Maintenance Time Prediction for Predictive Maintenance of Ship Engines

Abstract

Ships carrying large amounts of cargo and passengers are larger and slower than other modes of transportation. They are mostly foreign flagged and operate at sea far from coasts for 20 years or more, incurring more operating costs than construction costs. Therefore, an efficient maintenance system is necessary for stable, economical ship operation. Researchers are attempting to equip ships with predictive maintenance technology, which is used proactively in other modes of transportation to predict the maintenance time of machines through data monitoring and analysis. However, due to the nature of ship operation, data collection is difficult, and most studies focus on fault detection, hindering the application of predictive maintenance to ships. In this study, we developed a maintenance time prediction algorithm using the revision generator engine condition criterion (RGCCV) value and the cylinder exhaust gas temperature, as developed in a previous study for marine generator engines. And through comparison and verification using machine learning, the average mean absolute error (MAE) across all cylinders was 2.916 for the RGCCV-based method and 8.138 for the temperature-based method, demonstrating a 64% improvement. These findings establish a practical foundation for implementing predictive maintenance in ship engines by enabling more reliable and condition-based maintenance.

1. Introduction

Ships that carry large amounts of cargo and passengers are larger than other modes of transportation, incurring higher construction costs. Most of them are foreign flagged, which are typically slow and operate for more than 20 years at sea, far from coasts. Consequently, their operating costs, including labor, fuel, and maintenance, are approximately two to three times their construction costs, depending on ship type and size [1,2,3,4].

The unexpected operational interruptions of ships incur significant economic losses. Thus, shipping companies establish and operate maintenance systems that consider the operational characteristics of their ships for stable navigation. Because maintenance methods and levels significantly affect management costs, shipping companies strive to minimize management costs through optimal ship operation and efficient maintenance [5,6,7,8,9,10,11,12]. However, current preventive maintenance systems, which are based on uptime cycles, are excessively costly. Therefore, a more efficient maintenance system is required for stable, economical ship operation while avoiding service interruptions. Researchers are attempting to equip ships with predictive maintenance technology, which is used in various industries and other modes of transportation. It involves collecting and analyzing data on the operating condition of machinery in real time using artificial intelligence to identify operational abnormalities and predict the maintenance time [2,3,4,6,8,10,11,13,14]. However, ships perform preventive maintenance, hindering the collection of data on operational abnormalities, and operational characteristics limit data collection, increasing the difficulty of practical research.

Park et al. [6,8], Bae et al. [9], and Kim et al. [10,11] detected abnormal conditions by analyzing data collected from ship engines using machine learning and establishing abnormality criteria [6,8,9,10,11]. However, the research was conducted using data before and after major failures that stopped engine operation, or by analyzing key monitoring data for engine maintenance, establishing thresholds, and detecting abnormal operations by analyzing data collected during normal navigation; engine combustion mechanisms and the dynamic marine environment were not considered. In addition, several recent studies, including those by Yigin et al. [15] and Bukovac et al. [16], have attempted to collect and analyze operational data from ship engines to improve the efficiency of anomaly detection and fault diagnosis. However, research on predicting maintenance time after anomaly detection remains limited and underdeveloped in the maritime domain. As a result, the practical implementation of predictive maintenance is still challenging, especially under normal navigation conditions where identifying true anomalies in real time is inherently difficult. These limitations highlight the need for further improvements in predictive strategies tailored to the unique operational characteristics of marine engines.

Park and Oh [2,3,4] developed anomalous symptom detection algorithms for propulsion and generator engines (GEs) that can detect anomalous symptoms considering the engine status and operating conditions during normal operation, when the engine parameters are below the manufacturer’s recommended thresholds for exhaust gas and lubricant temperatures [2,3,4]. However, current studies on the predictive maintenance of ships mostly focus on detecting anomalous symptoms. Practical predictive maintenance requires (1) criteria for determining when to perform maintenance after detecting anomalous symptoms and (2) predictive methods for the maintenance time.

Therefore, in this study, we developed an algorithm for predicting the maintenance time of ship GEs using the revision GE condition criterion value (RGCCV) and the cylinder exhaust gas temperature, which were proposed in a previous study [4]. Then, through comparison and verification using machine learning, we established the foundation for the implementation of the effective predictive maintenance of marine GEs.

