A Fault Diagnosis Approach Utilizing Artificial Intelligence for Maritime Power Systems within an Integrated Digital Twin Framework

Abstract

This research focuses on enhancing the preventive maintenance strategies currently employed for induction motors within ship propulsion systems, advocating for a shift towards a predictive maintenance model. It introduces a real-time monitoring framework that continuously observes the induction motor, providing essential support to maintenance personnel. The motor operates under a range of environmental and operational conditions, including temperature fluctuations, rotational speeds, and mechanical loads. These variations can obscure the current time series data, potentially masking signs of actual damage and hindering effective damage detection. To tackle this issue, the proposed framework utilizes artificial intelligence (AI) technology, specifically the well-established autoencoder, in conjunction with the Mahalanobis statistical distance. This approach accounts for the diverse operating conditions during the training phase, allowing it to model complex, non-linear relationships and effectively differentiate between normal and anomalous states. The framework is integrated into a decision support platform designed for real-time operations in maritime settings, offering a sophisticated system architecture that aims to align advanced damage detection methodologies with the maritime industry’s need for real-time, user-friendly solutions.

1. Introduction

The shipping industry plays a crucial role in global trade, accounting for approximately 80% of the world’s freight transport and contributing 2.2% of the atmospheric air pollutants [1]. A report from the European Union in 2019 [2] indicates that, without significant intervention, the greenhouse gas (GHG) emissions from shipping could rise by 50% to 250% by the year 2050, relative to the levels recorded in 2008. To mitigate these emissions, the maritime sector has emphasized the adoption of data-driven strategies. The incorporation of digital twins (DTs) has significantly improved the operational efficiency and extended the operational lifespan of vessels [3]. However, most current methodologies focus on models that simulate the energy consumption during a ship’s operation. The main engine and auxiliary systems, including the thrusters and pumps, are powered by electric induction motors. Although these motors are generally dependable, they are susceptible to damage in components such as the shaft, bearings, and stator windings.

Preventive maintenance aims to reduce the likelihood of equipment failures; however, it can inadvertently lead to the early replacement of parts, which may shorten the lifespan of components and increase the overall costs. Furthermore, unexpected stresses can cause failures to occur prior to the planned maintenance schedule, potentially disrupting operations and creating safety hazards [4]. Research shows that around 32% of damage to induction motors is due to insulation failures resulting from short circuits, which produce high currents that threaten both motor integrity and operator safety [5]. This highlights the critical need for robust condition monitoring that focuses on vibrations, temperatures, and current levels to ensure the timely detection of faults [6].

Motor current signature analysis is promising for condition monitoring and fault diagnosis, but time-domain analysis often lacks effectiveness. Predictive maintenance increasingly relies on artificial intelligence (AI), particularly deep learning (DL), to extract insights from sensor data. DL models can preprocess data, identify features, and diagnose potential failures in critical vessel systems. They can also estimate the remaining useful life (RUL) for maintenance scheduling. Recent studies have advanced techniques like artificial neural networks (ANNs), fuzzy logic, and wavelet transformations [7]. For instance, Verma et al. [8] improved ANN classification for induction motor currents, while Duan et al. [9] developed a DL model for the diagnosis of transformer faults. Hsueh et al. [10] created a diagnostic system for three-phase induction motors using CNNs to classify fault types from the current analysis. Wang et al. [11] found that CNNs trained on scalograms from a continuous wavelet transform outperformed those using manual features. Agrawal et al. [12] compared ANNs and support vector machines (SVM) for the diagnosis of bearing faults, with SVMs showing greater efficiency.

Despite the great effort, comprehensive lifecycle solutions for ship power systems remain unrealized due to the significant technological challenges in implementing digital twins (DTs). Key issues include the absence of a collaborative framework for subsystem models, non-modular DT architectures that hinder technology integration, and difficulties in managing diverse real-time data. Additionally, deep learning models require extensive training datasets, which are difficult to compile, especially for rare anomalies [13].

