Non-linear principal components analysis with application to process fault detection

Principal component analysis has been used for the development of process performance monitoring schemes for both continuous and batch industrial processes. However, it is a linear technique and in this respect it is not necessarily the most appropriate methodology for handling industrial problems which exhibit nonlinear behaviour. A nonlinear principal component analysis methodology based upon the input-training neural network is proposed for the development of nonlinear process performance monitoring schemes. Kernel density estimation is then used to define the action and warning limits, and a differential contribution plot is derived which is capable of identifying the potential source of process faults in nonlinear situations. Finally, the methodology is evaluated through the development of a process performance monitoring scheme for an industrial fluidized bed reactor.