2. Materials and Methods

2.1. Experimental Equipment

The research ship was a 10,000-ton ship equipped with various technologies, including a dynamic positioning system and a controllable pitch propeller. The ship, whose specifications are detailed in

Table 1. Specifications of research ship.

Description	Specification
Gross tonnage (GT)	9196 tons
Length overall (LOA)	133.0 m
Breadth (Mlb)	19.4 m
Speed (design draft, 85% MCR with 15% SM)	17.7 knots
Range of endurance	14,500 n.miles
Power of propulsion engine	6618 kW/146 RPM
Controllable pitch propeller	D 4000 MM × 4 blades, Wärtsilä, Helsinki, FIN

, also had a propulsion engine with a rated power of 6618 kW and a cruising speed of 17.7 knots.

The target engine was a ship GE with a rated power of 1120 kW. It was a 900 rpm, constant-speed four-stroke diesel engine. The generated electric power is supplied to the ship’s main switchboard, which distributes electricity to various systems and equipment essential for ship operations. Depending on ship type and operation scenario, three to four generators are typically installed to flexibly manage power demand. Fuel supply is controlled via a governor to adjust the engine load accordingly. Normally, only one generator is operating during normal navigation, considering the operating time; one to three generators are operating in parallel during port entry and exit, berthing, and canal transit, and the engines are immediately used during port entry/exit and equipment loading/unloading. During navigation, the engine load fluctuates frequently due to systems that are repeatedly switched on and off according to the operating conditions, such as air compressors and the compressors used for air conditioners and refrigerators. Frequent load fluctuations also occur during port entry, port exit, and berthing due to port entry/exit operations and equipment loading/unloading. The detailed specifications of the GE are in

Table 2. Specifications of generator engine (GE).

Description	Specification
Type	HiMSEN 6H21/32
Manufacturer	Hyundai heavy industry, Ulsan, KR
Rated output	1120 kW/900 RPM
Number of cylinders	6
Cylinder diameter	210 mm
Stroke	320 mm
Compression ratio	17:1
Mean effective pressure	22.5 bar
Maximum cylinder pressure	179 bar
Turbo charger	ABB, 1 × A135-M65

.

Figure 1 shows the schematic structure of data acquisition from the experimental ship’s generator engine and the simulation-based learning data. Operating information from the generator engine is acquired via various sensors installed during ship construction, as designated by the shipyard and engine manufacturer. These sensors transmit data through an I/O module to the alarm monitoring system (AMS) located in the engine control room, where they are monitored and stored. The dataset consists of 42 types of time-series numerical data, including cylinder exhaust gas temperature, power, scavenge air pressure, and temperature, all recorded at 10 s intervals during ship operation.

Based on the specifications of the real generator engine, a digital asset was constructed using AVL CRUISE^TM M (R2021.1) to simulate abnormal operating conditions. The system consists of a data acquisition module that collects engine condition parameters (e.g., RPM, torque, exhaust gas temperature, maximum combustion pressure) based on the real ship generator engine. These data are utilized to configure a 1D engine simulation model, which includes major subsystems such as the turbocharger, fuel injection system, scavenge air path, and load profiles. The simulation was used to generate additional abnormal conditions data for model training. The detailed specifications of the software, including version and licensing information, are presented in

Table 3. Software and simulation environment specifications.

Item	Description
Simulation software	AVL CRUISE TM M
Software version	R2021.1
Developer	AVL List GmBH, Graz, AT
License type	Academic license
Simulation domain	1D engine system modeling and dynamic performance
Host environment	Window 10 Pro 64-bit
CPU	Core i9–12900 K, Intel, Santa Clara, CA, USA
GPU	GeForce RTX 2070 SUPER (8 GB), NVIDIA, Santa Clara, CA, USA
Python version	3.9.13

.

In total, 42 types of numerical data verifying the engine status during four voyages (VOYAGE 21101, 21102, 21111, and 21112) were analyzed. A voyage is a period during which the ship departs from its home port, takes on cargo at ports of call and loading ports, and returns to the home port to discharge the cargo. Although the criteria for classifying voyages vary between shipping companies and ships, a voyage serves as a reference point for ship scheduling and management. The frequency of generator operation among the collected data was high, and the minimum operating load exceeded 25% due to the continuous equipment operation during normal navigation and berthing. Thus, only data where the engine load exceeded 25% (i.e., above 257.9 kW) were used in this study.