Previous studies have explored the use of autoencoders in various applications, including an LSTM autoencoder for real-time pipeline leak detection using accelerometer signals and a framework for the identification of corrosion in hydropower facilities, achieving over 80% true positive detection [14,15]. Recent studies have shown that autoencoders effectively identify damage in electric motors. A comparative analysis found that MLP autoencoders outperformed MLP, CNN, and LSTM architectures in detecting damage from vibration signals, achieving an AUC of 99.11% [16]. Additionally, research developed a hybrid health indicator using current and vibration signals to monitor bearings and stator windings in motors, demonstrating excellent performance across various load conditions [17]. CNN autoencoders were also used for the rapid identification of electrical damage in induction motors, achieving over 95% accuracy. In our previous works, this unsupervised methodology proved to be especially advantageous in sectors such as maritime, where instances of damage data are frequently limited [18]. To this end, PRISMOID aims to develop a modular and scalable digital twin platform capable of collecting large amounts of heterogenous data from different sensors and systems onboard and provide the software with the means to reliably use the data onboard and onshore for early warnings and decision support reasoning. The present research focuses on the anomaly detection framework developed during the project, which utilizes autoencoders in conjunction with the Mahalanobis distance to distinguish between healthy and compromised states by analyzing time series current data, with a specific focus on damage occurring in stator windings [19].

The proposed framework is validated by simulation data from an induction motor under various conditions. Its integration into aging vessels enhances management and maintenance, offering easy damage detection that optimizes energy losses, reduces downtime, and improves the CAPEX and environmental sustainability.

This paper is organized as follows. Section 2 outlines the overall architecture of the digital twin platform, which includes the AI-driven framework. Section 3 details the simulation model for the motor, while Section 4 describes the proposed framework and Section 5 presents the validation results of the model. In Section 6, a discussion of the primary advantages of this approach and its applicability in business scenarios is offered, while Section 7 concludes with the key findings of the research.

2. Digital Twin Architecture

2.1. General Concept for Integrated Digital Twin Platform (IDTP)

An IDTP strives to enhance adaptability and resilience in an ever-changing operational setting, while promoting sustainable vessel operation to adopt a predictive approach. This maintenance approach heavily relies on Internet of Things (IoT) sensor measurements and integrated digital twins based on AI to estimate the condition of a vessel’s critical systems, such as its main engine and others, based on data-driven methods that perform timely prediction. A comprehensive illustration of the extended IDTP concept is provided in Figure 1.

The platform efficiently collects essential data from the ship sensors and operational history, analyzing them with advanced algorithms and multimodal fusion techniques. It offers two versions, onboard and onshore, both continuously assessing risks and supporting decision-making to reduce carbon emissions in the maritime industry. The onboard version focuses on environmental monitoring, energy efficiency, predictive maintenance, and performance enhancement, while the onshore version manages the vessel lifecycle, including design, production, and retrofitting evaluations. An IDTP is a vital resource for safety officers, authorities, crew, first responders, and maritime enterprises, providing comprehensive threat monitoring and real-time decision-making capabilities. Its customizable dashboards enable ongoing monitoring and mapping, making it a leader in integrated digital twin platforms in the shipping industry.

2.2. Modular and Scalable AI Agent Approach

There are two approaches to achieving reliable predictions using advanced algorithms and deep learning models. The first involves creating a deep learning platform that integrates big data technologies, collects data from onboard sensors, and uses large-scale processing for offline decision-making in shipping. This platform relies on effective data management and advanced deep learning models, as detailed in [20]. Figure 2 shows the architecture, centered around a distributed computing infrastructure that includes data preprocessing, training, and evaluation, allowing users to select various configurable parameters for data preprocessing and model training.

The design of the decision support system incorporates the concept of an application programming interface (API) gateway, which serves as a network component within a larger information system. This gateway facilitates the connection between system users’ applications and the system’s services, as well as enabling communication among the various applications within the system. A key benefit of this approach is its ability to facilitate interaction and organization among different applications according to their specific requirements. Furthermore, the system employs two technologies for data storage and management: a database and a data lake. The database is responsible for housing structured metadata, while the data lake is used to store preprocessed datasets and trained models.

The second focus area is a real-time operational module, as depicted in Figure 3. This module functions continuously and without interruption, ensuring timely assistance in decision-making processes. The trained autoencoders from the decision support platform aim to forecast the conditions of the vessel’s subsystems and identify any anomalies. The objective is to create a system capable of managing extensive data through distributed processing, providing real-time predictive analytics to various users of the platform. Oversight of the system is maintained by an administrator, who is responsible for the data lake and performs administrative tasks such as managing, monitoring, and configuring the deployed DL models. The database is employed to gather real-time data from the vessel, which subsequently enters the distributed computing infrastructure. Within this framework, two main services operate: the preprocess service, which prepares the data and converts them into an appropriate format for analysis, and the autoencoder prediction service, which receives the formatted data for real-time anomaly detection. The analysis results are accessible to end users through the cloud.

