A Novel Methodology for Process Monitoring by Using Deep Belief Network

Abstract

In recent years deep learning has been broadly used for process monitoring, fault diagnosis, pattern recognition, and image classification. In industrial systems there are many process systems that require ore advanced control technologies. Filling error control is an important but challenging problem in most production systems for a wide spectrum of liquid products. The major challenges behind this control problem include uncontrollable ambient factors, highly diversified products in modern manufacturing plants, limited budget for complex feedback control schemes, etc. To devise a low cost solution that is suitable massive replication, this paper proposes a data-driven approach for filling error control by using Just-In-Time Deep Belief Networks (JIT-DBN). The proposed method aims to construct a local DBN model based on historical data to suggest stopping time for the filling process by collectively considering the fluid viscosity, filling temperature and many other affecting factors. Based on the proposed method, a systematic framework for implementation is further devised. The proposed framework leverages the advantages of edge computing and cloud platforms to present a scalable solution with guaranteed computation efficiency and excellent adaptiveness to highly diversified products. In the validation experiments, both the proposed method and the proposed implementation framework are tested in the real-world filling production line for massive production. It is found that the proposed method can effectively reduce the mean filling errors and the filling uncertainties.