The use of batch processes is widespread across the manufacturing industries, dominating sectors such as pharmaceuticals, speciality chemicals and biochemicals. The main goal in batch production is to manufacture consistent, high quality batches with minimum rework or spoilage and also to achieve the optimum energy and feedstock usage. A common approach to monitoring a batch process to achieve this goal is to use a recipe-driven approach coupled with off-line laboratory analysis of the product. However, the large amount of data generated during batch manufacture mean that it is possible to monitor batch processes using a statistical model. Traditional multivariate statistical techniques such as principal component analysis and partial least squares were originally developed for use on continuous processes, which means they are less able to cope with the non-linear and dynamic behaviours inherent within a batch process without being adapted. Several approaches to dealing with batch behaviour in a multivariate framework have been proposed including multi-way principal component analysis. A more advanced approach designed to handle the typical characteristics of batch data is that of model-based principal component. It comprises of a mechanistic model combined with a multivariate statistical technique. More specifically, the technique uses a mechanistic model of the process to generate a set of residuals from the measured process variables. The theory being that the non-linear behaviour and the serial correlation in the process will be captured by the model, leaving a set of unstructured residuals to which principal component analysis (PCA) can be applied. This approach is benchmarked against the more standard approaches including multiway principal components analysis, batch observation level analysis. One limitation identified of the model-based approach is that if the mechanistic model of the process is of reduced complexity then the monitoring and fault detection abilities of the technique will be compromised. To address this issue, the model-based PCA technique has been extended to incorporate an additional error model which captures the differences between the mechanistic model and the process. This approach has been termed super model-based PCA (SMBPCA). A number of different error models are considered including partial least squares (linear, non-linear and dynamic), autoregressive with exogenous (ARX) variables model and dynamic canonical correlation analysis. Through the use of an exothermic batch reactor simulation, the SMBPCA approach has been investigated with respect to fault detection and capturing the non-linear and dynamic behaviour in the batch process. The robustness of the technique for application in an industrial situation is also discussed.
展开▼
