Adaptive Kalman Filter by Reinforcement Learning for Monitoring Aircraft Engines' Performance Against Abrupt Events

Abstract

Abstract. Engine performance's inverse problem is a well-known subject in the context of engine monitoring, particularly important for the aeronautics industry. In this framework, we aim to construct health/performance indicators (such as modular efficiencies and air mass flow rates) by leveraging operational data (i.e., sensors' measurements during flights) through the availability of a forward model (e.g., a thermodynamic simulator). An extensive literature is available on this topic, among which, Bayesian filtering—notably, Kalman filtering—is a dominant approach. However, even state-of-the-art methods still underperform in a scenario often found in practice: during its life, engine components not only degrade gradually over time due to wear but also can experience rare, abrupt changes in health states caused by uninformed maintenance or unknown external events such as Foreign Object Damages. In this work, we focus on this challenging scenario. We observe that Kalman filters (KF), when equipped with well-tuned a priori models, are capable of estimating the evolution of performance indicators due to degradations, but fail (if using the same a priori models) whenever an abrupt event occurs. To address this, we propose an adaptive filtering method, where parameters of the associated models are dynamically adjusted based on current estimates and observations. In particular, we propose a reinforcement learning (RL) agent, called single-filter reinforcement learning Kalman filter (RLKF), to control the noise covariance matrix of the transition function of a Kalman filter. Pushing one step further, we introduce a second agent, called double-filter RLKF, aided by launching alongside a nonadaptive filter predicting the moments of abrupt events. We conduct several experiments with simulated data of an in-house turbofan engine, and show the superiority of the adaptive filters with the proposed reinforcement learning agents.