Predictive Maintenance of Multi-Component Aircraft System using Convolutional Neural Networks and Deep Reinforcement Learning

Predictive maintenance is a new approach to replacing components based on the data-driven Remaining-Useful-Life (RUL) prognostics. However, implementing predictive maintenance remains challenging for aircraft. First, as aircraft maintenance requires high reliability, it is necessary to quantify the uncertainty of the predicted RUL. Moreover, the maintenance of multi-component systems should be planned considering the updated RUL distributions of individual components and complex cost models. This paper proposes an integrated method for the predictive maintenance of multi-component aircraft systems. We estimate the probability distribution of RUL using convolutional neural networks and Monte Carlo dropouts. Then, deep reinforcement learning (DRL) is applied to plan the replacement of multiple components based on individual RUL distributions. This method considers the uncertainty of RUL predictions, risk of component failure, time-varying maintenance costs, and maintenance slot costs. A case study on the predictive replacement of two turbofan engines illustrates the proposed method. By considering the probability distribution of RUL and grouping some replacements, the proposed DRL-based predictive maintenance provides lowered long-term maintenance cost.

Keywords: Aircraft maintenance, Predictive maintenance, Data-driven maintenance, Remaining-Useful-Life prognostics, Deep reinforcement learning.

Download PDF