Development of a Fault Prediction Algorithm for Marine Propulsion Energy Storage System

Abstract

The transition to environmentally sustainable maritime operations has gained urgency with the International Maritime Organization’s (IMO) 2023 GHG reduction strategy, aiming for net-zero emissions by 2050. While alternative fuels like LNG and methanol serve as transitional solutions, lithium-ion battery energy storage systems (ESSs) offer a viable low-emission alternative. However, safety concerns such as thermal runaway, overcharging, and internal faults pose significant risks to marine battery systems. This study presents an AI-based fault prediction algorithm to enhance the safety and reliability of lithium-ion battery systems used in electric propulsion ships. The research employs a Long Short-Term Memory (LSTM)-based predictive model, integrating electrochemical impedance spectroscopy (EIS) data and voltage deviation analyses to identify failure patterns. Bayesian optimization is applied to fine-tune hyperparameters, ensuring high predictive accuracy. Additionally, a recursive multi-step prediction model is developed to anticipate long-term battery performance trends. The proposed algorithm effectively detects voltage deviations and pre-emptively predicts battery failures, mitigating fire hazards and ensuring operational stability. The findings support the development of safer and more reliable energy storage solutions, contributing to the broader adoption of electric propulsion in maritime applications.

1. Introduction

The maritime industry is undergoing a significant transformation in response to global environmental regulations, particularly the International Maritime Organization’s (IMO) 2023 greenhouse gas (GHG) reduction strategy, which aims to achieve net-zero emissions in international shipping by 2050 [1]. Given that the maritime sector accounts for 90% of global freight volume, it also faces high fuel consumption and increasingly stringent carbon emission regulations [2]. In order to comply with these regulations, various alternative propulsion technologies, such as liquefied natural gas (LNG), methanol, and lithium-ion (Li-ion) battery energy storage systems (ESSs), are being explored as transitional solutions. Among these, Li-ion batteries are considered a promising option due to their superior energy density and efficiency compared to other ESS technologies like ultra-capacitors and flywheels. While ultra-capacitors provide high power density, their overall efficiency is lower. Similarly, both ultra-capacitors and flywheels are designed for high power output over short durations, a requirement that is seldom needed in most marine vessels. Although flywheels have a longer lifecycle, their power density remains lower than that of Li-ion batteries [3].

The advantages of battery propulsion over conventional diesel-powered vessels are evident in CO₂ emissions comparisons. Figure 1 illustrates the cumulative CO₂ emissions of both systems at a speed of 10 knots over a 30-year period. Diesel-powered vessels emit approximately 120,000 tons of CO₂, whereas battery-powered vessels emit around 60,000 tons, demonstrating a 50% reduction in CO₂ emissions when using battery systems. The growth rate of diesel emissions is nearly twice that of battery systems, reinforcing the environmental benefits of battery propulsion.

The incorporation of these elements into marine propulsion systems presents considerable safety challenges, including thermal runaway, overcharging, and fire hazards. These issues must be addressed to guarantee the reliability of the system [4,5]. Additionally, electric propulsion systems involve complex power electronics and require a holistic analytical approach to manage mixed power delivery modes effectively [6]. Given these challenges, researchers are focusing on developing predictive maintenance strategies and fault prevention algorithms to enhance the safety and performance of marine battery systems.

Numerous studies have explored fault prediction methodologies for Li-ion batteries in marine ESSs. One primary area of research involves the creation of AI-powered predictive maintenance models that utilize machine learning techniques to identify failure patterns and enhance battery lifecycle management [7]. Experimental studies, including electrochemical impedance spectroscopy (EIS) tests and fault tree analysis, have highlighted the critical role of continuous monitoring in evaluating battery performance and anticipating failures before they develop into hazardous incidents [8,9]. The implementation of battery management systems (BMSs) with advanced diagnostic functionalities has been identified as a crucial factor in mitigating risks associated with Li-ion batteries in maritime applications [10]. Moreover, research on high-power electronic conversion in hybrid electric propulsion systems highlights the necessity of integrating adaptive control frameworks, such as model predictive control (MPC), to enhance energy efficiency and ensure stable system operation [11]. In addition, comparative analyses of different battery safety measures suggest that large-scale systems tend to exhibit more stable performance within acceptable risk levels compared to smaller configurations, emphasizing the need for tailored safety protocols [12].

