Advancing Maritime Safety: A Literature Review on Machine Learning and Multi-Criteria Analysis in PSC Inspections

Abstract

This literature review provides a structured quantitative analysis of existing research on the application of machine learning models (MLMs) and multi-criteria decision-making methods (MCDM) in the context of port state control (PSC). The aim of the study is to capture current research trends, identify thematic priorities, and demonstrate how these analytical tools have been used to support decision-making and risk assessment in the maritime domain. Rather than evaluating the effectiveness of individual models, the study focuses on the distribution and frequency of their use and provides insights into the development of methodological approaches in this area. Although several studies suggest that the integration of MLMs and MCDM techniques can improve the objectivity and efficiency of PSC inspections, this report does not provide a comparative assessment of their performance. Instead, it lays the groundwork for future qualitative studies that will assess the practical benefits and challenges of such integration. The findings suggest a fragmented but growing research interest in data-driven approaches to PSC and highlight the potential of advanced analytics to support maritime safety and regulatory compliance.

1. Introduction

This paper analyzes the integration of machine learning (ML) models and multi-criteria analysis (MCA) in assessing vessel safety risks within the context of port state control (PSC) inspections, applying a systematic literature review (SLR).

The focus of this study is on a systematic literature review on the application of advanced ML and MCA methods to improve evaluation, forecasting, and planning processes in the context of PSC. The approach is primarily quantitative, with the intention of collecting and processing relevant literature documenting trends, scope, and specific applications of these methods in the maritime sector. The aim is not to critically evaluate each individual study or to assess the merits and limitations of particular techniques, but rather to systematically collect, categorize, and analyze existing work in order to assess the scope, dynamics, and focus areas of current research.

This provides an insight into the extent to which ML and MCA techniques have already been integrated into inspection risk assessment, resource planning in shipping companies, and the development of strategic regulatory approaches within international frameworks such as Memoranda of Understanding (MoU). While the practical relevance of these methods for improving maritime safety is obvious, the main contribution of this study is to take stock of existing knowledge and identify potential research gaps.

The study is part of a broader research project (PhD), with the next phase focusing on collecting newer and more comprehensive data sources to support further investigations while strengthening the application of advanced MLMs and MCA techniques in the context of PSC.

Maritime safety is a key component of the global transport system, and PSC [1,2] is an integral part of the international strategy aimed at reducing the risk of marine accidents. This system operates within the framework of the International Maritime Organization (IMO). It relies on regional PSC agreements, such as the Paris MoU and the Tokyo MoU, which facilitate the exchange of information and best practices among signatory states. PSC inspections play a crucial role in international maritime oversight, aiming to ensure safety and the protection of life at sea, as well as the preservation of the environment from potential maritime accidents [3,4,5,6].

The main objective of PSC inspections is to identify and eliminate safety risks and potential violations that could jeopardize the safety of the vessel, the crew, and the environment. Based on binding international conventions such as SOLAS (Safety of Life at Sea), MARPOL (Marine Pollution) and STCW (Standards of Training, Certification and Watchkeeping), PSC inspectors are authorized to inspect vessels in port, regardless of the flag under which the vessel sails, to check compliance with international standards. During these inspections, which may encompass various aspects of the vessel’s operation, inspectors assess the vessel’s condition, including its structural components, safety equipment, environmental protection systems, and the qualifications of its crew. In the event of non-compliance or irregularities, inspectors may decide to detain the vessel in port, which can have serious consequences for both the vessel owners and its operations.

Modern risk assessment methods for PSC vessel inspections extend far beyond traditional approaches, which primarily rely on statistical analyses and expert judgment. These are often limited by the subjective judgements of the inspectors and, from the shipowner’s perspective, reduce the ability to assess potential violations or the likelihood of the vessel being detained. The use of MLMs [7,8] enables the automated processing of large data sets [9], pattern recognition, and a more accurate determination of safety risks. At the same time, MCA provides a systematic framework for decision-making by considering multiple relevant factors simultaneously, which improves the interpretation of the results generated by MLMs.

The application of these methods enables a more precise prediction of inspection results. It is therefore extremely useful for both inspectors, who need to formulate requirements by legal standards, and shipowners, who want to minimize costs and optimize the time vessels spend in port. By utilizing advanced technologies, including artificial intelligence (AI), machine learning (ML), and algorithms, it is possible to optimize procedures for identifying potential risks based on historical data on violations and deficiencies on vessels. This approach not only improves the efficiency of inspections but also helps reduce the number of accidents caused by technical defects [10,11,12,13].

In this context, current research [14,15,16,17,18,19] focuses on the development of sophisticated MLMs that enable automatic pattern recognition in data from previous inspections and a more reliable prediction of future violations and safety risks. The main challenges in the field of PSC inspections and vessel safety are to increase the accuracy of predictions and improve the robustness of these models to ensure their applicability in real operational conditions.

The effectiveness of different ML and multi-criteria decision-making (MCDM) methods in PSC depends on the specific requirements of the inspection process. Methods such as decision trees are highly interpretable and easy for inspectors to understand, but often have lower accuracy in more complex scenarios. In contrast, methods such as Random Forests (RFs) and Support Vector Machines (SVMs) offer higher predictive performance and accuracy, but their complexity reduces transparency and makes the results more difficult to interpret. Bayesian network (BN) algorithms are fast and simple and are therefore suitable for initial analyses, but are limited by the assumption of feature independence. On the other hand, MCDM methods such as the Analytic Hierarchy Process (AHP) and the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) allow a structured ranking of vessels based on risk and other factors, which support informed decision-making, although they rely on subjective judgments that may influence the final outcome. In practice, an optimal approach often requires the combination of interpretable and powerful methods to find a balance between the reliability of the model and the confidence of the inspectors.

The combination of MLMs and MCA opens up the possibility of improving the efficiency of PSC inspections and increasing maritime safety [20,21,22,23].

