrate the systemic process by which risk evolves from latent managerial failures, through technical vulnerabilities and the actions of front-line personnel, to a major accident when triggered by specific environmental conditions. This work provides a dynamic, systematic network perspective for accident causation analysis, and its findings offer more precise intervention targets and process-based preventive strategies for maritime safety management.

Accurate vessel traffic flow prediction is essential for maritime navigation safety, port operations, and sustainable waterway management, as it supports proactive risk mitigation and efficient resource allocation. This study proposes SATM-ConvLSTM, a hybrid model that integrates Convolutional Long Short-Term Memory (ConvLSTM) networks with a Spatial Attention Mechanism (SAM) embedded in the temporal modeling process. By jointly capturing sequential dynamics and spatial dependencies across bridges, the proposed model effectively characterizes upstream–downstream interactions and localized flow variations. SATM-ConvLSTM is validated using real-world AIS data collected from the Yangtze River Estuary and its upstream bridge system, representing complex riverine traffic scenarios. Experimental results show that SATM-ConvLSTM consistently outperforms benchmark models, including LSTM, BiLSTM, ConvGRU, ConvLSTM, and Transformer, across four evaluation metrics, namely MAE, MAPE, RMSE, and R2. In particular, the proposed model shows superior capability in capturing peak flows and directional changes, where baseline methods often underperform. By incorporating spatial correlations and temporal patterns, SATM-ConvLSTM improves prediction accuracy and robustness under both single- and multi-input settings. These findings demonstrate its practical value for vessel scheduling, bridge operation planning, and maritime navigation safety, while also providing useful insights for coastal traffic management, risk assessment, and sustainable waterway governance in complex inland and estuarine environments.

Efficient and sustainable transportation systems are crucial. This paper introduces an innovative approach to optimizing ship scheduling in waterborne public transportation by leveraging deep learning-driven demand prediction. Employing the Temporal Fusion Transformer (TFT) model for superior forecasting performance, the study integrates dynamic and static data inputs to predict transportation demand accurately. A two-stage methodology comprising a TFT-based demand prediction model and a scheduling decision module is proposed, utilizing a mathematical model aimed at cost minimization. The model’s effectiveness is demonstrated through real-world data, showcasing optimized capacity adjustments that align closely with actual demand, thus reducing scheduling discrepancies and enhancing service quality. This work signifies a pivotal step toward utilizing deep learning for maritime transportation management, emphasizing the potential economic and societal benefits of AI applications in this field.

The surge in electric vehicles (EVs) is causing a structural disruption to high-density, short-sea Roll-on/Roll-off (RoRo) transportation, driven by stricter safety regulations and unique transport protocols. Operators like those in China's Qiongzhou Strait have implemented an 'EV Dedicated Service' (EVDS) strategy. This strategy involves a complex coordination problem between dedicated EV-carrying vessels and separate vessels for transporting drivers. However, this emerging scheduling paradigm has been insufficiently studied. This paper proposes a multi-objective mixed integer programming model for the RoRo fleet scheduling with a novel methodological approach to formulate EVDS mechanism. Additionally, we develop an Adaptive Large Neighborhood Search - based heuristic algorithm, featuring novel problem-specific neighborhood structures. Realistic instances validated the algorithm's performance against benchmark methods. The results also revealed the balance between economic efficiency and service levels across three different demand scenarios (Low-Season, Normal-Day, and Peak-Season). Furthermore, the analysis reveals the strategic value of flexible deployment for EV-certified vessels. We also introduce a method to quantify operational resilience by analyzing the impact of elastic capacity planning on alleviating port congestion. The findings provide a robust decision-support framework for RoRo operators and policymakers navigating the surge in EV transport demand.

