MOFSAut: Multi-task One-shot Feature Selection Autoencoder for Prognostics

Prognostic aims at predicting the lifetime at which a component or a system will be unable to perform a desired function. Datadriven models have been shown in recent literature to be reliable prognostic models. However, with the advancement of the Internet of Things, real-time monitoring of more features for components/systems under study has become easier. The availability of too many dimensions or features in the data acquisition stage has improved data-driven prognostic models, but it has also presented new challenges, including overfitting and lack of explainability. With a focus on feature selection, in this paper, we introduce the MOFSAut, the Multi-task One-shot Feature Selection Autoencoder, which is a deep learning-based model capable of generating lifetime predictions. The model selects from the original feature space the relevant features that are both globally representative and hold key information for lifetime predictions; this is achieved via multi-task learning. The one-shot training includes a feature selection task in the prognostic model training, eliminating the need for multi-step exploration for an adequate feature subset. This approach adds interpretability to the selected features which are shown to be important in predictability. The effectiveness of the approach is evaluated by applying it to a real-world case for the lifetime prediction of degrading turbofan engines.