MMoE-GAT: A Multi-Gate Mixture-of-Experts Boosted Graph Attention Network for Aircraft Engine Remaining Useful Life Prediction

Accurately estimating remaining useful life (RUL) is critical to reducing unplanned downtime, lowering maintenance costs, and improving safety and reliability in the field of prognostics and health management (PHM). At present, most of the data-driven RUL estimation methods are single-task learning models, i.e., the auxiliary tasks related to RUL are neglected, resulting in limited prediction accuracy. In this case, this study presents a multi-task learning (MTL) framework composed of a structure of multi-gate mixture-of-experts (MMoE) and a graph attention network (GAT) model, aiming to utilize the health state (HS) evaluation task to improve the prognostics accuracy. Specifically, GAT was employed to extract the intrinsic spatial information from the sensor network. A gating mechanism in the MMoE was utilized to adjust the parameters based on the distinctive features of different tasks. Moreover, we applied a learnable regularization term to deal with the fusion of HS loss and RUL loss. Experiments on the aircraft engine datasets reveal that the RUL prediction performances of MMoE-GAT are superior to those of available state-of-the-art (SOTA) methods.