The data of a voyage with anomalous symptoms (VOYAGE 21113) implemented through engine simulation using AVL CRUISE^TM M in a previous study [4], which showed an increase in the exhaust gas temperature (the current criterion for maintenance decisions), were utilized as data with anomalous symptoms in this study.

2.2. Experimental Methods

A dedicated data preprocessing algorithm was constructed for the generator engine to ensure reliable training data for machine learning-based prediction, as shown in Figure 2. Data from the ship’s AMS were initially grouped and merged by voyage and engine system unit (e.g., lubricating oil, exhaust gas, scavenge air, fuel consumption), followed by the removal of duplicate rows and null entries. Then, based on the operating characteristics of the generator engine, which runs at a constant rpm, error data such as non-operating engine segments, missing AMS values, and abnormal consumption values were excluded using predefined rules.

Additionally, the engine status was clarified into loading and unloading phases using power-based thresholds (GE gen power > 0 for loading and ≤0 for unloading) to distinguish idle periods. This preprocessing procedure ensures that the input dataset reflects normal engine operations and allows for accurate anomaly detection and maintenance time prediction.

Predicting the maintenance time after detecting anomalous symptoms is important for predictive maintenance. The proposed prediction algorithm for the GE maintenance time was built using the RGCCV. This value was derived as a standard generator condition factor in a detection algorithm for a GE’s anomalous symptoms in a previous study [4]. The RGCCV is defined by Equation (1) [4]. The performance of the constructed maintenance time prediction algorithm was analyzed as follows: The results of maintenance time prediction using only the cylinder exhaust gas temperature, anomalous symptom, and maintenance decision criteria for the GE were compared with the results of RGCCV-based maintenance time prediction, and the following prediction performance metrics were calculated: mean squared error (MSE), root MSE (RMSE), and mean absolute error (MAE).

$$RGCCV=GMDCT-\left\{\begin{array}{c}GCET\\ +\left[\begin{array}{c}\left(GELR\ast (GLC-GCL)\right)\\ +\left(GESTR\ast (GSTC-GCST)\right)\\ +\left(GESPR\ast (GSPC-GCSP)\right)\end{array}\right]\end{array}\right\},$$

(1)

• GMDCT: generator maintenance decision criteria temperature
• GLC: generator load criterion
• GSTC: generator scavenge temperature criterion
• GSPC: generator scavenge pressure criterion
• GELR: generator exhaust gas temperature/load regression coefficient
• GESTR: generator exhaust gas temperature/scavenge temperature regression coefficient
• GESPR: generator exhaust gas temperature/scavenge pressure regression coefficient
• GCL: generator collected load
• GCST: generator collected scavenge temperature
• GCSP: generator collected scavenge pressure
• GCET: generator collected exhaust gas temperature

The GE maintenance time prediction algorithm was created by analyzing the RGCCV change tendency of the voyage with the detected anomalous symptoms (VOYAGE 21113) or a specific time period. The created algorithm should be configured per cylinder, as the cylinders have different normal operating conditions, such as the exhaust gas temperature and maximum combustion pressure [3,4].

Regression analysis was conducted on the RGCCV change trend of VOYAGE 21113 according to the operating time of the GE. Then, the intercept of the derived regression coefficients (such as the prediction intercept by the RGCCV [PIRG]) and the slope (such as the prediction slope by the RGCCV [PSRG]) were utilized as factors for the prediction equation. The predicted maintenance time by the RGCCV (PTMRG), as expressed in Equation (2), is the predicted operating time of the engine until the maintenance criteria are met. The standard for the operation time was expressed in seconds (s) and divided by 3600 to obtain the number of hours.

$$PTMRG={\displaystyle \frac{\left[\frac{\left(MCRG-PIRG\right)}{PSRG}\right]}{3600}},$$

(2)

• MCRG: RGCCV as maintenance criterion
• PIRG: prediction intercept by the RGCCV
• PSRG: prediction slope by the RGCCV

The maintenance criterion RGCCV (MCRG) is the RGCCV at the point where the GE requires maintenance (i.e., the maintenance standard factor). The temperatures of all the cylinders of GE no. 2 under normal operating conditions (VOYAGE 21101–21112) were analyzed (

Table 4. Analysis data of cylinder exhaust gas temperature of GE no. 2.

