Neural networks for on-line parameter change detections in time series models

Time series models can be derived directly from a manufacturing process, assuming that variables are observed at equal time intervals. A theoretical autocorrelation function for the identified model is developed based on the estimated parameters. The Monte Carlo simulation technique is used to generate a synthetic series which has a similar autocorrelation function to the theoretical one. A feedforward neural network is trained to recognize the patterns exhibited in a sample autocorrelation function. An on-line sequential window obtained from a continuous process is propagated into the network to detect parameter changes for a given manufacturing process.