The aim of this paper is therefore to conduct a SLR for the period of the last ten years (from 2015 to 2025) to analyze existing approaches to the integration of MLMs and MCA in the assessment of vessel safety risks in the context of PSC reviews, to identify essential methods and research gaps, and to define possible directions for future research.

The paper is organized in such a way that the introductory section presents the basic concepts of the research, focusing on the importance of applying MLMs and MCA in assessing safety risks within the context of PSC inspections. The Materials and Methods chapter outlines the framework of the research methodology, including the search and analysis strategy for relevant literature, the criteria for selecting sources, and the data used for analysis. The Results and Discussion chapters present the analysis results, along with a discussion of the key findings and challenges encountered in implementing MLMs and MCA methods within the context of PSC inspections. Finally, the Conclusions chapter summarizes the main findings of the study. It formulates recommendations for future research and improvement of the methodological approach to safety risk assessment through the application of modern technological solutions and their implementation in the field of PSC inspections.

1.1. Maritime Safety and PSC Inspection

In the section “Maritime Safety and PSC Inspection”, the studies analyzed highlight the most important aspects of maritime safety and the role of PSC inspections in maintaining this safety. The main findings of these studies relate to the importance of PSC inspections, the impact of the COVID-19 pandemic in this area, the most commonly identified deficiencies in PSC inspections, human factors as key elements of maritime safety, and the researchers’ emphasis on the need for technological improvements and better regulations.

The data collected during PSC inspections form the basis for improving safety in the maritime industry and enable the development of strategies to improve inspection processes and predict potential risks on vessels [24,25].

Considering that PSC inspections are one of the most important mechanisms for ensuring maritime safety and reducing operational risks, their role in identifying violations on vessels is crucial, particularly in contributing to the reduction of marine casualties and environmental incidents.

Studies [26,27,28,29] confirm that PSC inspections play a key role in the implementation of international regulations such as SOLAS and MARPOL and ensure a high level of safety worldwide. Their effectiveness in reducing the risk of accidents through preventive measures and penalizing vessels that do not meet safety standards is particularly emphasized.

The COVID-19 pandemic has had a significant impact on PSC inspections, making them more challenging to conduct and reducing the scope of inspection oversight. Studies [30,31,32,33,34,35] show that the restrictions introduced during the pandemic have led to fewer inspections, difficulties in physically monitoring vessels, and adjustments to regulatory procedures to ensure uninterrupted shipping traffic. Due to the limited access of inspectors to vessels, the number of safety violations increased, particularly in areas such as technical maintenance and crew training. These problems also highlighted the need to digitize inspection procedures to ensure their continuous application even in exceptional circumstances.

The data analysis of PSC inspections reveals that the most common deficiencies and violations on vessels are related to technical aspects, safety procedures, and human factors. During Paris MOU inspections, problems related to safety equipment, navigation systems, and vessel maintenance are most frequently identified. Additionally, inadequate crew training and non-compliance with safety procedures often contribute to serious incidents at sea.

Research indicates that vessels with a history of repeated violations are inspected more frequently, underscoring the importance of consistently applying inspection protocols. The human factor is a crucial aspect of maritime safety, as most accidents at sea are attributed to human error. Inadequate training, crew fatigue, and poor management of safety procedures can significantly affect the safety of shipping. A particular problem is the lack of awareness of the importance of compliance with safety regulations, which emphasizes the need for continuous training and education of seafarers.

Current research [36,37,38,39] points to the need for technological improvements and better regulation in the area of PSC controls. The automation of data analysis, the use of digital technologies in vessel monitoring, and the improvement of international co-operation between port states can significantly increase the efficiency of inspection procedures. The introduction of digitalized inspection systems would certainly enable faster and more accurate identification of safety deficiencies. At the same time, improvements in regulations would contribute to more consistent enforcement of international safety standards.

The results of the previously cited studies and research confirm that PSC inspections play a crucial role in improving maritime safety, with particular emphasis on the importance of the human factor, the impact of global crises, and the need to modernize and optimize inspection methods to increase their efficiency and ensure a higher level of safety in the shipping industry [40].

In the context of PSC, it is important to understand that although the core process often involves physical vessel inspections, digitization can significantly enhance and support various aspects of the process. Physical inspection of vessels is still important, but digital tools and technologies can improve the efficiency and accuracy of the inspection process.

For example, digitization can enable faster access to data through electronic records of previous inspections, vessel tracking, and compliance data, and streamline data collection. In addition, advanced technologies such as remote monitoring, sensors, and AI-based predictive analytics can help inspectors identify potential issues before they become critical, increasing safety and reducing the risk of accidents.

While digitalization cannot completely replace physical inspections, it can improve the process by allowing inspectors to better analyze data, make informed decisions faster, and gain easier access to important information. This approach enables a more efficient and safer inspection process, contributing to the overall improvement of the system.

1.2. Application of MLMs in PSC

PSC inspections are an essential mechanism for monitoring maritime traffic, allowing ports to ensure that vessels comply with international safety, environmental, and labor standards. Traditional inspection methods rely on manual inspections, the inspector’s experience, and predefined lists of criteria. However, the growing amount of data generated from previous inspections, digital registers, and sensor systems opens up the possibility of applying MLMs aimed at optimizing inspection procedures, reducing operational risks, and improving predictive analysis in the identification of high-risk vessels [41,42,43,44,45,46,47,48,49].

The application of MLMs in the context of PSC helps recognize patterns in large datasets, reduces the need for manual intervention, and increases the efficiency of control mechanisms.

MLMs [50,51,52] can be divided into two main categories:

• Supervised learning methods (SLMs) use labeled data to train models that enable the prediction or classification of new data based on previous experience.
• Unsupervised learning methods (ULMs) do not require labeled data and are applied to detect hidden patterns and structures in the data, which helps discover new relationships or clusters among the data.