Item\Cylinder (°C)	1	2	3	4	5	6
Mean	459.5	456.5	471.9	453.1	456.0	480.7
Std	22.1	22.4	19.0	19.6	19.2	19.2
Min	343.8	343.8	348.8	334.9	333.9	364.9
25%	443.9	440.9	458.9	440.0	443.9	467.9
50%	460.9	456.0	473.9	456.0	459.9	482.0
75%	477.9	473.9	487.8	468.9	470.8	497.0
Max	523.0	524.9	536.9	511.8	515.0	549.0

). The RGCCV temperature criterion for maintenance decisions was set to 550 °C based on the maximum temperature (549 °C) of cylinder no. 6, which had the highest exhaust gas temperature among all cylinders [4]. Then, the MCRG was set considering the normal operating temperature range per cylinder.

The average temperature of cylinder no. 6 was 481 °C, deviating by approximately 70 °C from the temperature criterion for maintenance decisions of 550 °C, based on which the MCRG of cylinder no. 6 was set to 0. In other words, the maintenance time was reached when the average exhaust gas temperature of each cylinder rose by 70 °C (when the average temperature reached 550 °C, where the MCRG of cylinder no. 6 was 0). As the criterion for maintenance was the time the average exhaust gas temperature of each cylinder increased by 70 °C in normal operation, the MCRG was set differently per cylinder. Because anomalous symptoms were determined using the average data of the voyage, the MCRG was set based on the average exhaust gas temperature of the cylinders. The maintenance time (MCRG) must be determined according to the operating and maintenance standards and guidelines of each engine.

Given the GE’s current operating time (COT) and PTMRG, the proposed algorithm was developed to calculate the remaining operating time until maintenance, or the predicted time of remaining operation, using the PTMRG (PTROPRG), as seen in Equation (3).

$$PTROPRG=PTMRG-COT,$$

(3)

• COT: current operating time

The maintenance time predicted using the exhaust gas temperature (PTMET), or the time of generator maintenance predicted using only the exhaust gas temperature of the cylinder, was configured as follows using Equation (2). The MCRG value in Equation (2) was changed to the temperature criterion for maintenance (MCET), and the regression analysis of the engine operating time and cylinder exhaust gas temperature for VOYAGE 21113 was used to derive the regression coefficients. Here, the PTMET was defined as Equation (4), the remaining operating time until maintenance predicted using the PTMET (PTROPET) was defined as Equation (5), and the algorithm was constructed.

$$PTMET={\displaystyle \frac{\left[\frac{\left(MCET-PIET\right)}{PSET}\right]}{3600}},$$

(4)

• MCET: exhaust gas temperature as maintenance criterion
• PIET: prediction intercept by the exhaust gas temperature
• PSET: prediction slope by the exhaust gas temperature

$$PTROPET=PTMET-COT,$$

(5)

• COT: current operating time

The MSE, RMSE, and MAE were compared for a quantitative evaluation of the validity of the algorithm for predicting the maintenance time using machine learning. Moreover, a suitable performance metric was selected for assessing the GE maintenance time prediction algorithm by reviewing the characteristics of the data and the performance metrics.

The MSE is an important loss function that squares and averages the error between recorded and predicted values. Thus, it is sensitive to outliers due to squaring and depends on the data scale, with errors tending to increase with the data scale, as defined in Equation (6) [17,18]. The RMSE function uses the root of the MSE to reduce the error magnitude and convert it into values similar to the collected values, thereby facilitating analysis. It addresses the drawbacks of the MSE but remains dependent on the data scale and follows the standard deviation principle, as defined in Equation (7) [17,18,19]. The MAE function averages the absolute value of the error between the predicted and collected values; it is less affected by outliers than the other loss functions. This metric indicates the degree of the average error that depends on the data scale and is used to evaluate regression performance, as defined in Equation (8) [17,19,20,21,22].

$$MSE={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2,$$

(6)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2},$$

(7)

$$MAE={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n\left(\left|y_i-{\widehat{y}}_i\right|\right),$$

(8)

• $n$: sampling number
• $y_i$: collected value
• ${\overline{\widehat{y}}}_i$: predicted value

Algorithm development and machine learning were implemented using Python ^TM 3. Python easily uses various libraries needed for data visualization, image processing, statistics, and other purposes; with open-source packages, various machine learning libraries can be used directly in Python [23,24,25]. Python was used in this study due to its high development efficiency. Its simple syntax construction reduces error rates in complex code, its concise code construction accelerates work, and it easily links with other programming languages and libraries [23,24,25].

