Machine learning for wavelet-based valve stiction detection

Abstract

Valve stiction presents challenges in industrial process control, leading to oscillations and hindering the regulation of fluid flow. This thesis addresses the detection of valve stiction by exploring the limitations of current methods and proposes a novel approach that combines wavelet reconstruction and convolutional neural networks (CNN) to enhance stiction detection performance. The proposed method utilizes preprocessed process variable versus controller output (PV(OP)) plots as input to the CNN model, capitalizing on the distinctive characteristics of stiction. Training and evaluation employ both simulated and real-world data from the International Stiction Data Base (ISDB), with the F1 score serving as the primary performance metric. Results demonstrate that the modified PV(OP) input approach, in conjunction with the CNN classifier, achieves an impressive F1 score of 0.9, surpassing conventional methods. In-depth analysis of fault cases provides valuable insights into the strengths and limitations of the approach, emphasizing interpretability and robustness. The accurate detection of valve stiction enables proactive maintenance and targeted interventions, ultimately improving the performance of interconnected systems and enhancing safety and efficiency in industrial processes.