The most commonly used SLMs are as follows:

Linear regression is a statistical model used to predict continuous values based on the linear relationship between variables. In the context of PSC, linear regression is used to analyze quantitative variables that influence the frequency of incidents or inspection outcomes, such as vessel age, vessel size, and previous safety violations. Although this approach can be helpful in analyzing trends, its use is limited to situations where there is a linear relationship between the predictors and the target variable [19,53,54,55,56].

Logistic regression is a binary classification method that predicts the probability of an event occurring, i.e., it provides a yes/no answer (e.g., whether a vessel needs to be inspected or not). In PSC reviews, logistic regression helps to assess vessels with a higher probability of a safety risk based on historical accident data and non-compliance or other risk factors [16].

Decision trees are classification and regression techniques based on hierarchical data decomposition. They work by successively branching nodes based on selection criteria, with decisions being binary or multi-class depending on the algorithm used and the nature of the problem. In the context of PSC inspections, these methods enable a systematic analysis of relevant variables, such as vessel type and age, frequency and condition of previous inspections, among others, allowing for an objective assessment of risk and determination of the need for detailed inspections. Their interpretability and transparency make them an effective tool for making informed regulatory decisions [57,58].

SVMs belong to the group of ML and are based on models that are used for classification or regression by defining an optimal hyperplane that separates the data classes. There are several variants and extensions of SVMs, such as the Linear SVM, which is used when the data are linearly separable; non-linear SVM, which uses kernel functions to work with non-linearly separable data; Support Vector Regression (SVR), a version of SVM adapted for regression problems; and one-class SVM, which is used for anomaly and outlier detection. In the context of PSC inspections, an SVM enables the accurate classification of vessels based on their likelihood of meeting safety standards and identifies those that require priority inspection oversight. A key advantage of this method is its ability to efficiently analyze high-dimensional and non-linearly separable data, which is crucial for processing complex inspection criteria [49,59].

RF [53,60] is an ensemble method based on a set of decision trees that enables collective decision making to improve prediction accuracy and reduce the risk of overfitting. In the context of PSC inspections, RF is used to predict irregularities and risks on vessels, relying on a wide range of variables, including vessel type, inspection history, and other relevant factors. Its main advantage lies in its robustness and its ability to analyze complex and heterogeneous data. AdaBoost [61] is also an ensemble MLM that combines several weak classifiers (e.g., decision trees) into a more potent model. This technique uses labeled data to train the model and iteratively improves its performance by assigning a higher weight to samples that are difficult to classify. Applying these methods enhances predictions by sequentially learning and weighting the samples, resulting in improved precision in data analysis.

The most commonly used ULMs in the context of PSC inspections are as follows:

The K-Means algorithm is a clustering method that groups data into a predefined number of clusters based on their similarity to one another. In the context of PSC inspections, K-Means is used to classify vessels based on common characteristics, such as vessel type, age, or inspection history. This enables the more efficient allocation of inspection resources and helps identify hidden patterns among vessels that might otherwise go unnoticed using traditional analytical approaches [62].

Cluster analysis represents a group of MLMs and statistical analyses that allow segmentation of data into homogeneous groups based on their similarities. These methods are crucial for uncovering hidden patterns in large datasets, enabling informed decisions in various areas, including maritime safety and inspection processes. In the context of PSC inspections, cluster analysis is used to identify unsafe or high-risk groups of vessels based on their specific characteristics, such as the frequency of incidents, technical deficiencies, compliance with international standards, and frequency of environmental violations. By applying cluster analysis methods, it is not only possible to recognize potentially risky segments of the fleet but also to identify patterns that may indicate systemic failures in the maintenance and monitoring of vessels. Additionally, cluster analysis enables the identification of long-term trends in the maritime industry, which can significantly enhance inspection planning strategies. Based on the analysis results, it is possible to optimize the distribution of inspection resources and focus on the groups of vessels that pose the most significant risk to safety and environmental protection. By integrating cluster analysis with other machine learning methods, such as RF and AdaBoost, the accuracy of predictions can be further improved, enabling a proactive approach to maritime risk management [63].

Artificial neural networks (ANNs) [64,65,66,67] and deep learning [68,69] are highly developed models that imitate the human brain. ANNs can recognize patterns in complex and unstructured data. In the context of PSC inspections, ANNs can be utilized to identify complex relationships between various variables that impact the safety of vessels, including weather conditions, vessel type, previous violations or incidents, and others. The advantage of ANNs lies in their ability to learn and adapt to changes in the data, making them very effective in predicting risks and accidents. ANNs and their advanced variants, such as Convolutional Neural Networks (CNNs) [70,71] and Recurrent Neural Networks (RNNs) [72], are increasingly being used to analyze large data sets generated during PSC inspections.

BNs and Bayesian models (BMs) can be applied to both supervised and unsupervised learning, depending on their usage and the presence of labels in the data. When BNs are used for classifications or regressions where the data are already labeled (with output variables), they are categorized as supervised methods. For example, when they are used to predict the probability of a particular outcome based on known characteristics, such as predicting the risk of a shipping accident based on historical data. The application of BNs falls under unsupervised learning when they are used to uncover hidden structures in data without prior knowledge of the initial variables. In addition, BNs are often used in semi-supervised learning (SSL) when there is a combination of labeled and unlabeled data.

