Bayesian convolutional neural network models for uncertainty-aware bearing fault detection

Traditional classifiers are prone to overfitting and may exhibit spuriously high confidence, even when classifying low-quality data. Thus, it is important to have a reliable quantification of data and model uncertainty to make informed decisions about the system's condition. Probabilistic models, such as Bayesian, have been extensively researched to quantify uncertainty. However, recent developments have enabled us to successfully implement Bayesian in the backpropagation algorithm. By leveraging the properties of Bayesian and convolutional neural network, it is possible to classify directly from raw data and have a confidence band for each diagnosis. Here a Bayesian Convolutional Neural Network using back propagation has been implemented to classify common faults in bearings. The methodology has been evaluated using MAFAULDA open dataset which is composed of vibration signals from different experimentally simulated bearing conditions.