Ensuring the navigational safety of LNG carriers is crucial for maintaining orderly traffic management in port waters and for reducing the environmental risks associated with maritime accidents. In recent years, numerous scholars have conducted in-depth studies on the impact of LNG vessel navigation, and corresponding traffic control strategies within port areas have been continuously optimized. This paper, based on the current LNG traffic control measures implemented in Chinese ports, analyzes the allocation of channel resources during the inbound and outbound voyages of LNG carriers. It proposes a set of quantitative models grounded in the theory of spatiotemporal consumption, including a port channel capacity calculation model, a delay time model for other vessels, and a port throughput estimation model. These models are used to quantify the extent to which LNG carrier navigation affects the efficiency of other vessels. The proposed models are applied to a case study of the Dagang Port Area in Tianjin. Results show that due to traffic control restrictions, the channel capacity during LNG carrier navigation significantly decreases compared to normal operating conditions. Based on the forecasted LNG berth throughput, the study further estimates the vessel delays caused by varying numbers of LNG carriers, and identifies the key factors contributing to reduced channel capacity under special traffic control measures. These findings provide a theoretical basis for the planning, design, and expansion of port channels.

With the global energy system undergoing a transition toward green and low-carbon systems, the scale of liquefied natural gas (LNG) maritime transportation has expanded rapidly. Influenced by a combination of factors including the global economy, geopolitics, energy policies, and environmental conditions, the Liquefied Natural Gas Maritime Transportation Network (LMTN) exhibits a high degree of structural complexity and has gradually emerged as a prominent research focus in the field. This study provides a comprehensive review of the current research progress on LMTN. First, the concept of LMTN is introduced and the major stages of its research development are highlighted. Second, LMTN construction methods are systematically summarized with data sources, theoretical foundations, and application scenarios, thereby establishing a technical framework for global LMTN research. Subsequently, bibliometric analysis is also employed to extract representative publications and reveal the knowledge structure, historical evolution, and emerging research frontiers of the field. Finally, from three technical perspectives—methodology, data, and computational power—this study discusses existing limitations and challenges, and identifies future development trends of LMTN research driven by big data and artificial intelligence. Overall, this study aims to provide scientific guidance for future LMTN research and theoretical support for enhancing the security and resilience of global energy transportation systems.

Island-land roll-on/roll-off (ro-ro) transportation holds a critical position in maritime traffic. However, the volatility and uncertainty of multi-type transportation demands - encompassing trucks, cars, and passengers - present significant challenges to vessel scheduling. This paper addresses the island-land ro-ro transportation problem driven by multi-type demands and proposes a vessel scheduling optimization model aimed at maximizing cabin utilization. Leveraging actual operational data from the Qiongzhou Strait, a high-density liner model was developed, and an adaptive large neighborhood search algorithm was designed to solve this nonlinear mixed-integer programming problem, incorporating capacity resource constraints. Case study results reveal that the optimized schedule reduced the number of voyages by four while improving cabin utilization by 18.57 percentage points -from 72.66% to 91.23% - while fully meeting transportation demands. These findings offer theoretical support and practical guidance for enhancing resource efficiency in island-land ro-ro transportation.

Accurate prediction of multiple transportation demands - encompassing cars, passengers, and trucks - is crucial for enhancing the efficiency of intelligent transportation systems at port terminals and for optimizing roll-on rolloff (Ro-Ro) passenger ship scheduling. These demands series exhibit temporal variability, with fluctuations during holidays and seasonal events, and display inherent heterogeneity and interdependency across different types. Existing predictive research based on deep learning often treats historical data uniformly and relies solely on known date attributes, neglecting latent temporal features within time series data. To address these limitations, this study proposes the Dates Fine-Grained Temporal Fusion Transformer (DFTFT), a novel model enhancing prediction accuracy through innovative clustering-based date attribute encoding. This approach introduces two key innovations: firstly, the Date-Enhanced Time Series Clustering method, which employs Autoencoders and K-means clustering to uncover latent date-specific patterns from complex temporal dynamics. Secondly, the DFTFT architecture uniquely incorporates the novel attributes via an enhanced Variable Selection Network, and leverages a Temporal Fusion Decoder to integrate information across diverse time horizons. The architecture is further complemented by a Gated Residual Network for modeling nonlinear relationships and a Quantile Output Layer for robust uncertainty estimates. Experiments on real-world Ro-Ro data demonstrated DFTFT's high accuracy, robustness, and interpretability. Analysis indicated the superior importance of clustering-based date attributes over traditional ones and that demands from the preceding three days significantly influence future predictions. Consequently, the DFTFT model enables port terminals to more effectively anticipate transportation demands and optimize resource allocation.