2.3. Data Acquisition System

The initial phase of data-driven damage diagnostics involves collecting and safeguarding essential information, categorized into event data and condition data. Event data include historical records of faults, failures, repairs, and their causes, while condition data offer insights into asset health through measurements like vibration, oil quality, temperature, pressure, and humidity. This research used the PrismaSense™ data acquisition system [21], featuring a hybrid data fusion architecture for distributed processing and storage (Figure 4).

A comprehensive wired network was implemented at the power plant to facilitate wireless data transmission to the gateway. This setup incorporated signal collectors and power analyzers, while the wireless communication was based on the IEEE 802.15.4 Zigbee protocol, specifically designed for maritime applications. The gateways interfaced with the Prisma server through Ethernet and raw TCP, employing WebSocket protocols to ensure data security before transmitting information to the cloud for processing and visualization. The power analyzer was capable of monitoring a three-phase, four-conductor system with precision of 10 mA at a sampling frequency of 20 kHz, calculating effective values and other metrics every five minutes, with adjustable rates for users. Although Prisma Sense was utilized on an older, mostly intact vessel for two years, the validation of the algorithms was based on data obtained from a system simulator linked to the smart collectors via TCP/IP.

3. Simulation Test Case

The framework proposed in the previous section is evaluated regarding whether it can detect stator winding damage in induction motors. In practice, the experimental analysis of electromechanical systems under normal and damaged operation conditions is not always possible due to the need to damage the equipment. To address the above-mentioned problem, a 3-phase low-voltage squirrel-cage motor is simulated using MATLAB Simulink.

The simulation model utilized a 4 kW, 3-phase squirrel-cage induction motor (IM) and a fault block to replicate faulty operations. Figure 5 displays the finalized model, while

Table 1. Motor simulation details and specs.

Block Name	Parameter	Value
3-phase squirrel-cage IM (4 kW, 400 V, 50 Hz, 1430 rpm)	Stator resistance	1.4050 Ω
	Rotor resistance	1.3590 Ω
	Stator inductance	0.005839 H
	Rotor inductance	0.005839 H
	Pole pairs	2
	Friction factor	0.002985 N.m.s
	Inertia	0.0131 J/kg.m2
	Mutual inductance	0.1722 H
3-phase block of fault	Damage resistance	0.1 Ω
	Ground resistance	0.01 Ω
	Snubber resistance	106 Ω

outlines the key components, including the IM and fault block [22].

The selection of the 3-phase squirrel-cage motor is based on its extensive use in various electromechanical systems, including maritime applications. The model is developed to simulate the operation of the induction motor alongside a damaged block specifically designed to emulate faults. The detailed specifications of both the induction motor and the damage block are outlined in

Table 2. Different load conditions of step load torque.

Duration (s)	Load Torque (Nm)	Rotational Speed (rpm)
0–1	0	1499
1–2	26.72	1434
2–3	13.36	1468
3–4	6.68	1484
4–5	0	1499

, while Figure 2 illustrates the simulation setup within the MATLAB workspace. The focus of the damage analysis is on the stator winding during a short-circuit event. The Simulink 3-phase damage block is capable of simulating two types of faults: phase-to-phase and phase-to-ground short circuits. Specifically, six different scenarios are examined: phases A to ground (G), phases B to ground (G), phases C to ground (G), phases A–B, phases A–C, and phases C–B. Additionally, the level of damage, indicated by the resistance, is a crucial factor. The literature [23] suggests that lower resistance indicates a severe insulation failure, representing a complete interturn short circuit, whereas higher resistance can produce similar malfunctions that are more challenging to detect.

The simulation is conducted over 5 s, utilizing a sampling frequency of 20 kHz to capture the 3-phase stator currents. The torque of the induction motor is regulated by adjusting the load step function, which in turn affects the angular velocity, as detailed in Table 2.

