Abstract: With the rapid development of data collection and processing technologies, particularly in the field of machine/deep learning, significant progress has been made in manufacturing system failure diagnosis in recent years. However, there is currently limited research that integrates these failure diagnoses of the system into subsequent maintenance decision-making. On the other hand, existing maintenance optimization models often assume known system failure processes, which is not realistic in practical scenarios. In this study, we propose a data-driven maintenance planning approach for a single-unit system. Based on the system's condition information, we first employ a Long Short-Term Memory(LSTM) network to predict the probability of the system's failure in the next time period. Subsequently, we apply the estimated probabilities to a modified maintenance optimization model adapted for data-driven scenarios to determine the threshold at which preventive maintenance activities should be initiated when the system failure probability reaches a certain level. We validate the effectiveness of the proposed data-driven maintenance decision-making method using two turbo engine datasets. Compared to the classical periodic maintenance decision-making approach, this novel method significantly reduces the average maintenance costs of the system.

Key words: data-driven, deep learning, failure diagnosis, condition-based maintenance, turbo engine