In this study, a robust AI-based diagnostic algorithm is developed for fault prediction and prevention in marine lithium-ion ESSs. The proposed work involves analyzing battery failure data and conducting EIS tests to identify critical risk factors associated with thermal and electrical instabilities. This study employs Long Short-Term Memory (LSTM) models to predict voltage deviations and optimize failure detection accuracy through Bayesian hyperparameter optimization. Additionally, a recursive multi-step prediction model will be implemented to anticipate long-term battery performance trends and minimize cumulative errors in predictive analysis. By incorporating real-time monitoring capabilities and adaptive predictive maintenance strategies, this approach seeks to improve the reliability of marine propulsion battery systems. The findings from this study will facilitate the progression of safer and more efficient battery-powered vessels, thereby aiding the global transition towards sustainable maritime transportation.

2. Battery Cell Failure Analysis

Figure 2 illustrates the various failure scenarios within a BMS in an ESS, which are fundamental to developing an effective failure prevention and preservation algorithm. Battery failures can be categorized into three primary types: external short circuits, overcharge and over-discharge, and thermal runaway. External short circuits may result from severe physical damage, water infiltration, or battery penetration, leading to dangerous malfunctions. Overcharge and over-discharge conditions often arise due to BMS design flaws or inaccurate monitoring of battery parameters, ultimately causing excessive heating that accelerates battery degradation [13,14]. If overheating remains unchecked, it can damage the battery separator, leading to internal short-circuits that further exacerbate degradation and increase safety risks [15,16,17]. In order to ensure the stable and safe operation of an ESS, a proactive failure prevention algorithm is necessary to detect and mitigate battery degradation before thermal runaway occurs. By identifying failure conditions early, system reliability and safety can be preserved. An empirical investigation of these failure mechanisms was conducted by analyzing 30 battery cells under controlled conditions, including 2 normal cells, 8 short-circuited cells, 10 overcharged cells, and 10 over-discharged cells, each subjected to 10 charge–discharge cycles under normal operating conditions. The battery cell failure analysis was designed to simulate various failure scenarios, capturing battery behavior under extreme conditions. The framework of this analysis is illustrated in Figure 3.

2.1. Electrochemical Impedance Spectroscopy (EIS) Test

An EIS test was performed on the battery that underwent the failure test [9,18]. The tests were conducted to analyze electrochemical behavior distinguishing normal batteries from faulty ones under different failure conditions, including short circuit, overcharge, and over-discharge scenarios. The short-circuit conditions for the EIS test were evaluated at 80% and 100% SOC using 5 mΩ and 30 mΩ resistors, respectively. For the overcharge test, the voltage was systematically increased in 10% increments (0.17 V), starting from 110% (4.37 V) and reaching up to 150% (5.05 V). Similarly, the over-discharge experiment was conducted by systematically decreasing the voltage in the same 10% intervals, starting from 110% (2.33 V) and reaching 150% (1.65 V). A total of 28 cells were analyzed: 2 normal, 8 short-circuited, 8 overcharged (including 2 with Current Interrupt Device (CID) failure), and 10 over-discharged. The test aimed to assess variations in electrolyte resistance (R_s) and overall impedance behavior across different fault conditions, with a focus on identifying the causes of maximum voltage deviation in the low State of Charge (SOC) range. Constant current (CC) charging at a 0.5 C rate was applied to the batteries before discharging them to 2.8 V, ensuring uniform test conditions. A two-hour stabilization period followed, allowing the internal state of the battery to reach equilibrium. EIS measurements were taken across a frequency range of 10 kHz to 100 mHz at a test current of 50 mA, with the chamber temperature maintained at 25 °C to eliminate thermal variations. The EIS test profile is presented in

Table 1. EIS test profile.