Figure 6 illustrates the signals observed in both the healthy and damaged states, revealing that the transient phase at the start of the motor’s operation results in an overshoot. Furthermore, the fluctuations in the current time series align with the load pattern. The current signal for the damaged 3-phase motor, as shown in Figure 3, emphasizes the variations occurring between 1.4 and 1.7 s, which are indicative of the damage effect.

4. Damage Detection Framework

This section describes the damage detection framework, which involves two important components, the autoencoder and the Mahalanobis distance.

4.1. Proposed Framework

The damage detection framework consists of two distinct phases: the training phase (TP) and the inspection phase (IP). The TP involves the preprocessing of the training dataset, feature construction from the training dataset, the determination of the hyperparameters (number of neurons, activation function, etc.), the training of the autoencoder model, the computation of the Mahalanobis sample mean vector and sample covariance matrix, and finally the determination of a user-defined threshold. These actions are performed offline to establish a reference baseline for damage detection with which the new points are compared. The preprocessing uses a min–max scaler in the training data, while the feature construction uses the appropriate number of samples from the time series and feeds it into the autoencoder. In the TP, the user can compute the Mahalanobis distances of the training data and hence use them to derive information to help determine the threshold. On the other hand, the IP refers to the active deployment of the trained model for real-time processing, constantly calculating the Mahalanobis distance from the provided data. The procedures of preprocessing and feature construction are repeated in this phase in the same way as in the TP.

The health condition of the structure is assessed by comparing the computed Mahalanobis distance to the user-defined threshold set during the TP. If the Mahalanobis distance is less than or equal to the threshold, the condition is categorized as healthy. Conversely, the condition is classified as damaged if the distance is higher. Overall, this framework allows for automated decision-making, and the whole procedure is presented in Figure 7. The framework is implemented in the Python scripting language, and the TensorFlow [24] open-source library is utilized to train the neural networks.

The training phase was divided into four tasks.

TP Task 1: Feature Construction. This task generated features sensitive to damage from the preprocessed time series data, such as the phase current or voltage. A synchronized window technique selected m samples from each time series, forming a sequential feature vector with n components.

TP Task 2: Autoencoder Configuration. This task involved determining the input and output layer dimensions based on the feature vectors, typically including two layers between the input and bottleneck layers and between the bottleneck and output layers. Users can choose the number of neurons in these layers, following a guideline of reducing neurons from input to bottleneck and increasing from bottleneck to output. All layers except the output layer can have customizable activation functions, with the output layer using a rectified linear unit (ReLU).

TP Task 3: Mahalanobis Distance Calculation. This task calculates the sample mean vector and covariance using the Mahalanobis distance, which measures the dissimilarity based on the reconstruction error from a lower-dimensional error vector (k) relative to the feature vector (nxm). This distance relies on two hyperparameters, the sample mean and covariance matrix derived from k-dimensional error vectors in the training set, maintaining the framework’s unsupervised nature.

TP Task 4: User-Defined Threshold. Users can set decision thresholds based on intuition or statistical methods, potentially using the Mahalanobis distances from the training phase.

After model training, the inspection phase began with the collection of a new time series dataset matching the sample length from TP Task 1. These data were used to create a feature vector for the autoencoder’s reconstruction. The statistical distance between the reconstructed and original feature vectors was evaluated to assess the health status through a decision-making framework. The inspection phase included five tasks.

IP Task 1: Data Acquisition. Collect new time series data for each unique series, ensuring that the sampling rate matches the training samples.

IP Task 2: Feature Construction. Create a feature vector representing the current health status using the new time series data in the same order as during training.

IP Task 3: Compute k-Dimensional Error Vector. The autoencoder reconstructs the feature vector, generating a k-dimensional error vector by excluding nullified output values from the training data.

IP Task 4: Compute Mahalanobis Distance. Calculate the Mahalanobis distance using the k-dimensional error vector, sample mean, and covariance matrix from the training phase.

IP Task 5: Evaluate Structural Integrity. Assess the structure’s health by comparing the Mahalanobis distance to a user-defined threshold from training; if within the threshold, the structure is healthy; if not, it is deemed damaged.

The IP tasks are depicted in Figure 7, where f represents the feature vector and f hat represents the reconstructed feature vector.

