Stiction detection in industrial control valves using Poincaré plot-based convolutional neural networks

Abstract

Valve stiction is one of the major causes of poorly performing industrial control loops. Stiction occurs when the static friction exceeds the dynamic friction during a direction change or when the stem is at rest. Recently, machine learning techniques were employed to detect the presence of stiction. These techniques required the use of multiple signals from the control loop in order to extract the key features to distinguish stiction cases from healthy or other malfunctions cases. In this paper, a new image-generating method, named the Poincaré plot, is proposed to feed the convolutional neural network (CNN) that only needs one signal from the control loop. The Poincaré plot is a powerful technique that can reveal the complexity of the process by evaluating the correlation within a single time series. The proposed Poincaré plot-based CNN showed satisfactory results in detecting stiction in real industrial applications as compared to other machine learning techniques present in the literature.