Abstract
Objective: This work aims to use slow features extraction of time-varying signals to solve the unfavorable influences of traditional principal component analysis (PCA) method on feature extraction in Tennessee Eastman (TE) process.
Methods: Slow feature principal component analysis (SFPCA) method can obtain the slow features information of the observed data while considering variance maximization. The monitoring statistical indices are built on SFPCA method, and their confidence limits are computed by kernel density estimation (KDE), respectively.
Results: All the monitoring results of SFPCA are presented. The confidence limit for fault detection is set to 95%. The fault exists all the time from 161st sample by SFPCA method. Stochastic occurrence appears with relatively smaller amplitude in temperature of reactant feeding in fault 10. Monitoring chart based on SFPCA performs better with fault detection rate for T2 index reaching 93.13% and Q index 56.50%. In Table 2, the proposed method can detect most faults than PCA, especially for faults (4), (5), (8), (10), (11), (16), (17), (18), (19), (20), (21). In Table 3, for fault (2), (8), (10), (11), (13), (16), (17), (19), (21), SFPCA shows better detection performance than PCA. In fault 5, the positive step change in condenser cooling water temperature leads to a sharp increase in its flow rate which is measured by the 52nd variable.
Conclusion: SFPCA method demonstrates better performance than the traditional PCA method from the perspective of both fault diagnosis rate and fault diagnosis time in TE process.
Keywords: Fault diagnosis, kernel density estimation, principal component analysis, slow feature, tennessee eastman process, time-varying.
Graphical Abstract
28