On the implementation and performance assessment of an assumption-free methodology for batch process monitoring

Batch process monitoring is a challenging task due to the high variability of these type of processes. In order to ensure that the final product is of prescribed quality, statistical techniques have been developed for monitoring the process. Among these, multi-way principal component analysis is the most used. The standard methodology for monitoring batch processes requires that the data have the same number of samples. However, this is often not the case. Alignment methods for reaching this goal exist, but they are known for generating artifacts and being computationally demanding. Westad et al. (2015) proposed an assumption-free methodology that does not require any kind of data alignment. However, limited details were provided on the design and use of this methodology to perform process monitoring. Previous studies (Fracassetto, 2022; Sartori, 2023) have been done to understand how to exploit an assumptionfree model for process monitoring. In this thesis, further improvements on the topic have been carried out by providing an extensive set of guidelines on how to design the monitoring model. Furthermore, the assumptions made in the previous studies have been verified and a new methodology to build the control chart on the squared prediction error has been developed. In order to assess the monitoring performances of the assumption-free model, the obtained results have been compared to the ones reported by Sartori (2023) using a standard monitoring method on the same datasets. The comparison indicated that, on data which are already aligned, there is no clear evidence that a model performs better than the other. However, the assumption-free modelling outperformed the standard methodology on unaligned data in terms of both detection strength and detection speed.