3. Results and Discussion

The results of the maintenance time prediction algorithm based on the RGCCV and that based on the exhaust gas temperature of the cylinder were analyzed and compared, and the results are as follows. Figure 3 shows the exhaust gas temperature of cylinder no. 6 of GE no. 2 during operation, specifically its changes with the (a) operating time and (b) power output. In the AMS, the normal operating range for the cylinder exhaust gas in the GE is below 580 °C (yellow dotted line in Figure 3), and the data of VOYAGE 21113 cannot be identified based on the exhaust gas temperature alone [4].

Figure 4 shows the RGCCVs of the RGCCV-based algorithm considering the operating environment of cylinder no. 6 of GE no. 2 with the (a) operating time, along with the (b) voyage averages. In Figure 4, the yellow dotted line is the reference line for judging the concern stage (1) and the red dotted line is the reference line for judging the abnormal stage (2). Unlike the exhaust gas temperatures in Figure 3, which show simple changes, the RGCCVs in Figure 4a exhibit a decreasing trend versus the operating time for VOYAGE, indicating anomalous symptoms. These anomalies are further highlighted by the average voyage data in Figure 4b [4].

The MCRG, used to calculate the PTMRG (the result of the RGCCV-based maintenance time prediction algorithm), was selected for each cylinder (

Table 5. Calculation factors for predicted maintenance time by RGCCV (PTMRG).

Item\Cylinder	1	2	3	4	5	6
MCRG	20	20	10	20	20	0
PIRG	886.73	819.27	939.94	924.18	1037.33	1026.28
PSRG	−9.64 × 10−5	−8.79 × 10−5	−10.4 × 10−5	−10.0 × 10−5	− 11.4 × 10−5	−11.6 × 10−5

) by analyzing the exhaust gas temperature of the no. 2 GE in Table 4.

Figure 5 shows the (a) regression analysis results of the RGCCVs of cylinder no. 6 of GE no. 2 in VOYAGE 21113 versus the engine operating time, along with the (b) calculated PRMRG and PTROPRG. Figure 5a indicates that PIRG and PSRG (Table 5) are derived through the regression analysis of the engine operation time and RGCCV. Then, the GE maintenance time prediction algorithm is used to predict the maintenance time for cylinder no. 6 of GE no. 2 (Figure 5b).

The PTMRG value for cylinder no. 6 of GE no. 2 is 2465.9 h, and PTROPRG is 118.2 h. As for the maintenance time prediction algorithm based on the exhaust gas temperature of the cylinders, PTMET and PTROPET were calculated by setting MCET for each cylinder of GE no. 2 (

Table 6. Calculation factors for maintenance time predicted using exhaust gas temperature (PTMET).

Item\Cylinder	1	2	3	4	5	6
MCET	530	530	540	530	530	550
PIET	−467.35	−400.17	−499.28	−490.35	−597.72	−586.33
PSET	10.9 × 10−5	10.1 × 10−5	11.5 × 10−5	11.1 × 10−5	12.5 × 10−5	12.6 × 10−5

) based on the MCRG (‘X’ mark in Figure 5b) values in Table 5. As shown in Figure 6a, PIET and PSET were set through the regression analysis of the exhaust gas temperature and the engine operation time for cylinder no. 6 of GE no. 2 in VOYAGE 21113 (Table 6). PTMET and PTROPET were calculated, as shown in Figure 6b based on the MCET (‘X’ mark in Figure 6b) values in Table 6.

For cylinder no. 2 of GE no. 6, the PTMET is 2498.3 h and the PTROPET is 150.6 h, predicting a maintenance point 32.4 h later than the PTMRG (2465.9 h) and PTROPRG (118.2 h). As shown in

Table 7. Prediction of maintenance time of all cylinders of GE no. 2.

