Data-driven Prognostics of Alternating Current Solenoid Valves

Abstract

Solenoid valves are critical components in many process control systems, as their failure is often a root cause for plant shutdowns. Therefore, the ability to predict the remaining useful life (RUL) of solenoid valves is highly desirable. In this paper, a novel data-driven RUL prediction methodology for solenoid valves is proposed, by training deep neural networks on images constructed from raw current signatures. The performance is compared to shallow machine learning algorithms, trained on features which are obtained using domain expertise on the valves. We show that an ensemble of CNN subnetworks (constructed by the AdaNet algorithm) achieves a predictive performance comparable to the feature -based approaches (p = 0.584 for linear regression, p = 0.321 for Gradient Boosted Regression Trees). By focusing on reducing the modeling effort required for constructing features and fixing neural network architectures, this study offers a new and promising approach for solenoid valve prognostics, and thus even more complex systems in the future.