State	Controls	Limits
0	Rest	Time > 1.000 h → Next sequence
1	CC Cha (I = 2.400 A)	Ecell > 4.200 V → Next sequence
2	CV 4.200 V vs. Ref	I < 50.000 mA → Next sequence
3	Rest	Time > 90.000 min → Next sequence
4	CC Dis (I = −2.400 A)	Ecell < 2.800 V → Next sequence
5	Rest	Time > 2.000 h → Next sequence
6	GEIS 50.000 mA from 10.000 kHz to 100.000 mHz
7	Rest	Time > 10.000 min → Next sequence

.

2.1.1. EIS Test Results

Electrolyte resistance (R_s) is a critical parameter in assessing battery performance and degradation trends, and to ensure accurate analysis, R_s values below 0 were excluded, as they were more influenced by external wiring effects than by the battery’s internal characteristics. This filtering minimized distortions caused by circuit influences, allowing for a more precise evaluation of electrochemical behavior. When configuring an electrochemical equivalent circuit, an evaluation of the relationship between open-circuit voltage (OCV) and terminal voltage confirmed that terminal voltage decreases as R_s increases, indicating reduced ionic conductivity due to electrolyte depletion, which directly impacts battery performance. Additionally, since the EIS test was conducted in the low SOC range, failures in individual cells within a battery pack led to localized voltage drops, potentially contributing to an overall voltage deviation. The R_s values obtained under various test conditions are presented in Figure 4, illustrating the impact of different failure scenarios on electrolyte resistance.

As shown in Figure 4a,b,d, the results indicate no significant changes in electrolyte resistance, suggesting that normal, short-circuit, and over-discharge conditions did not substantially impact R_s within the tested SOC range. This implies that while external resistance may influence current flow and thermal characteristics, it does not necessarily cause immediate shifts in the battery’s internal resistance. In contrast, the R_s values obtained under overcharge conditions (Figure 4c) exhibited a different trend. At 130% overcharge (4.71 V), a notable impedance shift was detected just before failure, indicating that as the battery approached a critical overcharge threshold, its internal resistance increased significantly. This rise in resistance is attributed to electrolyte decomposition and structural degradation, both of which negatively affect battery performance and stability.

For overcharge conditions at 140% (4.88 V) and 150% (5.05 V), impedance measurements were not feasible, as the battery’s built-in protection mechanism was immediately triggered upon current application. This safety feature effectively prevented further overcharging, safeguarding the system from catastrophic failure. These findings underscore the crucial role of protective mechanisms in mitigating hazardous conditions associated with excessive voltage application.

2.1.2. EIS Test Data Analysis

The EIS test results from overcharge and over-discharge experiments provide crucial insights into battery failure mechanisms and their impact on performance and safety. During the 130% overcharge condition, R_s exhibited a significant increase, indicating a reduction in ionic conductivity. This suggests that excessive voltage application led to electrolyte degradation, causing the breakdown of essential chemical components within the cell. The resulting rise in internal resistance limits efficient charge transport, degrading overall battery performance. As electrolyte depletion progresses, impedance increases, reducing energy transfer efficiency and further accelerating deterioration. If left unaddressed, these changes can lead to increased heat generation and heightened safety risks, including thermal runaway.

