Quality prediction of industrial polyethylene process with robust long short term memory model

Abstract

A challenge for construction of accurate soft sensors in the process industries is that industrial process data often contains various noise and outliers. A robust long short term memory (LSTM) neural network with the maximum correntropy criterion (MCC) is proposed to build a reliable soft sensor model. The proposed model employs an MCC-based objective function centred on a Gaussian kernel. Without tedious preprocessing approaches for process data, the proposed model can assign smaller weights to outliers to reduce their negative effects on prediction. Consequently, it can achieve better prediction performance compared with the traditional LSTM soft sensor without robust strategy. The quality prediction results on an industrial polyethylene process demonstrate its effectiveness and advantages.