Operating condition-invariant neural network-based prognostics methods applied on turbofan aircraft engines

Abstract

Neural networks in their many flavors have been widely used in prognostics of engineered systems due to their versatility and increasing potential, especially with recent breakthroughs in Deep Learning and specialized architectures. Despite these advances, some problems can still significantly benefit from a solid exploratory analysis and simple taskspecific data/target transformations. In this work, popular architectures including Feedforward, Convolutional and LSTM (Long Short-Term Memory) networks are evaluated in a case study of RUL (Remaining Useful Life) prediction for turbofan aircraft engines, using data from publicly available repositories. A robust set of over 20; 000 model configurations are tested, evaluating the effects of several hyper-parameters and design choices. The latter includes a maximum prediction horizon, revealing a trade-off between prediction accuracy and timeliness which can have significant impact in realworld applications. An operating condition-specific standardization scheme is also evaluated, in order to minimize the impact of normal changes in operating regimes which obfuscate the fault degradation patterns. A comparison with existing works in literature shows some simple policies for operating condition-invariance have lead to results which outperform the current state-of-the-art methods for some of the data subsets with multiple operating conditions.