The progression of internal short circuits (ISCs) in batteries can be categorized into three distinct levels based on voltage stability and thermal behavior [19]. In Level I, a slow decrease in voltage is observed due to self-discharge, with no significant heat generation. This phase is primarily governed by electrical mechanisms, and the failure remains self-contained without external intervention. As the ISC progresses to Level II, a rapid voltage drop occurs, accompanied by noticeable joule heating. The effectiveness of heat dissipation determines whether the failure can be managed, with electrical and thermal coupling effects becoming dominant, indicating a more severe degradation state. In Level III, the voltage drops to zero, and an uncontrollable rise in temperature occurs due to a combination of joule heating and chemical reactions. This leads to massive heat release, accelerating separator collapse and ultimately resulting in thermal runaway. At this stage, failure is irreversible, posing severe safety risks. This progression highlights the importance of early detection and proactive thermal management strategies to prevent catastrophic battery failure. Monitoring internal resistance fluctuations, particularly in the low SOC region, is critical for identifying early ISC onset and mitigating cascading failures within battery packs. Moreover, fluctuations in R_s in the low SOC region had a pronounced effect on voltage stability. The experiments also revealed a noticeable voltage drop trend during over-discharge conditions, suggesting the onset of ISCs. If unmonitored, these internal failures can worsen over time, leading to irreversible damage and severe safety hazards, such as unexpected power loss and complete battery failure [19]. The presence of ISCs indicates localized separator degradation, facilitating unintended electron pathways and increasing self-discharge. Such conditions elevate the risk of cascading failures within battery packs, emphasizing the need for proactive monitoring and early intervention strategies to mitigate potential hazards and extend battery lifespan.

3. Estimation Model for Maximum Voltage Deviation

3.1. Target Definition for the Design of the Fault Prediction Algorithm

An ESS is not composed of a single battery but rather a combination of multiple batteries, including modules, racks, and banks. Therefore, failure analysis should not be limited to individual battery cells but should instead focus on the entire battery pack or larger units. Voltage deviations within a battery pack are influenced by several key factors that can impact both performance and safety. One of the primary causes of these deviations is internal chemical state changes within battery cells, which occur naturally over time and throughout charge–discharge cycles. Additionally, external conditions, such as temperature fluctuations, can further exacerbate voltage imbalances, leading to increased instability within the system. Figure 5 illustrates the voltage variations observed in a battery pack over time, providing insight into the fluctuations and deviations that occur during operation.

Fault conditions, including overcharging and over-discharging, also contribute significantly to voltage deviations, often accelerating cell degradation. Furthermore, if voltage deviations persist and accumulate over prolonged periods without corrective measures, they can compound, resulting in a more pronounced decline in system stability and efficiency. To manage these deviations, the BMS plays a crucial role in maintaining balance across the cells during ESS operation. By continuously monitoring and adjusting the voltage levels, the BMS ensures that deviations remain within acceptable limits, thus preventing excessive imbalances. Additionally, during the system’s idle period, voltage deviations tend to subside naturally to some extent, allowing the cells to regain a more stable state. However, if deviations are not properly addressed and continue accumulating over time, the overall performance of the ESS deteriorates, reducing its efficiency and lifespan.

In more severe cases, excessive voltage deviations caused by overcharging or over-discharging can lead to critical failures, significantly increasing the risk of hazardous incidents such as thermal runaway and fire accidents. Figure 6 illustrates the variation in capacity reduction among different cells within two battery packs (Pack A and Pack B) over multiple charge–discharge cycles. In both cases, individual cells degrade at different rates, leading to non-uniform capacity fading. In Figure 6a (Pack A), some cells retain higher capacity while others degrade more rapidly, indicating inconsistencies in aging behavior. Similarly, Pack B (Figure 6b) shows widening capacity differences among cells as the cycles increase, suggesting faster degradation in certain cells due to factors like internal resistance, thermal effects, or initial capacity variations. This uneven degradation can cause imbalance, reduce overall efficiency, and impact reliability, highlighting the need for effective battery management strategies to ensure the longevity of ESSs. To enhance system safety and reliability, the prediction of cell-to-cell voltage deviations has been identified as a key focus for fault detection and prevention. By implementing advanced fault prediction algorithms, potential failures can be detected early, allowing for proactive measures to mitigate risks and ensure the stable operation of the ESS.

