Development of a Deep Learning-based Uncertainty-aware Predictive Maintenance Method

Abstract

Research on predictive maintenance strategies for industrial electronics has improved from corrective and preventive strategies, which require a large amount of time and cost, to data-driven intelligent systems. In particular, data-driven methods monitor the real-time condition and health status of industrial electronics to facilitate efficient maintenance. Based on the real data collected from electronics, machine learning methods have been widely employed to analyze the health status of industrial electronics in advance of the unpredicted shutdown. Recently, the use of deep learning-based methods that have higher expressive power and adaptability is increasing in order to yield more accurate and stable prognostic results. However, there still exist some problems in the current form of predictive maintenance strategies. First, prognostic results are mostly point estimates, which lack not only predictive uncertainties but also interpretabilities for domain experts on site. Furthermore, some existing uncertainty-aware prediction methods are not scalable to large-scale multivariate data collected from industrial electronics. In this research, a Bayesian learning-based prognostic method is proposed for predictive maintenance of industrial electronics. The proposed method provides uncertainty-aware prediction results that are scalable to large-scale datasets. The effectiveness of the proposed prognostic method is validated using multiple datasets as well as is compared with existing deterministic and probabilistic methods.