Item\Cylinder	1	2	3	4	5	6
PTMRG (h)	2497.7	2523.5	2477.7	2509.2	2478.3	2465.9
PTROPRG (h)	150.0	175.9	130.0	161.6	130.7	118.2
PTMET (h)	2537.2	2562.5	2512.0	2543.8	2510.4	2498.3
PTROPET (h)	189.6	214.8	164.4	196.1	162.8	150.6

, the maintenance points predicted for all cylinders using PTROPET are, on average, 35.3 h later (by approximately 24%) than those obtained using PTROPRG. The PTMRG (PTROPRG) is the time predicted using values adjusted for the operation condition based on the GE operating environment using the RGCCV, reflecting the current operating status of each cylinder. This is more practical and reliable than predicting the maintenance time using changes in the exhaust gas temperature of the cylinder, which is the criterion suggested by engine manufacturers for maintenance decisions [3,4].

The prediction performance of the maintenance time prediction algorithms based on the RGCCV and the exhaust gas temperature was compared, as shown in Figure 7, Figure 8, Figure 9, Figure 10, Figure 11 and Figure 12. The predicted maintenance times for each cylinder were graphed by setting 75% (25%) of the 21,887 data points of VOYAGE 21113 as training (test) data. Then, the evaluation metrics (MSE, RMSE, and MAE) were calculated.

Figure 7 shows the results for cylinder no. 6. Figure 7a shows the results of the PTMRG-based algorithm, where the predicted data follow the (collected) test data and are distributed close to the test data. By contrast, the PTMET results in Figure 7b show that the predicted data follow the test data but are widely and sporadically distributed relative to the PTMRG results, with large differences between the predicted and test data within certain maximum and minimum temperature ranges in the test data. The PTMRG- and PTMET-based algorithms have MAE values of 2.313 and 9.774, MSE values of 10.358 and 141.62, and RMSE values of 3.218 and 11.9, respectively. The prediction evaluation metrics for the PTMET are relatively poor, confirming that the PTMRG, which uses the RGCCV, has better prediction performance.

Depending on the engine specification and manufacturer, the cylinder exhaust gas temperature fluctuates widely between 250 and 500 °C during normal operation with load variations; temperature changes of approximately 10 °C or more may not be problematic, depending on the operating environment, and the scale of temperature data is large. In addition, due to temperature sensors, detection systems, and motor operation, abnormal data unrelated to motor abnormalities may be collected in real time. Therefore, the MAE, MSE, and RMSE, which were used here to evaluate the maintenance time prediction algorithms, are suitable for assessing such algorithms for ship engines’ exhaust gas temperatures.

Figure 8, Figure 9, Figure 10, Figure 11 and Figure 12 show the performance comparison results and the calculated MSE, RMSE, and MAE of the prediction algorithms for cylinder nos. 1–5. Although the cylinders have different operating temperature ranges, as shown in the results for cylinder no. 6, Figure 8a, Figure 9a, Figure 10a, Figure 11a and Figure 12a for cylinder nos. 1–5 shows that the prediction data of the PTMRG-based algorithm follow the (collected) test data and are distributed near the test data. Figure 8b, Figure 9b, Figure 10b, Figure 11b and Figure 12b shows the PTMET results; the prediction data follow the test data but are widely and sporadically distributed compared with the PTMRG results. The MAE, MSE, and RMSE values also confirm that the PTMRG (using the RGCCV) has better prediction performance than the PTMET, which is based on the exhaust gas temperature.

Table 8. Evaluation of PTMRG and PTMET predictions per cylinder.

Item\Cylinder	1	2	3	4	5	6
PTMRG
MSE	10.358	11.789	18.594	14.415	25.331	18.852
RMSE	3.218	3.433	4.312	3.797	5.033	4.342
MAE	2.313	2.529	3.237	2.639	3.54	3.235
PTMET
MSE	141.62	100.222	80.541	88.762	122.58	86.393
RMSE	11.9	10.011	8.974	9.421	11.072	9.295
MAE	9.774	8.31	7.124	7.382	8.707	7.529

is a summary of the prediction evaluation metrics for the PTMRG and PTMET for all cylinders. The average MAE values for the PTMET and PTMRG for cylinder nos. 1–6 are 8.138 and 2.916, respectively, indicating a 64% better MAE for the PTMRG than for the PTMET.

