Aircraft maintenance is critical to an airline's operations to ensure the reliability, availability, and safety of their assets. Recently, the approach of using component prognostics in aircraft maintenance has received increasing attention in academic- and industrial research. Predictive maintenance has demonstrated promising results in using sensor-based prognostics for maintenance decisions. In this paper, we propose a novel predictive maintenance framework that is capable of mapping the individual component degradation levels to an optimal maintenance decision. The independent component degradation levels are computed by a supervised learning model, called "Long Short-Term Memory Networks". Subsequently, the computed degradation levels are utilized in a multi-component maintenance decision framework, by using a model-free reinforcement learning technique named "Deep Q-Learning". The predictive maintenance framework aims to minimize a cost objective based on the type and frequency of a maintenance action. In addition, we analyzed several key performance indicators, such as the number of components used, the component utilization level, as well as the wasted component lifetime. The predictive maintenance framework was evaluated using NASA's turbofan degradation dataset. Ultimately, the results of the numerical experiments showed that the proposed predictive maintenance framework resulted in lower costs than when using a time-based and corrective maintenance policy and competitive costs compared to an ideal maintenance policy. The proposed predictive maintenance framework opens new directions for multi-component sensor-based maintenance decisions. The results found form the basis for application suggestions and future research directions in practice.

Over the recent years a significant amount of research has been conducted to develop models which are able to estimate a components Remaining Useful Life (RUL) based on available sensor readings. In this research a deep learning (DL) model in combination with a similarity-based curve matching technique is used to estimate the RUL of a component. The data-driven RUL estimation scheme consist of an online and offline step. In the online step, a 1-dimensional Convolutional Auto-Encoder (1DConvNet-AE) is trained in an unsupervised way to convert multi-sensor readings, collected from historical run-to-failure instances, into a 1-dimensional reconstruction error vector. The set of generated 1-dimensional reconstruction error vectors is used to generate 1-dimensional Health Index (HI) curves which represent the various degradation paths of the run-to-failure instances. The HI curves based on historical run-to-failure instances are stored in the HI library. In the online step, multi-sensor readings of a test instance are converted into an online HI curve, representing the degradation path of an operational component. By using the similarity-based curve matching technique, the online generated HI curve is matched with the HI curves situated in the HI library. The RUL is estimated based on a weighted average of a set of matches which pass a set similarity threshold value. The proposed procedure is tested on an operational dataset. This research shows the existence of a relation between the increase of reconstruction error and health deterioration over time. The RUL estimation performance is considered in an operational evaluation procedure in which a similarity threshold value is included. This research shows a usable RUL can be estimated directly from raw multi-sensor input data by the proposed model.

The competition in the airline industry has rapidly increased during the last decades, especially with the entrance in the market of low-cost carriers. The high costs incurred in Maintenance, Repair and Overhaul (MRO) activities are generating a great interest in the improvement of maintenance operations as a way to stay competitive in the market. At the same time, the new generations of aircraft are increasingly being equipped with sensors that monitor the component's health condition. This stimulates the shift towards data-driven predictive aircraft maintenance, which is enabled by prognostics. This study proposes a model for maintenance scheduling of a fleet of aircraft based on component Remaining-Useful-Life prognostics and a limited stock of available spares. A discrete-time, rolling horizon approach is proposed, resulting in a sequence of scheduling time windows. For each time window, the goal is to find an optimal maintenance schedule. Moreover, the scheduling model considers three stages in decreasing order of maintenance priority, from critical aircraft leading to grounded condition, to predictive alerts, to non-critical failures. The results show that a cost-efficient maintenance schedule for a large fleet of aircraft is generated with an outstanding computational performance. Moreover, the aircraft operating costs are significantly reduced in the long-run, even when considering limited spares.