3.2. Data Organization and Feature Extraction for Algorithm Design

In ESSs, data transformation is crucial for effective monitoring and analysis. The 24S1P battery pack structure is converted into an ESS data structure, enabling comprehensive assessment and control. To facilitate detailed analysis, the charging and discharging sections are treated separately, allowing for the extraction of voltage and current data specific to each phase. Within the charging section, the minimum, maximum, and average voltage values are extracted to further evaluate voltage deviations and identify potential irregularities. Analyzing individual cell voltages within the battery pack data is essential to detect inconsistencies and ensure system stability.

For medium-to-large-scale batteries, particularly in ESS applications, limiting the SOC within a controlled range (typically 10% to 90%) is a common practice to enhance system safety and prolong battery lifespan. It has been observed that the largest voltage deviations occur in the low SOC region, particularly in the 0–10% range. However, since ESS operation scenarios typically exclude this range due to its instability and potential risks, these deviations are considered inapplicable. A detailed analysis of a 24S1P pack dataset confirms that no significant deviations are present within the typical SOC range used in real-world ESS applications.

Voltage deviation analysis by SOC range, performed using a Min-Max scaler approach, reveals that the highest deviation within the applicable SOC range occurs in the 10–20% range. Interestingly, this deviation demonstrates a strong correlation (0.995) with the difference observed in the 0–10% SOC range, suggesting a pattern in the charge cycle behavior. Further investigation into the constant current (CC) charging area highlights that the key feature for predicting maximum voltage deviation is the difference between the minimum voltage in the CC area and the maximum voltage in the CC area (Max-Min CC Area). This difference serves as a critical input parameter (feature), while the maximum voltage deviation (Max Deviation) is designated as the corresponding output (label) for predictive modeling and system optimization. Max Deviation and Max-Min CC Area across the training and testing datasets over the cycle are shown in Figure 7.

3.3. Long Short-Term Memory (LSTM)-Based Maximum Voltage Deviation Estimation Model

The design of a maximum voltage deviation estimation model utilizes a Long Short-Term Memory (LSTM) model structured in a one-to-one configuration, consisting of 110.84 units. LSTM effectively mitigates the vanishing gradient problem inherent in traditional Recurrent Neural Networks (RNNs), allowing for better modeling of long-term temporal dependencies without extensive manual feature engineering [20,21]. Although simpler alternatives like the Gated Recurrent Unit (GRU) sometimes achieve comparable results, LSTM typically performs better for sequences with very long dependencies [22,23]. Compared to Convolutional Neural Network (CNN)-based methods, LSTM naturally captures long-range dependencies more effectively unless CNN architectures are specifically designed with dilated or stacked convolutions [24,25,26]. Additionally, while classical machine learning methods such as SVMs and decision trees may perform well when carefully tuned with handcrafted features, LSTM inherently learns complex temporal dynamics end-to-end [25,27]. These characteristics make LSTM particularly suitable for predictive tasks involving sequential data, such as voltage deviation forecasting in energy storage systems, offering a robust framework to capture intricate temporal patterns efficiently and intuitively. In line with the hyperparameter tuning method for ANNs outlined in [28], which involved optimizing key factors such as neuron count, hidden layers, activation functions, and learning rates to accurately forecast compressor performance, this research employs Bayesian optimization to meticulously adjust hyperparameters. This approach guarantees reliable and generalized predictions of voltage deviations. The systematic tuning process significantly improves both the precision and reliability of predictions, especially regarding intricate, nonlinear battery degradation patterns. The optimized hyperparameters include the number of layers, learning rate, and dropout rate. The learning process of the maximum voltage deviation estimation model and the hyperparameters optimized using the Bayesian optimization method are presented in Figure 8 and

Table 2. Optimized hyperparameters for model training.

Dropout Rate	Learning Rate	Units
0.1935	0.0095	110.84

, respectively.

