A lognormal-normal mixture model for unsupervised health indicator construction and its application into gear remaining useful life prediction

Accurately predicting the remaining useful life (RUL) of a key component, such as gear, is significant for guaranteeing the safe operation of mechanical equipment and making a proper maintenance plan. The health indicator (HI) plays an essential role in the data-driven RUL prediction technique. HI can be constructed from the perspective of the data distribution discrepancy. However, some existing methods cannot utilize different types of distributions to estimate the distribution discrepancy in various domains. In addition, the constructed HI may not comprehensively describe the tendency of performance degradation by using a type of distribution to obtain the distribution discrepancy in a domain. To overcome these challenging problems, a novel lognormal-normal mixture model (LNMM) that utilizes lognormal and normal distributions is constructed to estimate data distributions from two data domains, including the raw data domain and exponentially transformed data domain. Then, the distribution contact ratio metric (DCRM) is applied to calculate the discrepancies between benchmark distribution of healthy data and distributions of whole life-cycle data in two domains. The gear HI is generated without supervision by combining the DCRMs of two domains. The developed unsupervised HI is employed to estimate gear's RUL via an improved multi-hierarchical long-term memory augmented network. Finally, the experimental results indicate the feasibility and merit of the developed LNMM in gear HI construction. The LNMM-based HI has a better predictive efficacy than the conventional and state-of-the-art unsupervised HIs.