Deep learning-based intelligent multilevel predictive maintenance framework considering comprehensive cost

Due to the increase in the series-parallel multi-state system (MSS) complexity caused by the nonlinear change of parameters, the traditional model-based maintenance methods are becoming less effective and obsolete. This study proposes a novel deep learning-based intelligent multilevel predictive maintenance (MPM) framework for series-parallel MSS considering comprehensive cost. A new adaptive convolution-transformer (C-Transformer) was constructed to predict component remaining useful life (RUL) through extracting features adaptively. Based on this, the component failure probability was obtained through convolutional neural network (CNN). Then, to directly reflect the operating conditions of MSSs, multilevel maintenance was customized with multilevel failure through the trial-and-error learning method. During the intermission breaks, an intelligent dynamic decision-making optimization model was proposed by introducing multilevel maintenance to improve the system's state in a future mission, which was solved by a new artificial bee colony algorithm (called MDU-ABC-K) to minimize the comprehensive cost under economic dependence and critical component constraints, thus simultaneously balancing maintenance time and cost. The proposed approach was compared with other models through turbofan engine data set by NASA. The comparison results indicate that the proposed intelligent MPM framework can offer a more reasonable and superior maintenance strategy.