The application of BNs and BMs to PSC inspections [38,73,74,75,76,77,78,79,80,81,82] represents one of the most advanced solutions for probabilistic modeling and decision-making under uncertainty. These methods enable the formalization of complex dependencies between variables and provide a robust framework for analyzing dynamic systems with incomplete and uncertain data. Their application to PSC inspections contributes to a more accurate risk assessment, improvement of inspection strategies, and optimization of resource allocation, thus increasing the efficiency and effectiveness of marine surveillance. The cited theoretical and empirical studies confirm that BNs facilitate a more in-depth analysis of the factors that influence the outcomes of inspections, including the likelihood of detention, identification of high-risk vessels, and assessment of compliance. By using probabilistic models, researchers are developing sophisticated methods for predicting security threats and optimizing inspection procedures, reducing uncertainty in the decision-making process. Some papers focus on improving vessel liability prediction algorithms, while others investigate the integration of BNs with MCA to achieve more accurate modeling of risk scenarios and increase regulatory efficiency. The common goal of these studies is to improve maritime safety through the use of advanced MLMs. BNs are utilized as key tools for building intelligent risk assessment systems and facilitating the development of sophisticated, data-driven inspection strategies. Their application not only increases the efficiency of PSC procedures but also allows for the more precise identification of potential threats and the making of optimal regulatory decisions. The synergy of these studies reflects the goal of developing predictive models that improve preventive measures and enhance global vessel safety.

The application of MLMs to PSC inspections represents a significant advance in improving the efficiency, accuracy, and predictive analysis of vessel safety and regulatory compliance. Monitored methods enable accurate vessel classification based on historical inspection data, optimizing the identification of high-risk vessels and increasing the efficiency of inspection processes. Non-monitored techniques, on the other hand, uncover hidden patterns in the data, enable proactive detection of potential irregularities, and optimize the allocation of inspection resources. Integrating these methods into PSC inspections not only improves the detection of violations but also significantly contributes to increasing vessel safety, reducing operational risks, and minimizing the likelihood of accidents at sea through more accurate and informed regulatory decisions.

1.3. Application of MCDMs in PSC Inspections

MCA and MCDM methods [83,84] serve as basic approaches for optimizing inspection procedures in the context of PSC inspections. Their application facilitates a structured and quantitatively sound evaluation of alternative solutions, considering multiple, often conflicting criteria, and provides a methodologically rigorous framework for decision-making in complex regulatory and operational environments. Unlike MLMs, which rely on algorithmically generated patterns derived from data, MCA is based on explicit decision modeling that enables transparent analysis of the determinants influencing inspection outcomes.

In the context of PSC inspections, the implementation of MCA involves defining relevant evaluation criteria, selecting and analyzing alternatives based on predefined parameters, determining the relative weighting of the criteria through expert judgment or empirical data, and integrating the results using advanced aggregation methods. This analytical approach enables the formal quantification of safety risks, the objectification of inspection decisions, and the optimization of resource allocation, which ultimately increases the efficiency of regulatory oversight and reduces uncertainty in assessing safety risks [85].

In the risk assessment of vessels, MCA enables the integration of various factors, including technical specifications, operational parameters, inspection history, flag state, detention frequency, and compliance with international conventions.

The best-known methods used in this context include the AHP, the TOPSIS, PROMETHEE (Preference Ranking Organization Method for Enrichment Evaluations), and ELECTRE (Elimination and Choice Expressing Reality). Each of these methods offers different advantages in the multi-criteria optimization process.

The AHP method [86] facilitates the hierarchical decomposition of decision problems and uses pairwise comparison analysis to quantify the relative importance of criteria.

The TOPSIS method [76,87] ranks alternatives based on their proximity to ideal and anti-ideal solutions, ensuring the selection of an option that optimally balances competing criteria.

The PROMETHEE [88] method incorporates preference functions that enable a more adaptive and dynamic evaluation of alternatives, aligning with the decision-makers’ subjective preferences.

The ELECTRE [89] method employs a process of eliminating dominant alternatives, supporting decision-making in highly uncertain environments with competing criteria. This is particularly beneficial for inspection procedures that require a comprehensive approach to safety risk assessment.

Additionally, the synergistic integration of MCDM with probabilistic models, such as Bayesian networks (BNs), facilitates the formal consideration of uncertainties and dependencies between key inspection parameters. This approach enhances the accuracy of risk prediction, optimizes inspection strategies, and contributes to the overall improvement of maritime safety [8,26,33,40,76].

The application of MCA methods in the context of PSC inspections is a key element of the modern approach to safety risk assessment and regulatory oversight in the maritime sector. Its application enables the systematic analysis and quantification of risk factors, refines inspection methodologies, and improves the efficient allocation of inspection resources, ultimately promoting a strategically oriented and science-based decision-making model for maritime safety.

1.4. Application of Combined ML and MCA Methods in the Context of PSC Inspections

In today’s PSC inspection practice, there is a growing need for more efficient and accurate methods to analyze safety risks on vessels. Traditional approaches that rely on human judgment and simple algorithmic techniques are being gradually replaced by advanced, integrated, and novel methods known as “hybrid” or “combined” methods. These methods enable the efficient processing of large amounts of data and decision-making based on complex analysis. In this context, the combination of MLMs and MCA represents a significant step towards improving the accuracy and efficiency of safety assessments for vessels subject to PSC review.

The combination of MLMs and MCA in PSC inspections improves the assessment of vessel safety risks. DL methods, such as stacked autoencoders [90], have proven to be highly effective in extracting relevant features from high-dimensional inspection data. Furthermore, algorithms such as XGBoost, which are based on gradient boosting methods, are characterized by iterative model refinement in risk classification and prediction. In addition to these methods, the application of reinforcement learning (RL) techniques enables adaptive decision making in inspections by iteratively learning optimal strategies from previous results [58,91,92].

The application of the Gale–Shapley algorithm [46] plays a crucial role in optimizing the allocation of inspection resources. This algorithm facilitates the allocation of inspection capacity based on a multi-criteria evaluation of vessels, ensuring a balance between inspection efficiency and accuracy. Consequently, it optimizes the allocation of inspectors to the vessels that pose the highest safety risk. In addition, principal component analysis (PCA) [58] helps to reduce the dimensionality of the data and identify the most critical risk factors, allowing further study and a more accurate assessment of safety parameters.

