Analyzing process monitoring data is very difficult in that the data usually consist of many variables correlated with each other. The traditional multiple regression approach is known to be inappropriate for analyzing this type of data.
This thesis considers partial least squares (PLS) method and artificial neural network (ANN) for analyzing the process monitoring data. PLS is known to be useful for constructing prediction equations as well as reducing the effects of multicollinearity. An ANN is an information-processing system inspired by human nervous systems, and has been widely used for describing nonlinear input-output relationships. However, there are no systematic procedures of determining ANN parameters (i.e., the number of hidden layers and neurons, learning rate and momentum). In this thesis, an experimental design approach is developed for finding optimal settings of ANN parameters.
The prediction abilities of PLS and ANN are compared in terms of the root mean of predicted squared error (RMPSE) for five sets of process monitoring data. Computational results indicate that PLS generally performs better than ANN. PLS shows especially good performance when there are many explanatory variables and relatively little observations.