Deep Learning Damage Forecasting for Rotating Devices Using Bayesian Sub-predictors

Predictive maintenance is a growing element of the Industry 4.0 strategy, and it involves detecting the possibility of failure as soon as possible. Very often, vibration data, appropriately reprocessed using such signal analysis methods as Fourier transform or wavelet transform, is used to detect drive failures. This paper presents the results of a study of a neural Deep Learning classifier system. The problem arising from the occurrence of noise in the learning and validation data is defined. It this method a use of Bayesian subpredictor was proposed to marginalize noise influence on the networks learning and validation performance. The designed classifier with Bayesian sub predictor achieved an accuracy of 96.08% on the test set. The proposed solution can be effectively applied to wind turbines, water turbines and many other rotating machines present in modern industry.