Comparison between Monte Carlo Dropout and Variational Inference Techniques for Bayesian Neural Network Models applied to Rotating Machinery Diagnostics

Resumo

Rotating components play a crucial role in mechanical systems and are present in several industrial
areas. These systems suffer from the adverse actions of loads and environmental condition. Thus, condition-based
maintenance of these components is a fundamental technique to synchronize and support maintenance schedules,
and machine learning algorithms are nowadays supporting these tasks. This paper aims to present a comparison
between Monte Carlo Dropout and Variational Inference techniques applied to Bayesian neural network models in
damage dignostics of ball bearings. The Bayesian Convolutional Neural Networks were tested, evaluated, validated
against the physical data, and their prediction performances were compared. Results showed that both models had
high performance in diagnosis. The comparison between the methods applied to Bayesian neural network models
showed similar results but each method present their own characteristics that could provides advantages in certain
specific situations.