The integration of MLMs with MCA techniques such as AHP and TOPSIS enables decision-making based on the simultaneous assessment of multiple aspects of vessel safety. These methods enable the incorporation of expert judgment and quantitative data, ensuring objectivity in the evaluation process [76,87].

The combined application of MLMs and MCA significantly improves the ability to identify high-risk vessels, optimizes inspection processes, and contributes to a more efficient allocation of inspector capacity. This approach enables more accurate risk assessment and better resource allocation, thereby increasing the overall efficiency of inspections and maritime safety.

In summary, the integration of MLMs and MCA, along with other advanced statistical and mathematical methods in PSC inspections, not only enables a more accurate risk assessment but also optimizes the allocation of inspection resources [93]. This ensures that capacity is targeted to those vessels that pose the most significant risk to maritime safety. This approach enhances the accuracy and speed of decision-making while also providing greater transparency and accountability in the inspection process, ultimately contributing to improved global maritime safety.

Table 1. Distribution of references according to thematic focus in connection with PSC inspections.

Subject	References
PSC inspection in general	[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23]
Maritime safety and PSC inspection	[24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40]
Application of MLMs in PSC	[16,19,38,41,42,43,44,45,46,47,48,49,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82]
Application of MCDMs in PSC Inspections	[8,26,33,40,76,83,84,85,86,87,88,89]
Application of combined ML and MCA methods in the context of PSC Inspections	[46,58,76,90,91,92,93]
ML in general	[50,51,68]

provides a brief quantitative summary of the academic studies, organized by their thematic focus on PSC and associated methods. The largest group of studies (40) addresses the use of MLMs in PSC, reflecting a strong research focus on improving inspection processes through advanced technologies, data analytics, and algorithm-based decision-making.

The second most represented category includes papers addressing the general aspects of PSC inspections (23 sources), highlighting the continued academic focus on understanding the regulatory framework, operational structure, and core functions of the PSC system.

Maritime safety in the context of PSC inspections is represented by 17 sources, confirming the crucial role of PSCs in ensuring the safety of life, property, and the marine environment.

The use of MCDM methods in the context of PSCs is documented in 10 sources, while combined approaches involving both ML and MCDM techniques appear in six references. This reflects a growing, albeit still ongoing, trend towards integrated analytical models in this area.

The least represented group is “ML in general” with only three sources, suggesting that general ML concepts in this paper are explored through their specific application to PSC rather than as a standalone research area.

Overall, Table 1 illustrates a clear shift towards digitalization and data-driven methods in PSC, with a focus on automating decision making, improving safety procedures, and reducing subjectivity in vessel condition assessment. This thematic division shows an emerging research trend towards integrated, data-driven approaches for the evaluation and optimization of PSC performance. The focus is on analytical decision support systems that aim to improve the objectivity and efficiency of inspection processes across the maritime sector.

Given the complexity and scope of the topic, this work represents the first stage of a broader research initiative. The present work focuses on mapping the research landscape and identifying key trends and patterns based on frequency, distribution, and thematic grouping. A subsequent qualitative analysis is already planned, which will examine the specific methods, data sets, and results of the identified studies in detail.

However, some initial qualitative reflections on the performance and characteristics of the methods used have already been made—particularly in the papers cited as references [58,67], which provide methodological insights into selected approaches. We believe that this multi-layered methodology allows for a more structured and meaningful synthesis of the literature, starting with a macro-level overview and moving to a more fine-grained analysis in the next phase.

2. Materials and Methods

Data processing in this study was performed using R-4.5.0 packages for Windows [94], which provide a robust framework for efficient data analysis and manipulation. R, a versatile tool for statistical analysis and data visualization, offers a comprehensive selection of packages that support various analysis techniques, from basic statistical tests to advanced modeling approaches. Key packages, such as dplyr (1.1.4), ggplot2 (3.5.1), and tidyr (1.3.1), have been utilised to process, clean, and visualise the data, enabling insightful interpretation and comprehensive analysis of the results.

2.1. Bibliography Data Description

In this study, the input dataset comprises 93 sources stored in an Excel spreadsheet, which contains the collected and relevant literature on PSC inspections, MLMs, and MCA. Only relevant sources were selected, organized, and systematized. The table is sorted alphabetically by the last name of the first author and contains the following key information: Column 1—Source ID; Column 2—Full source citation (Author 1, A.B.; Author 2, C.D. Title of the article. Abbreviated name of the journal, Year, volume, pages); Column 3—List of authors; Column 4—Title of the article; Column 5—Keywords; Column 6—Year of publication; Column 7—Abbreviated name of the journal/conference; Column 8—Full name of the journal/conference; Columns 9–12—Indexing data in relevant databases (Web of Science, Scopus, Google Scholar, ScienceDirect), where 1 means that the source is indexed and 0 otherwise; Column 13—Total number of authors of the publication. This structured breakdown enables efficient filtering and categorization of the sources according to the criteria of the PICOC method, ensures a comprehensive analysis, and serves as a basis for further discussion and interpretation of the results.

Keywords were generated to identify relevant terms and phrases related to the research topic. Terms were selected that most accurately reflect the content of the sources, thereby improving findability in search engines and databases. The selected keywords were strategically integrated into the title, summary, and paragraphs of the text to maximize the article’s visibility and accessibility to the target audience. Abbreviations were introduced for the keywords in the table, while all synonyms were standardized to ensure that each term appeared only once. For example, vessel detention is abbreviated as vd, while artificial neural network (ANN) is shortened to “nn”, which also applies to other terms. This approach enhances searchability, ensures consistency, and facilitates the efficient retrieval of relevant terms, thereby making the study more accessible to researchers and experts in the field.

Table 2. Frequency of different source types.

