Unsupervised machinery prognostics approach based on wavelet packet decomposition and variational autoencoder

The prognosis of rotating machinery has been very prominent in recent years thanks to the advances in digital signal processing and intelligent systems. Unsupervised machine learning methods have been adopted along with signal processing techniques in both time and frequency domain to build indicators that describe the degradation of mechanical systems. This paper proposes a novel method for generating a degradation indicator to estimate the remaining useful life of rotating machinery critical components, based on a beta variational autoencoder neural network that processes statistical distributions in a feature hyperspace whose coordinates combine time-domain analysis and wavelet packet decomposition features from vibration signals. Indicators are obtained using bearing data from a publicly available dataset, aiming to enhance the observability of monotonic trends, and are used to assess different hyperparameter configurations of the proposed methodology. Based on the comparison with recently published results on the same dataset, the proposed method produced more robust indicators capable of detecting early changes in degradation processes, determining more accurate RUL estimates.