Remaining Useful Life (RUL) estimation is directly related with the application of predictive maintenance. When RUL estimation is performed via data-driven methods and Artificial Intelligence algorithms, explainability and interpretability of the model are necessary for trusted predictions. This is especially important when predictive maintenance is applied to gas turbines or aeroengines, as they have high operational and maintenance costs, while their safety standards are strict and highly regulated. The objective of this work is to study the explainability of a Deep Neural Network (DNN) RUL prediction model. An open-source database is used, which is composed by computed measurements through a thermodynamic model for a given turbofan engine, considering non-linear degradation and data points for every second of a full flight cycle. First, the necessary data pre-processing is performed, and a DNN is used for the regression model. The selection of its hyper-parameters is done using random search and Bayesian optimisation. Tests considering the feature selection and the requirements of additional virtual sensors are discussed. The generalisability of the model is performed, showing that the type of faults as well as the dominant degradation has an important effect on the overall accuracy of the model. The explainability and interpretability aspects are studied, following the Local Interpretable Model-agnostic Explanations (LIME) method. The outcomes are showing that for simple data sets, the model can better understand physics, and LIME can give a good explanation. However, as the complexity of the data increases, both the accuracy of the model drops but also LIME seems to have difficulties in giving satisfactory explanations.
Estimating the remaining useful life (RUL) of an asset lies at the heart of prognostics and health management (PHM) of many operations-critical industries such as aviation. Mod- ern methods of RUL estimation adopt techniques from deep learning (DL). However, most of these contemporary tech- niques deliver only single-point estimates for the RUL without reporting on the confidence of the prediction. This practice usually provides overly confident predictions that can have severe consequences in operational disruptions or even safety. To address this issue, we propose a technique for uncertainty quantification (UQ) based on Bayesian deep learning (BDL). The hyperparameters of the framework are tuned using a novel bi-objective Bayesian optimization method with objectives the predictive performance and predictive uncertainty. The method also integrates the data pre-processing steps into the hyperparameter optimization (HPO) stage, models the RUL as a Weibull distribution, and returns the survival curves of the monitored assets to allow informed decision-making. We vali- date this method on the widely used C-MAPSS dataset against a single-objective HPO baseline that aggregates the two ob- jectives through the harmonic mean (HM). We demonstrate the existence of trade-offs between the predictive performance and the predictive uncertainty and observe that the bi-objective HPO returns a larger number of hyperparameter configurations compared to the single-objective baseline. Furthermore, we see that with the proposed approach, it is possible to configure models for RUL estimation that exhibit better or comparable performance to the single-objective baseline when validated on the test sets.