In order to quantitatively evaluate the performance of the designed model, an assessment index is derived based on several key error metrics. The Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE) are used as primary evaluation criteria, with lower values indicating better performance. Additionally, the coefficient of determination (R²) serves as a complementary metric, where values closer to 1 signify a stronger predictive capability. By employing these evaluation indices, the effectiveness of the LSTM model in estimating maximum voltage deviation can be rigorously validated, ensuring its reliability in real-world ESS applications. The performance of the maximum voltage deviation estimation model, optimized through hyperparameter tuning, and its evaluation metrics are presented in Figure 9 and

Table 3. Performance metrics for maximum voltage deviation estimation model.

MAE	MSE	RMSE	R2
0.0142	0.00039	0.0199	0.9823

, respectively.

4. Prediction Model for Maximum Voltage Deviation

4.1. Maximum Voltage Deviation Prediction Model Data Configuration

The maximum voltage deviation prediction algorithm is designed to estimate the number of cycles required to reach a critical voltage deviation threshold, helping to mitigate failure or fire risks in ESSs. This model builds upon the previously developed maximum voltage deviation estimation model, using its output as an input to predict deviations in the following cycle. The prediction process involves a sequential sliding mechanism where the maximum voltage deviation at cycle K is used to estimate the deviation at cycle K + 1. The deviation is determined by subtracting the minimum voltage value from the maximum voltage value within a single cycle, with calculations extending up to 257 cycles. To achieve accurate predictions, the sliding window method is employed, allowing the model to analyze sequential patterns in voltage deviations. This method structures the dataset into overlapping segments, ensuring that temporal trends and anomalies are effectively captured. In this implementation, a window size of 2 is used, meaning two consecutive data points serve as input features for predicting the next cycle’s voltage deviation. With 257 cycles in total, this results in 256 rows of structured data, enhancing the model’s ability to learn time-dependent relationships. By leveraging this approach, the model improves its predictive accuracy and provides a reliable framework for monitoring voltage deviations in ESSs.

4.2. Data Smoothing Techniques for Model Performance Enhancement

To improve the stability and accuracy of the voltage deviation estimation model, smoothing techniques such as the Moving Average are applied. These techniques help maintain the overall trend of voltage deviation while minimizing noise and fluctuations, preventing short-term variations from adversely affecting predictions. By employing a window size of 10, the model effectively captures long-term patterns in the dataset, enhancing its generalization and reliability in ESS applications. The dataset is structured into two subsets for training and evaluation. The training dataset comprises 179 cycles, enabling the model to learn historical voltage deviation patterns, while the test dataset extends to 257 cycles, providing a robust framework for assessing the model’s predictive accuracy on unseen data. Since the voltage deviation data does not show a consistent increasing trend, direct prediction without smoothing may lead to inaccurate results, particularly when encountering fluctuations. The Moving Average method is used to stabilize the prediction trend by averaging data points over a defined period, ensuring that the model focuses on the general deviation pattern rather than minor variations. This structured approach allows for more effective failure prediction and preservation, enhancing the model’s applicability in real-world ESS monitoring and management. The application of a smoothing technique to the voltage deviation prediction model to enhance its performance is presented in Figure 10.

4.3. Long Short-Term Memory (LSTM)-Based Single-Step Prediction Model

The maximum voltage deviation prediction model is designed using an LSTM network with a one-to-one structure, incorporating 110.84 units. This architecture is chosen for its ability to effectively capture sequential dependencies in voltage deviation data, making it well suited for time-series forecasting in ESSs. The model ensures accurate predictions of voltage deviations, aiding in the proactive management of battery performance and failure prevention. To further enhance predictive accuracy, hyperparameter optimization is conducted, focusing on key parameters such as the number of layers, learning rate, and dropout rate. Bayesian optimization is employed as a systematic approach to identifying the optimal combination of these hyperparameters, similar to the process used for the estimation model. The learning process of the maximum voltage deviation estimation model, with hyperparameters optimized using the Bayesian optimization method, is presented in Figure 11.

