Process monitoring based on dissimilarity of time series data

For process monitoring, principal component analysis (PCA) has been widely used. Since PCA can describe the correlation among variables, PC-based monitoring outperforms traditional statistical process control methods, such as the Shewhart chart. Nevertheless, PC-based monitoring cannot detect changes in the correlation while the indices monitored are within their control limits. In the present work, a new monitoring method based on distributions of data is proposed, noting that distributions of data reflect the corresponding operational conditions. In order to quantitatively evaluate differences between two data sets, dissimilarity is defined and calculated by applying PCA to transformed-data matrices. The proposed monitoring method and the traditional PC-based method are compared using simulated data. The results of this study clearly indicate the advantage of the proposed method.