ENBIS-18 in Nancy

2 – 25 September 2018; Ecoles des Mines, Nancy (France) Abstract submission: 20 December 2017 – 4 June 2018

Fault Detection for Batch Processes Using k Nearest Neighbours and Dynamic Time Warping

3 September 2018, 14:20 – 14:40

Abstract

Submitted by
Max Spooner
Authors
Max Spooner (Technical University of Denmark), Murat Kulahci (Technical University of Denmark)
Abstract
Online statistical process monitoring of batch processes is challenging due to the three-way structure of the data. Typically, J variables (pH, temperature, concentrations, etc.) are measured at K times points (e.g. every minute) throughout each of I batches. To monitor a new batch, one established approach is to first model the variation of batches from normal operating conditions (NOC) with multiway PCA. As the new batch progresses, it is compared to this model and an alarm is given if it deviates too greatly from the model. This is especially suited to processes where batch-to-batch variation of the variables at each time-point is approximately normally distributed, but less so if there is clustering of batches due to, e.g., changes in suppliers of raw materials. A new data-driven k-nearest neighbour method for online monitoring of batch processes is presented. This method uses the dynamic time warping (DTW) distance between an ongoing batch, and past NOC batches and signals an alarm if the distance becomes too great. The DTW distance has the advantage that it is not sensitive to minor differences in rates of progress between the ongoing and past NOC batches. The method is demonstrated using an extensive dataset of NOC and faulty batches from a simulated penicillin batch process, and shown to be flexible to local structures in NOC batches, such as clustering.

Return to programme