The model’s performance is thoroughly assessed using multiple evaluation metrics, including MAE, MSE, and RMSE, where lower values indicate higher predictive accuracy. Additionally, the R² score is used to measure the model’s explanatory power, with values approaching 1 reflecting a strong correlation between predicted and actual voltage deviations. The final evaluation results show that the model achieves MAE, MSE, and RMSE values close to zero, with an R² score of 0.9976, demonstrating high reliability. By integrating LSTM-based predictive modeling with Bayesian optimization, the model provides an effective solution for monitoring and forecasting voltage deviations in ESSs, contributing to enhanced operational stability and failure prevention. The performance of the maximum voltage deviation prediction model, optimized through hyperparameter tuning, and its evaluation metrics are presented in Figure 12 and

Table 4. Performance metrics for maximum voltage deviation prediction model.

MAE	MSE	RMSE	R2
0.0048	0.0004	0.0059	0.9976

, respectively.

4.4. LSTM-Based Recursive Multi-Step Prediction Model

The prediction of voltage deviation progression in energy storage systems (ESSs) is conducted using a Long Short-Term Memory (LSTM) model trained with historical deviation data. This model enables sequential forecasting by predicting the deviation at each subsequent step based on previously observed values. The approach is particularly effective in capturing time-dependent fluctuations, which play a critical role in ESS performance and safety. The input feature for the model is the voltage deviation at the previous time step, while the label represents the deviation at the next time step. By iteratively feeding predicted values back into the model, a continuous sequence of estimated deviations is generated. To enhance predictive accuracy, a recursive multi-step prediction model is implemented. This technique leverages the output from the LSTM-based one-step voltage prediction model and recursively reuses each predicted value as input for the next prediction. The process continues until a predefined voltage deviation threshold is reached. The recursive multi-step prediction model is illustrated in Figure 13. This recursive approach improves data efficiency by reducing computational cost while ensuring the temporal flow of time-series data is effectively captured. In this study, the stopping criterion for the prediction process is set at 0.5195 V, beyond which system failure risk significantly increases. Once this threshold is met, the model halts further predictions, issuing an early warning for potential failures. By utilizing this recursive prediction strategy, the model enables proactive maintenance and enhances the safety of ESS operations.

5. Development of Failure Prevention Algorithm

5.1. Voltage Deviation Prediction Model Enhancement

In the overall aging process of the battery pack, the lowest voltage was continuously observed for one battery as the voltage deviation accelerated. When analyzing the battery pack data used in this algorithm, the batteries in which the lowest voltage was observed varied from the initial 1 cycle to the 179 cycle, which is the time when the voltage deviation accelerated, but after that, it was confirmed that battery 11 continuously maintained the lowest voltage as the voltage deviation accelerated. The contents are shown in Figure 14. In addition, it is necessary to analyze the phenomenon from the viewpoint of predicting the maximum voltage deviation based on artificial intelligence. Figure 15 shows that the prediction results are similar when the maximum voltage deviation value at the 100 cycle point is input and when the maximum voltage deviation value at the 179 cycle point where the voltage deviation is accelerated is input. From the perspective of the pre-diction model, the maximum voltage deviation value at the 100 cycle point or the voltage deviation value at the 179 cycle point is similar, so the prediction result that the maximum voltage deviation increases regardless of the time point is output.

In order to prevent this phenomenon and improve the performance of the maximum voltage deviation prediction model, it is necessary to define the time point at which the voltage deviation is accelerated according to the cycle. Therefore, before the maximum voltage prediction model operates, an algorithm was added to determine whether the location of the battery where the lowest voltage in the battery pack was observed in the cycle data converged to one battery.

5.2. Design and Verification of Failure Prevention Algorithm

By integrating the maximum voltage deviation estimation model within the cycle, the maximum voltage deviation prediction model to the dangerous voltage deviation point, and the Min. Cell Number convergence algorithm, a failure prevention preservation algorithm was designed. To verify the failure preservation algorithm, data after the 179 cycle were not used in the training data, and the held pack data were converted into an ESS data structure to simulate the situation in which ESS data are input. Unlike laboratory-based data, the data output from the BMS installed in the actual ESS provides only limited information. This is because the actual ESS structure is a large-scale battery system composed of Bank, Rack, Pack, Module, and Cell, so there is a limit to providing information for all cells. Therefore, the converted structure is Max Cell Voltage, Max Voltage Cell Number, Min. Cell Voltage, Min. Voltage Cell Number, Avg Voltage, Current, and Pack Voltage, which are the same as the information given in the actual ESS.

