선박 디젤엔진의 인공지능 기반 대기오염물질 배출량 예측 및 시뮬레이션을 활용한 잔여 수명 평가

다국어 초록 (Multilingual Abstract)

Marine diesel engines, widely used as the primary power source for ship propulsion, are considered a major contributor to air pollution. Moreover, the deterioration of engine components affects combustion conditions, exhaust emissions, and operational reliability. With recent tightening of emission regulations, concerns about the reliability of conventional emission estimation methods have increased. Therefore, a data-driven framework that predicts pollutant emissions and assesses engine health conditions is required to improve emission management and maintenance efficiency. This study proposes an artificial intelligence‐driven methodology for predicting air pollutant emissions and evaluating the engine degradation and remaining useful life（RUL） of marine diesel engines. Pollutant emissions were measured using a portable emission measurement system to develop emission prediction models for carbon dioxide（CO2）, carbon monoxide（CO）, nitrogen oxides（NOX）, and total hydrocarbons（THC）. After evaluating several machine learning and deep learning models, a stacking meta model combining a
convolutional neural network and XGBoost was proposed to improve prediction accuracy. In addition, simulation-based failure scenarios representing the degradation of major engine components were generated using AVL CRUISETM M. A health index for estimating engine degradation and RUL was subsequently derived using principal component analysis and Mahalanobis distance. The proposed stacking meta model predicted CO2 and NOX emissions with high accuracy under various operating conditions, with R2-adj values exceeding 0.95. The proposed model captured transient and nonlinear emission behaviors more effectively than a conventional emission estimation method, resulting in more than a 70% reduction in estimation errors for CO2 and THC emissions and an approximately 50% reduction in estimation errors for CO and NOX emissions. The simulation‐based degradation assessment successfully identified different fault severities and provided a quantitative relation between the health index and degradation level. Based on the estimated degradation trend, the proposed framework effectively predicts the time required to reach predefined fault thresholds, thereby supporting RUL estimation. The proposed methodology can contribute to the implementation of condition-based maintenance and intelligent engine management systems for smart ships.