4.2. Autoencoder Description

The framework is based upon the autoencoder, an auto-associative neural network that is designed to reconstruct its input as the output. The autoencoder is trained using only healthy data and ‘memorizes’ them to capture and retain the effects of the environmental and operating conditions (EOCs) in its weights and biases. Once a signal under a damaged structure is acquired and fed into the autoencoder, the error between the reconstruction and the input is higher than those from healthy signals and hence the anomalous case is differentiated from the nominal signal.

The autoencoder is composed of five sequential parts: the input, the encoder, the bottleneck, the decoder, and the output (Figure 8). The layers of the encoder and decoder are mirrored via the bottleneck, hence resembling an hourglass. The data are passed through the encoder, which has a descending number of neurons, and reach the bottleneck layers, where the data are compressed in a low-dimensional latent space. Subsequently, the decoder uses the data in the latent space to reconstruct the input of the model as the output. The number of layers in the encoder and decoder is selected to be two based on our previous study [15].

Neural networks, including autoencoders, rely on activation and loss functions. Activation functions modify neuron outputs using nonlinear methods to capture complex data relationships. Training an autoencoder minimizes the loss function, which measures the difference between the actual output and the input. Common loss functions are the mean absolute error and mean squared error, with lower values indicating better training effectiveness. The number of neurons, the activation function, and the learning rate are optimized for this study using the Bayesian optimizer and are presented in detail in

Table 3. Autoencoder architecture details.

Layer	Dimensionality	Activation Function
Input	2000	-
Encoder 1	1800	ReLU
Encoder 2	1400	ReLU
Bottleneck	600	Tanh
Decoder 1	1400	ReLU
Decoder 2	1800	ReLU
Output	2000	Sigmoid

. The autoencoder uses the Adam training algorithm and the loss function is the mean absolute error (MAE).

4.3. Mahalanobis Distance

The Mahalanobis distance is a statistical tool used to measure and quantize the distance between a multidimensional point and a multivariate distribution. In this study, the Mahalanobis distance is used to calculate the error vector of the new feature vector fed into the autoencoder compared to the error vectors of the training data. From the error vectors of the training data, the sample mean vector and the sample covariance matrix are calculated and used to compute the Mahalanobis distance. The benefit of the Mahalanobis distance is its ability to contribute to the robustness of the framework, consider the EOCs present in the system, model it in the sample mean and sample covariance matrix, and hence enhance the performance of the damage detection framework.

The Mahalanobis distance D equation is the following:

$$\mathrm{D}=\sqrt{{\left(x-\mu \right)}^{{\rm T}} {\left[\Sigma \right]}^{-1} \left(x-\mu \right)},$$

(1)

where x is the multidimensional error vector of the point, μ is the sample mean vector, and Σ is the semi-positive sample covariance matrix of the multivariate distribution of the error vectors.

This study utilized the Mahalanobis distance to evaluate the relationship between a feature vector and a collection of feature vectors within a multivariate distribution. The error vector represents the discrepancy between the reconstructed and original feature vectors, with its magnitude being less than that of the original. The sample mean and covariance matrix were calculated from the k-dimensional error vectors in the training dataset, illustrating that the damage detection framework functioned in an unsupervised manner.

Damage detection is determined by comparing the Mahalanobis distance (D) to a predefined threshold (lim). If D is less than or equal to lim, the feature vector signifies healthy signals and an intact structure. In contrast, if D surpasses lim, it indicates potential damage and structural issues. This straightforward framework, as described in Equation (2), effectively assesses the structural health through the Mahalanobis distance metric.

$$\begin{gathered} \mathrm{D} \le lim \to \mathrm{Healthy} \\ \mathrm{D} > lim \to \mathrm{Damaged} \end{gathered}$$

(2)

5. Results

The validation of the framework relies on analyzing a single current signal of phase A. The simulation time to produce this signal is 5 s and the sampling frequency is 20 kHz, resulting in a signal length of 100,000 samples.

The first segment of the signal, consisting of 20,000 samples (1 s), is deliberately omitted to disregard the transient response of the system, which leads to overshoot and non-stationarity, as already presented in Figure 3. This decision enhances the results and the framework’s detectability.