As shown above, predicting the maintenance time using an indicator that reflects engine conditions under varying operational environments (PRMRG) is more practical and reliable than relying only on changes in cylinder exhaust gas temperature (PTMET). In a broader context, system-level prediction techniques have been widely studied in other engineering domains as well. For instance, Panic et al. [26] proposed a hybrid FSO/RF communication model employing receive diversity under atmospheric turbulence, highlighting the effectiveness of analytical prediction approaches in dynamically changing environments.

Similarly, the approach presented in this study—using real-time engine data for predictive maintenance—shares the same underlying goal of improving system reliability in the presence of fluctuating operating conditions. Efficient predictive maintenance can be conducted by identifying engine operation abnormalities under normal operating conditions using a previously developed engine data preprocessing and anomalous symptom detection algorithm [4], combined with the proposed maintenance time prediction algorithm.

It should be noted that the engine model used in this study is a single model installed on a specific vessel. As such, the proposed algorithm is optimized for that particular engine configuration, and its applicability to other engine types remains an open question that requires further investigation. Nonetheless, the preprocessing framework and performance evaluation method based on engine operating conditions proposed in this study are expected to be applicable to a wide range of systems with further validation.

Future research will focus on acquiring additional engine models to evaluate the generalizability and applicability of the proposed approach across various marine power systems and include comparative studies with classical prediction models to further validate the robustness of the proposed approach.

4. Conclusions

In this study, algorithms for predicting maintenance times for ship engines were developed and tested using data from an operating ship generator for practical predictive maintenance. An indicator of anomalous GE symptoms (RGCCV) and the cylinder exhaust gas temperature were used as algorithm bases, and the prediction performances of both algorithms were analyzed and compared. The highlights of this study are as follows:

• A GE maintenance time prediction algorithm was developed by defining the PTMRG equation, which predicts the maintenance time via regression analysis based on the RGCCV, and defining the PTROPRG equation, which predicts the remaining operating time until maintenance.
• Another prediction algorithm was developed by defining the PTMET equation, which predicts the maintenance time via regression analysis based on the exhaust gas temperature of the GE’s cylinders, and defining the PTROPET equation, which predicts the remaining operating time until maintenance.
• The PTROPET-predicted maintenance times for all cylinders were, on average, 35.3 h later than those predicted using PTROPRG (by approximately 24%). PTROPRG (PTMRG) is the time predicted using values (RGCCV) adjusted for the status based on the engine operating environment, reflecting the operating status of each cylinder. This approach is more practical and reliable than predicting the maintenance time using changes in the exhaust gas temperature of the cylinder, which is presently the criterion suggested by engine manufacturers for maintenance decisions.
• The comparison of prediction performance between the PTMRG and PTMET showed that the MAE, MSE, and RMSE values for all cylinders indicated better prediction performance for the PTMRG (using the RGCCV) than for the PTMET (using the exhaust gas temperature). The average MAE values for the PTMET and PTMRG across all cylinders were 8.138 and 2.916, respectively, showing a 64% improvement in MAE for the PTMRG compared to the PTMET.
• Between the three metrics used to assess the maintenance time prediction algorithms, the MAE, which is less affected by outliers, is appropriate, considering the characteristics of engine data, such as wide normal operating temperature ranges, frequent fluctuations, and large temperature data ranges.

In conclusion, by defining the necessary equations for predicting the maintenance time of engines and developing prediction algorithms, we established a foundation for implementing effective predictive maintenance for ship GEs through comparison and verification using machine learning.

In future work, we plan to conduct comparative studies between the proposed maintenance time prediction algorithm and conventional time-based prediction methods. This will allow us to further validate the reliability and performance of our approach. Furthermore, we aim to develop an integrated predictive maintenance platform for ship engines, combining the abnormal symptom detection and maintenance time prediction algorithms proposed in this and previous studies. This platform will be applied to actual operating vessels and verified using real-time engine data, enabling us to evaluate its effectiveness and reliability in practical maritime environments. We expect that the methodologies and results derived from this study, along with our previous research, can contribute meaningfully to the practical implementation of predictive maintenance strategies—not only for marine engines but also for mechanical systems in various industrial domains.