Sensor Data Prediction in Process Industry by Capturing Mixed Length of Time Dependencies

Abstract

Sensor Data prediction has been an interesting and practical topic in many domains. In the process industry, sensor data prediction can help us detect, diagnose and even predict possible failures to reduce unnecessary losses. Due to the complex relationship among multiple sensors, it is challenging to accurately predict the time series of multivariate sensors. In this research, we aim to solve the problem of predicting the time series of several related sensor data and proposed a novel structure for addressing with this provocative problem. More specifically, several proposed mixed length dilation layers and recurrent cells are used to capture mixed length of time dependencies. Experiments demonstrate that our proposed model indicates competitiveness in predicting comparing with other baseline methods.