Online Feature Selection Method Based on Change Rate for Process Monitoring

In most existing process monitoring methods, the features are selected according to only the amount of information they carry, but ignoring the information that is beneficial to monitoring carried by measurement variables with small variance, and hence some key information is lost and the features cannot be updated online. A change-rate principal component analysis (CR-PCA) is proposed to select the features that are most conducive to process monitoring. The data that change along the direction of the loading vector is reflected in the score matrix. The proposed method reduces information loss by considering the fluctuation of score vectors and takes into account historical fault information. A sliding window algorithm that adaptively changes the window width is introduced, which can realize the self-adjustment of features update speed and reduce the influence of random errors to a certain extent. The CR-PCA method includes the establishment of an offline model and the selection of an online feature. Finally, this research is applied to Tennessee Eastman (TE) process, which validates that with the update of features, the fault detection rate (FDR) has improved.