Towards Remaining Useful Life Prediction in Rotating Machine Fault Prognosis: An Exponential Degradation Model

Estimating the Remaining Useful Life (RUL) of critical assets is not only a basic requirement of condition-based maintenance, but is also central to system prognostics for cost efficiency. A well professed definition of prognostics in the existing literature is the ability to use automated methods to assess system condition, estimate functional parameters and forecast degradation. Rotating shafts are a critical component to most modern day machinery and are at a constant risk of failure given the harsh working environment they are subjected to. The main aim of this paper is to propose a data-driven prognostic approach combining a machine learning method like Principal Component Analysis (PCA) with an exponential degradation model to accurately predict the RUL of a rotating shaft. For this purpose, vibration data collected off a faulty shaft over many days is analyzed in both time and frequency domains to extract descriptive fault features. Following feature post-processing for noise reduction and data training, it is observed that Kurtosis ranks the highest in terms of feature importance by quantifying its merit amongst all other features based on the metrics of monotonicity and trendability. Following feature normalization, a PCA model is employed for dimensionality reduction and feature fusion to improve the accuracy of the prognosis system. As a good indicator of deteriorating health, the PCA-based fused health indicator is combined with the previous top feature, Kurtosis, to be used as a mathematical input for a physical-behavior degradation model. Unlike most practical cases, the selection of threshold for a degradation slope in the proposed model is independent of historical data and is capable of evaluating the significance of slope by relying on observed data instead. Results indicate that parameter distribution is updated on a real-time basis by selecting an arbitrary slope parameter every time a significant variance in health is detected. The final output includes probability density function (PDF) of RUL, Estimated & True RUL, confidence intervals and prognostic performance analysis plots indicating better performance of the proposed degradation model in predicting shaft failure.