The rapid development of artificial intelligence and big data has elevated autonomous ships to a prominent position in maritime research, introducing significant challenges in automatic path planning. This study presents a systematic framework for global ship path planning by integrating heuristic search algorithms with maritime route extraction methods. The framework consists of three key components: trajectory compression, maritime traffic route extraction, and global path planning. An adaptive threshold model combined with a sliding window algorithm processes ship trajectories to capture motion state changes accurately. The subsequent turn-point identification algorithm maps turning points precisely, while waypoints are optimised through Cluster-based Kernel Density Estimation to enhance route topology. A path matching search algorithm incorporating Dynamic Time Warping, Traffic Separation Schemes, and historical ship data generates practical and smooth navigation paths. Experimental validation using Automatic Identification System (AIS) data from the Yangtze River Estuary and Ningbo-Zhoushan Port areas demonstrates the superior performance of the proposed framework. The generated paths show improved compliance with Traffic Separation Schemes requirements and enhanced smoothness compared to traditional path planning algorithms.

Ice-Strengthened vessel structures are essential for ensuring safe navigation in ice-covered waters. While the China Classification Society currently mandates B-class ice strengthening for vessels operating in domestic ice-covered waters, recent years have witnessed ice conditions in northern coastal waters approaching B3-class severity. This study proposes a finite element simulation-based methodology for structural strength verification to assess the navigational safety of B-class strengthened vessels under B3 ice conditions, using a 100,000-ton oil tanker as a case study.The investigation reveals that although stress concentration leads to maximum stress values exceeding allowable limits, they remain below the yield stress of high-strength steel. The research demonstrates that reinforcing transitional structures, particularly bracket connections, can effectively mitigate stress concentration. B3-class ice loads primarily affect the midship section, with relatively minor impact on the overall hull structure and contributing marginally to the total stress in cargo holds. Through enhanced bracket structures, the analysis confirms an 8.97% reduction in maximum equivalent strain and a 9.23% decrease in maximum equivalent stress, thereby ensuring safe navigation under B3 ice conditions for B-class strengthened vessels.

Island-land roll-on/roll-off (Ro-Ro) transportation is critical component of regional logistics. However, volatile multi-type demands for trucks, cars, and passengers poses scheduling challenges, causing capacity mismatches and resource redundancy. To address these inefficiencies, this study proposes a demand-driven scheduling optimization framework. A nonlinear mixed-integer programming model is developed to maximize vehicle capacity utilization, incorporating a penalty method to pragmatically manage demand fulfillment. A customized adaptive large neighborhood search algorithm is designed to solve the large-scale problem. The framework's effectiveness is validated using real-world operational data from the Qiongzhou Strait, a high-density maritime corridor. The proposed ALNS-based approach achieves a 91.95 % capacity utilization rate, significantly outperforming benchmark heuristics while maintaining high demand fulfillment. Furthermore, analysis of an embedded elastic capacity coefficient reveals its strategic function in enabling trade-offs between operational efficiency and service robustness across different planning periods. The proposed framework resolves critical scheduling imbalances and provides port authorities with a quantitative tool for enhancing operational efficiency and sustainability in Ro-Ro networks.

Sensor fusion will undoubtedly remain a cornerstone of enhanced perception and decision-making in numerous intelligent systems. As we all know, each type of sensor has advantages and disadvantages. The data obtained from a single sensor is frequently incomplete and unreliable. The complementary information from multi-sensor improves the reliability and robustness of the system, particularly while working in challenging conditions. To improve the situation awareness ability for intelligent surface vehicles (ISVs) under complex navigational conditions, many efforts have been devoted to developing advanced multi-sensor data fusion methods. However, these methods often suffer from the high computational cost and high latency on compute-constrained platforms. It is thus necessary to develop intelligent edge computing frameworks to accelerate the data fusion methods, making real-time environmental perception, behavior decision, and navigation control for ISVs. In this work, we will introduce the various sensors used in ISVs and discuss their advantages and disadvantages. We will also present the key steps and typical methods for multi-sensor heterogeneous data fusion suitable for ISVs. Furthermore, we present a review and recent advances for model lightweight. These methods will help reduce model parameters and accelerate model computation, thereby facilitating more efficient model deployment in ISVs.