Non-linear multi-way principal components analysis for process performance monitoring

Multi-way principal component analysis is increasingly being used for the monitoring of the performance of batch operations. In common with its principal components analysis counterpart, it is a linear technique and in this respect is not necessarily the most appropriate tool for handling industrial problems which exhibit significant non-linear behaviour. A non-linear principal component analysis methodology is proposed based upon an Input-Training neural network. Multivariate statistical process control charts based upon non-linear principal components are then used for process performance monitoring and non-parametric control limits defined. These overcome the limitations of the conventional approaches for action and warning limits which assume the data is normally distributed. Finally the methodology is evaluated on a benchmark multi-stage batch reactor simulation.