Nonlinear industrial process monitoring based on Deep Residual PCA-SVM model

Machine learning and deep learning methods for fault detection of industrial processes have been widely used. However, for the one-layer or "two-step" multivariate statistical methods and traditional machine learning industrial process feature extraction is shallow, deep learning related methods are uninterpretable and require a large number of samples. For the above problems, this paper proposed the DRPCA-SVM model, after multilayer PCA extracted residual contains feature information, which is rarely attention, multilayer residual PCA for linear data removal, the last layer residual contains strong non-linear characteristics, the residual information as the input of SVM, SVM has a good generalization ability, non-linear feature extraction, the interpretability of the overall model is also retained. The idea of "deep" is introduced to fully excavate the characteristic information and make the performance index of fault monitoring better. Finally, the validity of the proposed method is confirmed with the Tennessee Eastman (TE) process case.