Developing ladder network for intelligent evaluation system: Case of remaining useful life prediction for centrifugal pumps

Abstract Intelligent evaluation system has been widely used in industries to estimate essential indexes which are unable to be measured directly through physical devices. Due to the complexity of labeling samples, common data-driven techniques such as supervised learning are developed on a small number of labeled data, while a large amount of unlabeled data is discarded. The amount of labeled information greatly limits the improvement of prediction accuracies. Furthermore, conventional evaluation approaches have only static structures, which makes the dynamic characteristics of parameters difficult to be presented. This paper proposes a ladder network (LN) based semi-supervised learning model to evaluate parameter dynamics, and a case of remaining useful life (RUL) prediction for centrifugal pumps is illustrated. LN datasets comprise a small part of labeled data and a large amount of unlabeled data. We exploited fluid-structure interaction (FSI) numerical simulation to replace actual monitoring, as well as built a RUL prediction model to annotate useful life for offline datasets. After that, the RUL was performed in the online stage by substituting real-time monitored variables into the network. The case study indicates that the LN-based intelligent evaluation system identifies the real-time RUL profile and achieves better predictive outcomes than supervised learning approaches.