Online monitoring of multi-phase batch processes using phase-based multivariate statistical process control

Batch and semi-batch modes of production are common in a number of high value-added industries including specialty chemicals, pharmaceuticals, and biologics. Online monitoring of such processes seeks to detect run-to-run deviations, anomalies or faulty conditions by analyzing real-time measurements so that they can be corrected quickly and the batch recovered. Multivariate statistical methods have been widely used for batch process monitoring; however, most of the existing techniques can be used only at the completion of the batch, i.e., they are not suitable for online use. A unique characteristic of batch processes is that the correlation structure between the process variables can change as a function of time – different phases can have different correlation structures. Further, the length of each phase and the time at which the phase (correlation structure) changes can also vary from one run to another. We account for both these issues in the proposed online monitoring method. First, we propose a method for detecting the occurrence of phase changes. Landmarks, called singular points, in real-time measurement are used as indicators of phase change. Their occurrence is used to identify the current phase in an ongoing batch. We use dynamic time warping to account for changes in the length of a phase. Singular points and dynamic time warping together thus provide the mapping in real-time between the current run and a reference golden batch. This mapping is used with phase-specific multivariate monitoring models to monitor the process and detect abnormalities in real-time. The application to a well-known fermentation process simulation reveals that the proposed method significantly improves performance and sensitivity-both false positives and false negatives are reduced – which justifies the extra effort in developing phase-specific monitoring models.