Remaining Useful Life Estimation for Anti-friction Bearing Prognosis Based on Envelope Spectrum and Variational Autoencoder

Anti-friction bearings (AFB) are crucial structural components conveying rotating motions in a variety of mechanical systems. To avoid unscheduled breakdowns and fatal failures, remaining useful life (RUL) prediction is of great practical significance in industrial practice for prognostics health management, e.g., optimizing maintenance plan for component replacements. Recently, the artificial intelligence (AI) advancements have provided effective data-driven models for bearing prognostics using machine learning. In this paper, using the variational auto-encoder (VAE) networks as the regression backbone, the bearing RUL is estimated using envelope spectra via measured vibrational data. First, the envelope spectra are utilized for bearing fault detection and the network input features. After the fault is detected, the VAE is used for learning the probabilistic mapping from the spectral input to the estimated RUL value, given its good probabilistic and generative properties over the classical auto-encoder (AE) in content generation and variational inference. The application of the method to the run-to-failure measured vibration data from the experimental rig available online have shown its efficacy in bearing RUL estimation.