When the structural-transformation-completed battery pack data are input to the failure prediction and preservation algorithm, the maximum voltage deviation estimation and maximum voltage deviation prediction model will not operate until the 179 cycle by the cell convergence algorithm. When data from the 179-cycle time are input, the model operates as if the battery where the lowest voltage in the data is observed converges to Battery 11. Then, the SOC 10% to 20% section in the 179-cycle data is calculated, and if it is confirmed that a positive current is applied, the area under the maximum voltage curve and the area under the lowest voltage curve in the section are calculated. The corresponding voltage value is calculated by subtracting the area under the maximum voltage curve and the area under the lowest voltage curve.

It is input to the deviation estimation model, and accordingly, the maximum voltage deviation value in the data is output. The output maximum voltage deviation value is input to the maximum voltage deviation prediction model and calculates the number of cycles that it takes to reach the set threshold. As a result, the failure prediction and preservation algorithm predicts and tells the user whether the ESS needs failure preservation, the location of the cell requiring preservation in the ESS, and the time point taken to reach the voltage deviation of the dangerous level. The operation flow of the fault prediction conservation algorithm is shown in Figure 16.

The fault prediction maintenance algorithm provides crucial insights into battery health by analyzing voltage deviations within the pack. When cells with large voltage deviations converge, it signals a potential risk, highlighting the need for failure prediction maintenance. Additionally, the algorithm identifies the specific cell at risk by determining its position within the pack, allowing for targeted monitoring and preventive action. Furthermore, it predicts when the voltage deviation will reach a critical threshold, estimating the number of cycles before failure occurs. This information is essential for end users, enabling proactive maintenance strategies to enhance battery reliability and prevent unexpected failures.

6. Conclusions

This study presents the development of an AI-based fault prediction algorithm for lithium-ion battery energy storage systems (ESSs) in marine applications. With increasing environmental regulations, battery-powered vessels are becoming a key transitional solution to reducing greenhouse gas emissions. However, safety concerns such as thermal runaway, overcharging, and voltage deviations necessitate predictive maintenance strategies. By leveraging experimental data and electrochemical impedance spectroscopy (EIS) analysis, this research identified voltage deviation as a critical indicator of battery failure.

The proposed algorithm integrates a maximum voltage deviation estimation model and a maximum voltage deviation prediction model, employing Long Short-Term Memory (LSTM) networks optimized through Bayesian hyperparameter tuning. The recursive multi-step prediction model was implemented to enhance long-term failure forecasting by minimizing cumulative prediction errors. The model’s hyperparameters were optimized to 110.84 LSTM units, a learning rate of 0.0095, and a dropout rate of 0.1935. Performance evaluation showed a Mean Absolute Error (MAE) of 0.0048, a Root Mean Squared Error (RMSE) of 0.0059, and an R² score of 0.9976, indicating high predictive accuracy. The recursive multi-step model initially underestimated failure onset, predicting 57 cycles to reach a 0.5195 V deviation threshold, but post-error correction, this improved to 76 cycles, closely aligning with the actual 78-cycle failure onset.

Various factors, including battery geometry, composition, and electrochemical properties, influence failure patterns, making it essential to refine the model using diverse datasets and real-world operating conditions. This study highlights the necessity of continuous learning and adaptive algorithms to improve fault detection accuracy and mitigate safety risks. Given the increasing incidence of lithium-ion battery-related fire hazards, ensuring battery system reliability is crucial. The proposed approach, if further optimized, holds significant potential for enhancing maritime battery safety, reducing operational risks, and supporting the global transition to sustainable electric propulsion systems.