Improved anomaly detection based on integrated multi-scale principal component analysis using wavelets: An application to high dimensional processes

Monitoring processes is becoming extremely crucial to maintain reliable and safe plant operation. Anomaly detection has proven to be easier through use of data driven methods that rely on historical data available from a process. Principal Component Analysis (PCA), a data-driven method, has been applied for detecting anomalies in industrial processes over last few decades. Since most processes are subjected to harsh environment, data collected is inherited with large amount of noise. Multi-scale filtering using wavelets has been well established for handling noisy data. In this paper, a new anomaly detection strategy which integrates multi-scale PCA(MS-PCA) and generalized likelihood ratio (GLR) test is proposed. To enhance MS-PCA efficacy, a novel method for computing optimum decomposition depth is proposed which is based on developing PCA model at each decomposition depth. The performance of the proposed strategy is demonstrated on high dimensional processes such as benchmark Tennessee Eastman process and an experimental quadruple tank process. The results shows that proposed strategy is having advantages in terms of better detection of faults and fewer false alarms.