Prediction of early failure of TOC analyzer usgin deep learning


Thanaphat Patravinij, Prabhas Chongstitvatana

5th National Conference of Science and Technology
15-16 January 2021

Abstract

Prediction of early failure of the total organic carbon analyzer (TOC) is important. The analyzer is used in the production of chlorine from brine.  TOC in brine is measured by the analyzer.  Too much TOC can cause damage and clog the production machine.  The ability to predict the early failure of the analyzer will reduce the loss from production.  An analyzer consists of many sensors.  There are 26 parameters reading from the analyzer.  All parameters are collected every 15 minutes.  The cycle will restart once the machine is stopped for maintenance.  This is called one cycle.  Remaining useful lifetime (RUL) can be calculated from all data from one cycle.  It is classified into three classes: Class 0, 1 and 2. This is useful to notify the user.  This work proposed using deep learning to learn RUL class from data of the analyzer collected from the real machines in used.  The result shows the prediction accuracy of 80.5%.