Anomaly detection using multi-scale dynamic principal component analysis for Tenneesse Eastman Process

Effective and accurate detection of anomalies is crucial to ensure safe and reliable operation in complex chemical processes. Principal Component Analysis (PCA) technique has been a popular choice for anomaly detection vowing to its capability in handling complex multiple variables. However, data extracted from chemical process plants is correlated in time due to the process dynamics, hereby making the standard PCA approach ineffective in detecting of anomalies. Also, with most industrial process data embedded with heavy noise content, the detection ability of the standard PCA strategy is greatly effected. The dynamics present in the industrial data can be incorporated using dynamic PCA modeling technique where existing PCA model is embedded with lagged variales which would aid in capturing plant dynamics. Further, the effect of measurement noise can be reduced by using wavelet-based multi-scale representation of data which extracts useful information from the data in time and frequency domain simultaneously. In the current paper, a multi-scale dynamic PCA (MSDPCA) is used as the modeling frame-work while T-2 and Q statistics are used as fault detection indicators. The efficacy of the developed MSDPCA strategy is demonstrated on a bench-mark Tenneesse Eastman(TE) industrial process. The simulation results clearly shows that the MSDPCA strategy is able to detect anomalies effectively in the multi-variate dynamic TE process.