Journal or Source Type	Frequency
Journal	86
Proceedings	2
Book or chapter in book	2
PhD Dissertation	1
Master of Science thesis	1
Web link	1

shows the frequency distribution of the different types of sources in the collected literature. It categorizes the various types of sources (e.g., journals, conference papers, books, dissertations) and indicates how often each type appears in the dataset.

2.2. Literature Review Methodology

There are various methods for conducting literature reviews. The most commonly used methods include the systematic literature review (SLR), the narrative literature review (NLR), meta-analysis, scoping review (SR), critical review (CR), and the population, intervention, comparison, outcome, and context (PICOC) method. Each of these methods has specific applications and advantages. However, given the aim and purpose of this study, the SLR method with the application of PICOC is considered the most appropriate. The SLR method enables a rigorous, systematic, and objective analysis of existing research, which is crucial for a topic that integrates MLMs and MCA in the context of PSC security risk analysis. This approach enables the selection of relevant sources based on predefined criteria, reducing subjectivity and ensuring the accuracy of the analysis [95,96].

The combination of the SLR method with PICOC enables a precise formulation of parameters for literature selection, thus increasing the precision and relevance of the analysis. The PICOC method is essential for formulating the research question and filtering out relevant literature sources through five key components:

• (P)opulation refers to all studies, papers, dissertations, and articles related to PSC inspections.
• (I)ntervention focuses on the application of methods such as MLMs and MCA in analyzing PSC inspections.
• (C)omparison analyzes different approaches to the application of MLMs and MCA in the field of PSC inspections and compares them with other methods of safety risk analysis.
• (O)utcome identifies and evaluates safety risks on vessels and the improvement of existing methods and models in the context of PSC inspections.
• (C)ontext refers to the application of these methods in the field of maritime inspections, with a particular focus on vessel safety research and the implementation of the relevant technologies.

By employing the SLR method in conjunction with PICOC, this paper offers a comprehensive analysis of methodological approaches, identifies research gaps, and explores the connections between studies in this area. This approach enables the systematic integration of MLMs and MCA methods within the context of PSC inspections, which is crucial for improving existing models and developing new strategies in this area.

Table 3. Step activity description.

1. DEFINITION OF THE RESEARCH QUESTION	Formulate a clear and concise research question using the PICOC method. Focus on the safety risks associated with PSC controls, including MLMs and MCA.
2. FORMULATION OF CRITERIA FOR STUDY SELECTION	Use the PICOC framework to define specific criteria for selecting relevant studies. Define the inclusion and exclusion criteria for appropriate studies, based on factors such as research area, methodology, and publication year.
3. LITERATURE SEARCH	Systematic search in relevant databases, including Web of Science, Scopus, Google Scholar, and ScienceDirect. Use of specific keywords and phrases related to MLMs, MCA, and PSC inspections.
4. STUDY SELECTION	Application of the previously defined inclusion and exclusion criteria to select only relevant studies. This step involves reviewing abstracts, keywords, and conclusions.
5. ASSESSMENT OF THE QUALITY OF THE STUDIES	Assessment of the quality of the selected studies based on methodological criteria.
6. DATA EXTRACTION	Systematic extraction of key data from the selected studies, including methodology, results, variables analyzed, and applications of MLMs and MCA in relation to PSC inspections.
7. DATA ANALYSIS AND SYNTHESIS	Analyzing and synthesizing the data from all studies to identify common themes, methodological approaches, and research gaps. Integrate findings into a unified framework that links MLMs, MCA, and PSC inspections.
8. WRITING THE REPORT	Produce a report summarizing the key findings of the literature review. This report will include a critical review of methodological approaches, identification of research gaps, and suggestions for future areas of research.
9. DISCUSSION AND CONCLUSIONS	Discussion on the importance of integrating MLMs and MCA in analyzing PSC inspections and the benefits and challenges of using these methods. Suggestions for further development and improvement of research in this area.

presents a comprehensive breakdown of the steps involved in the systematic research process. The initial phase focuses on formulating a well-defined research question using the PICOC methodology, with an emphasis on PSC inspections and the application of MLMs and MCA. This is followed by the establishment of rigorous selection criteria, which clearly define inclusion and exclusion parameters based on the research scope, methodology, and publication years.

The next stage involves conducting a systematic search across key academic databases, including Web of Science, Scopus, Google Scholar, and ScienceDirect, utilizing carefully selected keywords and phrases related to MLM, MCA, and PSC inspections. Once the studies are selected, their quality is critically evaluated against predefined methodological criteria. Subsequently, essential data are extracted from the chosen studies, including research methodology, key findings, analyzed variables, and the implementation of MLMs and MCA within the context of PSC inspections.

The extracted dataset is then systematically analyzed and synthesized to identify prevailing thematic patterns, methodological trends, and existing research gaps. Finally, a structured report is compiled, offering a synthesized overview of key insights from the systematic literature review, a critical evaluation of methodological approaches, identification of unresolved research gaps, and recommendations for future investigations in this domain.

The methodological steps (Table 3) provide a structured and comprehensive application of the SLR methodology, using PICOC for the precise formulation of the research question and criteria. This approach ensures objectivity, rigor, and thoroughness in the literature review, which is essential for integrating different approaches in assessing safety risks during PSC inspections.

3. Results

Table 4. Distribution of authors per paper in the dataset.

Number_of_Authors	Frequency
3	24
4	24
2	17
1	11
5	10
6	3
7	2
9	2
Total	93

illustrates the distribution of the number of authors per paper in the collected sources. Papers with three and four authors are the most common (24), suggesting that research in this area tends to be conducted in smaller teams, allowing for effective collaboration and coverage of different areas of expertise. This distribution suggests that research in this area often requires the collaboration of multiple researchers, but without the need for excessively large teams. Single-authored articles (11) are relatively rare, and articles with five or more authors are even rarer, suggesting that a larger number of authors is not always necessary to conduct high-quality research in this area.