Additionally, white noise is added to the healthy signal to increase the number of healthy signals, prevent overfitting, and improve the framework’s robustness to variations. This strategy results in a dataset containing 10 healthy signals (80,000 samples per signal), with each signal divided into segments of 2000 samples (the feature vector size), finally producing a total of 400 healthy feature vectors. The 320 feature vectors (80%) are used for the training of the model and the remaining 80 (20%) for the validation of the model. An identical strategy of white noise addition is employed for the signals under damage, resulting in a dataset of eight signals, each consisting of 6000 samples (0.3 s), which yield 24 feature vectors representing the damage state. Overall, the assessment of the framework is based on 80 healthy and 24 damage inspection cases (104 in total).

The outcomes of the damage detection framework are illustrated in the scatter plot in Figure 9. In the scatter plot, a user-defined threshold of 10,000 is depicted with a red dotted line. The distances pertaining to healthy feature vectors, either training (blue color) or validation (green color), are under the threshold, while the feature vectors under damage (red color) result in distances that are over the threshold. This translates into detection accuracy of 100%. In Figure 4, it is worth noting that the interval between the distances of the healthy and damaged states is significant, indicating the high detectability of the damage and clear separability between the healthy and damaged states.

Figure 10 shows the confusion matrix, indicating that the model successfully identified only true positives and true negatives, as the results are concentrated along the diagonal. In addition,

Table 4. Metric calculations for AE framework for different white noise values.

Measure	Value W–N = 60 Eb/No.	Value W–N = 78 Eb/No.	Derivation
Sensitivity	0.9195	0.9195	TPR = TP/(TP + FN)
Specificity	1.0000	1.0000	SPC = TN/(FP + TN)
Precision	1.0000	1.0000	PPV = TP/(TP + FP)
Negative Predictive Value	0.9195	0.9195	NPV = TN/(TN + FN)
False Positive Rate	0.0000	0.0000	FPR = FP/(FP + TN)
False Discovery Rate	0.0000	0.0000	FDR = FP/(FP + TP)
False Negative Rate	0.0805	0.0805	FNR = FN/(FN + TP)
Accuracy	0.9581	0.9581	ACC = (TP + TN)/(P + N)
F1 Score	0.9581	0.9581	F1 = 2TP/(2TP + FP + FN)

TP: True Positive, FP: False Positive, TN: True Negative, FN: False Negative.

presents, in a quantitative way, the values of different measures for the sensitivity, precision, and accuracy of the framework for the two different values of white noise used: 60 and 78 Eb/No.

Table 4 illustrates that the framework operates with high efficiency, even under extreme white noise conditions (W–N = 78 Eb/No), demonstrating a strong capability for the detection of significant damage. Future investigations will involve testing various noise types and analyzing real-world data. However, the impressive performance of the proposed framework prompted a final phase in our research, which entailed comparing its results with an alternative method that has been validated for the same type of generator [22]. This alternative method utilizes a hybrid CNN-wavelet algorithm, where the initial signal is first transformed from the time domain to the frequency–time domain via wavelet analysis, followed by the application of a 2D CNN algorithm to the results of the wavelet analysis for damage identification. Figure 11 highlights the differences in sensitivity between the two methods when subjected to varying levels of white noise.

Sensitivity has been identified as the most suitable metric, as accurately detecting true positives in damage identification is critical for this specific application. The results indicate that the proposed framework demonstrates strong performance in most scenarios, whereas the alternative model exhibits a decline in sensitivity when faced with high levels of white noise.

6. Discussion

This section describes the business application of the damage detection framework previously presented. Additionally, future directions and industry solutions for predictive maintenance are recommended and discussed.

6.1. Main Benefits of the Architecture

Our research demonstrates that the suggested methodology serves as an effective approach for digitalization tools aimed at reducing the carbon footprints of vessels and enhancing their operational efficiency. Additionally, a significant advantage of the proposed framework lies in its ability to deliver a practical solution for onboard applications. Major challenges, such as data quality and the need for substantial computational power, are addressed through the distribution of tasks among self-contained agents, as outlined in this study. The primary innovations that facilitate this comprehensive solution are summarized below.

Novel technologies for data collection and fusion. 1. Handling data nearer to their origin can resolve various challenges related to the data size, volume, and speed. Additionally, organizations can gain advantages in several areas: (a) enhanced business agility—the quick creation and implementation of fog applications, along with the provision of Mobility as a Service by equipment manufacturers; (b) improved cybersecurity—similar strategies can be applied to safeguard fog nodes as in other segments of the IT infrastructure; (c) better privacy management—sensitive information is processed locally rather than transmitted to the cloud; (d) reduced operational costs—local processing helps to conserve the network bandwidth.

