Detection and Severity Identification of control valve stiction in industrial loops using integrated partially retrained CNN-PCA frameworks

The wear and tear of control valves is a common problem encountered on process plants, owing to continuous movements of the valves. The aging of control valves leads to operational problems, such as valve stiction. Detection and severity identification of valve stiction remains an extensive research area even today, since the behavior of these values tends to be nonlinear and is not necessarily easy to detect. Recent neural network based stiction detection methods published are only able to perform either stiction detection or quantification, which open up an area of research to propose a simplified algorithm to simultaneously detect and quantify stiction with high generalization capability. In this thesis, an integrated framework using such partially retrained convolutional neural networks in conjunction with principal component analysis (CNN-PCA) is proposed for simultaneous control valve stiction detection and automatic identification of the severity of the problem. In essence, features are extracted from segments of control valve signals or time series data accumulated via moving window and these features are subsequently used as a basis for monitoring of the behaviour of the valve with a PCA model in a standard multivariate process monitoring framework. The ability of the partially retrained CNN-PCA method to detect stiction and identify its severity with a smaller window size allows predictive monitoring to be performed 88% faster than the recently published SDN method. The detection results on 78 benchmark industrial loops show the ability of the proposed method to retain the generalization property and balance of false-positive and false-negative detections of the latest methods published in the literature, while having the key advantage of being readily extendible to the identification of the severity of stiction. Results based on simulated data with also show the promising capability of the proposed method to be used in online predictive monitoring for process plants which may be beneficial to alert the instrumentation/maintenance teams on the current and future health of their valves. In addition, results based on industrial data achieve 71% accuracy in stiction detection and 80% accuracy in severity identification.