The focus was on identifying sources dealing with PSC and the application of MCA and MLMs in the context of PSC risk assessment and decision-making. The combined search terms included terms such as PSC, PSC inspection, vessel detention, ML, effectiveness, risk, safety, BN, and the like, with the term “PSC” necessarily included. These terms can also be seen in Figure 1, which helps to explain the frequency of certain terms in the visualization.

The top 10 keywords (Figure 1) represent the most frequently used terms in the publications, listed in order of occurrence. The most frequently used keywords are PSC, with 83 occurrences, and VD (vessel detention), with 31. ML (machine learning) also occurs 13 times. Other keywords such as effectiveness, safety, and risk appear 13 times each, BN (Bayesian network) 12 times, prediction 11 times, MCA (multi-criteria analysis) 10 times, and NN (neural network) 9 times. This visualisation gives an overview of the predominant topics in the analyzed publications and reflects the most important concepts investigated in this area.

The distribution of publications by database (Figure 2) shows the number of publications in four well-known academic databases. Google Scholar contains the most significant number of publications, with 89, followed by Scopus with 88 and Web of Science with 87. The lowest number of publications is found in the ScienceDirect database, with 58. This diagram illustrates the distribution of publications across various databases, providing a better understanding of research coverage and availability within the academic community.

The frequency of authors with two or more publications (Figure 3) shows the number of publications by individuals who have at least two works in the dataset. The leading author is Wang, S. et al. with 12 publications, closely followed by Yan, R. et al. with 11. Authors Yang, Zh. and Yang, Za., with seven, and Wan, C. et al., Yin, J. et al., and Yu, Q. et al. have each published five papers.

The distribution of publications by year (Figure 4) and the distribution of publications by year and source (Figure 5) illustrate the trends in publication output over time. The highest number of publications was recorded in 2023 with a total of 19 publications, while the lowest number was observed in 2017 with only 1 publication. These visualizations provide a clear overview of the fluctuations in the number of publications over the years, allowing for the identification of periods with the highest and lowest publication activity.

Figure 5 further refines this analysis by plotting the years on the x-axis and the abbreviated titles of the different publication types (such as journals, dissertations, conference proceedings, and similar sources) on the y-axis. The size of the circles in this graph represents the number of publications: larger circles indicate years with a higher number of publications. In comparison, smaller circles highlight years with fewer publications. This visualization allows for a nuanced analysis of publication trends over time and shows the years with the most and least publications.

Figure 6 focuses on the distribution of publications across specific journals and proceedings. The x-axis shows the number of publications per year, while the y-axis lists the different journals and proceedings in which these papers were published. For example, Marine Policy contains 11 publications, the Journal of Marine Science and Engineering 10, Marine Policy and Management 8, and Reliability Engineering and Systems 6. This visualization provides insight into the authors’ preferences in choosing publication venues and highlights the most essential and relevant journals and proceedings for specific years.

The author collaboration network with cluster visualization (Figure 7) depicts the scientific collaboration structure among authors based on their shared publications. It comprises 41 distinct clusters, where nodes represent individual authors and 435 edges reflect their co-authorship relationships. The clusters are visually distinguished by different colors, facilitating the identification of research groups with stronger collaboration. This type of analysis offers valuable insights into the patterns of collaboration and highlights key contributors within the scientific community.

The author collaboration network with a focus on the surnames of first authors (Figure 8) provides a more granular view of scientific collaboration, concentrating on the primary researchers behind the publications. In this network, nodes represent the first authors, while the 426 connections illustrate their co-authorship links. Structural analysis of this network reveals the distribution of authorship and the contributions of individual researchers, thereby allowing the identification of dominant collaboration patterns within the dataset under examination.

4. Discussion

The data analysis in this study was performed using the R package, which enables efficient data analysis, manipulation, and visualization. R has proven its exceptional capabilities as a tool for statistical analysis, offering a wide range of available packages that support various methodological approaches, from basic statistical tests to advanced modeling techniques. This tool enabled the extraction of relevant results that were subsequently analyzed as part of the ongoing research.

The analysis of the data revealed several vital conclusions. Most of the papers in the collected sources involve three or four authors, suggesting that smaller research teams predominantly conduct research in this area. These teams tend to foster effective collaboration and involve a wide range of expertise. Single-authored papers are relatively rare, while publications with five or more authors are even rarer. This distribution suggests that a large number of authors is not necessarily a prerequisite for high-quality research results in this field.

Table 5. Comparative analysis of methods in the context of PSC inspections.

Method	Advantages	Limitations	Effectiveness in PSC Context
Decision Trees	High interpretability; easy to apply	Lower accuracy in complex scenarios	Suitable for training and quick decision-making, but limited in complex cases
RF	High accuracy; robust to noisy data	Low transparency; complex structure	Very effective for risk detection, but requires additional explanation of results
SVM	Excellent performance on complex datasets	Difficult to interpret for non-experts	Accurate, but less applicable without expert support in operational settings
BN	Fast and simple; suitable for small or text-based data	Assumes feature independence, which is often unrealistic	Useful for preliminary classification and filtering
AHP	Structured multi-criteria decision-making	Subjectivity in determining weighting factors	Effective when decision priorities are clearly defined
TOPSIS	Clear rankings; focuses on proximity to the ideal solution	Limited flexibility for unstructured or dynamic data	Useful for ranking vessels based on risk level

provides a structured comparative overview of selected ML and MCDM methods in terms of their applicability to PSC inspections. Each method is evaluated based on its main advantages, limitations, and overall effectiveness in supporting risk assessment, decision-making, and PSC compliance. The analysis aims to highlight the suitability of each approach for operational implementation in maritime safety inspections.

