Using Bayesian Autoencoders for Health Indicator Construction in Remaining Useful Life Prediction on Ball Bearings

Kurzfassung

The useful life of electromechanical actuators (EMA) for the actuation of primary and secondary flight control surfaces in aircrafts is limited by mechanical failures of its components such as bearings. Trustworthy and reliable prediction of the Remaining Useful Life (RUL) of those components is therefore crucial for safety, maintenance and sustainability reasons. One challenge in the RUL prediction process is the construction of a trustworthy health indicator (HI) from the measured vibration data. Using deep autoencoders (AE) to construct the HI has been shown to be a promising approach. However, those approaches neglect the uncertainty in the data as well as in the neural network (NN) model. Therefore, this paper proposes using a Bayesian Adversarial Autoencoder (BAAE) to construct the HI as a probability distribution incorporating uncertainty rather than as a point estimation. The AE is formulated as a Bayesian Neural Network (BNN) to quantify the model and data uncertainty and is trained on bearing run-to-failure datasets for rotating bearings. Variational inference is used to approximate the posterior distribution of the BNN. The HI is constructed by taking the Mahalanobis distance between a healthy baseline and the current measurement. The proposed method is validated on the PRONOSTIA run-to-failure dataset. The constructed HIs show promising results regarding the suitability for RUL prediction and additionally the uncertainty of the constructed HIs can be quantified.