Agglomeration Detection in Gas-Phase Ethylene Polymerization Based on Multi-scale Convolutional Neural Network

Abstract

Fault will affect product quality, damage the reaction device, and cause property damage, so fault detection is a crucial part of the industrial production process. The traditional multivariate statistics method is gradually limited because the industrial data for inspection is mostly time series with the characteristics of difficult modeling and large noise interference. Multi-Scale Convolutional Neural Network (MCNN) has achieved remarkable results in time series processing and the computational efficiency. This paper applies MCNN to the fault detection and classification of the industry process. MCNN incorporates the feature extraction and the classification in a single framework. It will lead to further feature representations and superior fault detection performance at the industrial process. MCNN is conducted in the TensorFlow framework, and its fault detection performance is evaluated with existing BP neural network on a large amount of time series industrial data from a real gas-phase ethylene polymerization industry.