Enhanced Restricted Boltzmann Machine With Prognosability Regularization for Prognostics and Health Assessment

In the Internet-of-Things environment, it is critical to bridge the gap between business decision-making and real-time factory data to let companies transfer from condition-based maintenance service to predictive maintenance service. Condition monitoring systems have been widely applied to many industries to acquire operation and equipment related data, through which machine health state can be evaluated. One of the challenges of predicting future machine health lies in extracting the right features that are correlated well with the fault progression/degradation. We propose an enhanced restricted Boltzmann machine with a novel regularization term to automatically generate features that are suitable for remaining useful life prediction. The regularization term tries to maximize the trendability of the output features, which potentially better represent the degradation pattern of a system. The proposed method is benchmarked with regular restricted Boltzmann machine algorithm and principal component analysis. The generated features are used as input to a similarity-based method for life prediction. Run-to-failure datasets collected from two rotating systems are used for validation.