International Conference on AI and Big Data in Engineering Applications, 15 June 2021, pp.227-239
Bearings are widely used in the machinery industry. In particular, they undertake important functions at many critical points. Any failure in these parts causes the machine to stop and leads to unexpected production losses. These unplanned downtimes increase the costs. Therefore, the goal of predictive maintenance is to estimate the remaining useful life of critical machine parts, such as bearings. The main purpose of this study is to determine the machine health indicator from vibration data using machine learning methodology and then use the computed health indicator in the remaining useful life estimation of bearings. The steps required to compute the machine health indicator are listed as feature extraction, feature postprocessing, dimension reduction, and feature fusion. The vibration signal is collected in the time-domain and is transformed to the frequency-domain using the Welch method. The traditional statistical indicators such as mean value, standard deviation, skewness, kurtosis, etc. are utilized to extract features in both domains. Monotonicity analysis is employed to select the most useful features, and they are fused into a single health indicator by the principal component analysis. An exponential degradation model for estimating remaining useful life is developed using the forecasting machine learning method. The proposed methodology is evaluated through a case study involving vibration measurements generated using a ball bearing signal simulator.