Artificial intelligence for predictive performance and maintenance management. The maritime industry faces challenges like vessel malfunctions, poor marine system performance, ship emissions, and rising operational costs due to system degradation. While extensive historical data from ship monitoring systems exist, they are often too complex to address the current environmental issues. Although planned maintenance systems (PMS) are widely used, only 2% of vessels adopt condition-based maintenance (CBM). Module 1 will concentrate on data collection, machine learning, and advanced modeling strategies.

Automated real-time lifecycle management decision support. While several companies have initiated the provision of vessel performance and efficiency monitoring, these offerings remain proprietary and non-reusable. To date, the literature has only presented conceptual descriptions of relevant methodologies. The proposed solution stands out as the sole collaborative option, accessible to all participants in the supply chain, with a business model centered on advanced applications utilizing AI and ML techniques.

6.2. Business Applications

Numerous use cases have been identified that underscore the importance of the proposed framework utilizing autoencoders. In addition to detecting damage in ship motors, identifying issues in the hull and diesel generators is also critical. The hull is prone to exhibit faulty behavior multiple times during a ship’s operational life, primarily due to extreme loading, challenging operating conditions, and aging, which often leads to lengthy maintenance processes. These factors necessitate regular maintenance interventions by regulatory authorities. Furthermore, ships frequently encounter harsh marine environments throughout their lifespan, which accelerates corrosion and biofouling—two prevalent types of long-term damage that can result in catastrophic incidents, environmental pollution, and significant economic repercussions.

7. Conclusions

Regular maintenance is crucial for the smooth operation of diesel generators and for the prevention of potential breakdowns. Historically, preventive maintenance has been utilized to enhance the reliability and reduce the repair costs. However, this approach can result in excessive maintenance expenses, which could be alleviated through an effective predictive maintenance platform. To address these issues, we propose a comprehensive predictive maintenance platform that leverages integrated digital twins, merging advanced technologies with innovative strategies. This platform is designed for application across ship fleets, to improve the efficiency of predictive maintenance practices. Our aim is to develop a platform that is not only innovative and cost-effective but also well accepted by ship personnel. It will feature state-of-the-art techniques for damage detection, diagnosis, and remaining useful life (RUL) estimation, specifically designed for vessels and their subsystems.

This study presents a module of the digital twin platform focused on damage detection. The module is designed as a versatile framework that operates efficiently with limited computational resources and does not require extensive training data. By training solely under healthy conditions, this framework is capable of addressing infrequent damage occurrences, such as those encountered during the operational lives of onboard power generators, which can have significant socioeconomic repercussions. To validate the framework, simulation data from an induction motor were utilized, specifically examining damage related to short-circuits. An inverse analysis approach [25] was employed to enhance the simulated data with varying levels of white noise. The findings demonstrated the framework’s proficiency in differentiating between healthy and damaged signals, achieving 100% accuracy in predicting the motor’s condition across most noise levels, while maintaining up to 99% accuracy in extreme noise scenarios.

An essential advantage of this framework is its ability to connect academic advancements in damage detection algorithms with the practical requirements of the industry. To this end, a comprehensive system architecture is introduced, featuring a deep learning training platform alongside a real-time module. Prior to achieving a turnkey solution, several additional steps must be undertaken, with the two most crucial being the validation of the framework’s accuracy using real-world data and the enhancement of the software platform to include a tool that allows users to efficiently train and validate models offline. Subsequently, these models can be implemented in the real-time module for immediate anomaly detection. Additionally, a distributed computing framework is suggested to facilitate horizontal scaling, ensuring timely data processing and bolstering the system’s resilience by mitigating the effects of potential failures. Our overarching goal is to develop a comprehensive predictive maintenance platform that monitors all critical systems of a ship through an integrated digital twin. This platform has the potential to deliver economic advantages to ship stakeholders while helping to prevent catastrophic incidents that could result in the loss of life and harm to the maritime environment.

During the preparation of this work, the authors used the Ahrefs tool to edit the English language of the text. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.