Probabilistic Gear Fault Diagnosis Using Bayesian Convolutional Neural Network

Vibration measurement-based gear fault diagnoses have shown the promise aspects, where the deep learning methods have been harnessed. However, the traditional deep learning methods are deterministic in nature, and will be prone to false prediction when uncertainties are involved, such as time varying condition and measurement noise. To address these challenges, the fault pattern recognition needs to be performed in a probabilistic manner. Considering the features in vibration time-series usually are massive, in this research we develop a Bayesian convolutional neural network (BCNN) to conduct the gear fault diagnosis under uncertainties. The predictive distribution yielded facilitates the decision making with confidence level, leading to the robustness enhancement of the fault diagnosis. Comprehensive case studies are carried out to validate the proposed methodology.