All recursivemodeling was performedmaintaining the temporal ordering of the samples to simulate the application of the models in on-line process monitoring. For validating the different calibration methods, the root mean square error (RMSE) for all the validation samples was calculated using always the prediction given by the “old” model, i.e. the model updated with all the samples preceding the current sample. Thus, the RMSE describes well the general prediction ability that can be achieved in real applications. However, since the main interest of the spectral analysis is to detect sudden major process changes as soon as possible, also the RMSE during the copper content increase related to the process failure in the validation data was used to compare the modeling performances. This error (RMSE2) was calculated as the root mean
square prediction error for the eight validation samples where the copper content was greater than 1.5% due to the process failure. These samples are referred in the following as the test period