The TriLS Approach for Drift-Aware Time-Series Prediction in IIoT Environment
Elena Uchiteleva, Serguei Primak, Marco Luccini, Ahmed Refaey, Abdallah Shami
Abstract
This article presents a novel drift-aware approach to multivariate time-series modeling in the nonstationary industrial Internet of Things environments. The three-layered three-state (TriLS) system enables cooperation between the gateway and the cloud toward the timely adjustment of a lightweight predictive model. Concept drift is detected by the cloud with the use of the extended adaptive windowing algorithm that operates on statistics of time sequences tracked by the gateway. This system is geared toward providing accurate predictions of nonstationary industrial processes for intelligent factory automation and safety. The proposed TriLS system is evaluated on records of recurring chemical processes collected at two plants and implemented on a Raspberry Pi board. TriLS achieves a lower prediction error than the reference adaptive schemes while reducing the computational effort and memory requirements for adaptation at the gateway by over 66% and 48%, respectively. It also reduces the volume of shared data between the gateway and the cloud by 40% –72% that is a significant cut on communications overhead.