The thematic focus of the publications analyzed is made clear by the use of key terms. The most frequently occurring key terms, such as PSC (83 occurrences) and VD (31 occurrences), indicate important research directions, while terms such as ML and efficiency appear in fewer publications. This distribution of key terms likely reflects the current challenges and technological advancements in the maritime industry, as well as the increasing use of advanced methods to solve complex problems.

The distribution of publications across different databases reveals that most papers are indexed in the Google Scholar database (89 entries). At the same time, Scopus and Web of Science contain nearly identical numbers of publications (88 and 87, respectively). This indicates a wide availability of research results in various academic sources, which enhances the visibility and accessibility of relevant information within the scientific community.

The trend in the number of publications shows a significant increase in 2023, when the majority of papers were published. This increase could be attributed to increased research activities, the initiation of new projects, and a growing interest in the topic. An analysis of publication trends over time provides valuable insights into the periodic fluctuations in research intensity, which may indicate progress and developments within the field.

The authorship network, comprising 41 clusters and 435 links between authors, provides a detailed insight into the structure of scientific collaboration. This network exhibits a clear concentration of cooperation, a crucial factor for progress in the field. Additionally, identifying the dominant research teams and their interactions offers valuable insights into collaboration practices and key players within the scientific community. Such information is crucial for understanding the dynamics of research collaborations and is likely to play a significant role in shaping future research efforts.

In the coming years, significant progress is expected in the application of MLMs to the risk analysis of vessels, particularly in the context of PSC. A systematic review of the available literature reveals that the application of MLMs can provide significant benefits in enhancing the accuracy, efficiency, and speed of vessel security risk assessments. While these techniques promise to improve on existing methods, several critical challenges still need to be overcome to realize their potential fully.

One of the main problems is the availability and quality of data, as information on vessel accidents and inspection results is often not standardized or publicly available in sufficient quantity. This lack of high-quality data is an obstacle to the application of ML techniques, as these methods require large amounts of accurate and up-to-date information. Without adequate data, it is challenging to develop reliable models that accurately predict safety risks, which limits the application of machine learning (ML) in the context of PSC inspections.

Although MCA facilitates systematic decision-making in complex situations, its integration with MLMs has not been sufficiently explored, particularly in the context of PSC inspections. The synergy between these methodological approaches promises more accurate and efficient assessments, but requires further research to fully realize its benefits in the context of vessel safety analysis.

Another challenge that needs to be addressed concerns the interpretation of the results arising from the combination of MLMs and MCA methods. Many MLMs operate as “black boxes”, i.e., it is not always possible to clearly explain the reasons for categorizing a vessel as high-risk. This lack of transparency can lead to uncertainty among inspectors and weaken confidence in automated decisions. The literature recognizes the need to develop more transparent and interpretable models that enable inspectors to better understand the decisions made and facilitate the verification and validation of these decisions in practice.

To overcome these obstacles, further research is needed in the coming years in the areas of data collection, standardization, and development of new methodological approaches to integrate MLMs and MCA techniques.

In addition, the development of transparent and interpretable models that provide clear explanations for decisions is crucial to increase confidence in these technologies and their application in vessel surveys. The transparency of the models not only improves the understanding of the decision-making process but also helps to increase accountability and accuracy in the application of the review criteria.

Future research directions should therefore focus on further optimizing the methods of data collection and validation to increase the accuracy of bibliometric analyses. Particular attention should be paid to enhancing MLMs for assessing security risks, thereby increasing the efficiency and accuracy of testing procedures.

Although the results of this study provide valuable insights into the research trends and key players in this field, the challenges related to data quality and standardization remain a significant obstacle to the further development and implementation of advanced analytical approaches in vessel safety verification.

5. Conclusions

To improve the interpretability of machine learning models in the context of PSC, it is critical to apply techniques that improve model transparency and facilitate understanding by inspectors and regulators. Models with inherent interpretability, such as decision trees and rule-based systems, provide explicit and understandable decision rules that align well with operational requirements and allow inspectors to follow the decision process step-by-step. This transparency is critical to gaining the trust and acceptance of practitioners who rely on clear explanations to justify inspection results.

For more complex MLMs, which often work as a “black box”, post-hoc explanation methods are used to explain the behavior of the model without compromising the predictive performance. Techniques such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) generate interpretable explanations by quantifying the contribution of individual input features to individual predictions. These methods allow inspectors to understand which factors influenced the model’s decision, increasing transparency and facilitating compliance.

In addition, improving interpretability not only promotes inspector confidence but also helps to identify potential biases or errors in the model, contributing to continuous improvement and safer vessel operations. Therefore, incorporating interpretability techniques is critical to the effective use of machine learning tools in PSC inspections to balance predictive accuracy with the practical need for understandable and justifiable decisions.

This systematic literature review examined the integration of MLMs and MCA within the framework of PSC inspections. By applying the SLR and PICOC methodologies, the study identified and synthesized key contributions, methodological trends, and research gaps in the existing literature.

Findings indicate that MLMs, particularly in the form of predictive models, have strong potential to enhance risk identification and prioritization in PSC inspections. These models allow for more accurate forecasting of vessel detention probabilities and improved assessment of safety-related factors. When combined with MCA, they support structured and transparent decision-making by enabling the evaluation of multiple, often conflicting, risk indicators.

Despite these opportunities, the review also highlighted critical challenges. One of the main limitations is the scarcity of empirical studies that apply MLM and MCA to real-world PSC data. Moreover, methodological heterogeneity and lack of standardization in current approaches hinder broader applicability and reproducibility.

Future research should focus on developing advanced MLM techniques—such as DL and ANNs—and on improving the operational integration of MCA in PSC environments. Emphasis should also be placed on collaborative efforts with maritime authorities to validate these models in practice, ensuring they contribute meaningfully to the improvement of maritime safety and regulatory compliance.