One-class classification techniques for improving phase I analysis in statistical process control

We propose to use a one-class classification technique to improve phase I analysis in statistical process control. Phase I analysis attempts to isolate the in-control data from the historical data, usually unlabelled to construct the reliable control charts. A traditional phase I method recursively removes the observations, which exceed the control limits until no out-of-control observations are detected. This recursive method requires a distributional assumption because the control limits are determined based on a certain parametric distribution. The proposed oneclassification technique does not require such distributional assumption. The effectiveness of the proposed approach is demonstrated through the simulated data set.