A convolutional neural network (CNN)‐based direct method to detect stiction in control valves

Abstract

AbstractControl valves are considered important capital assets in any process industry. A properly maintained control valve can have a significant impact on how well the process is controlled as well as the overall cost of the plant. However, control valves can suffer from poor control performance due to valve non‐linearities. One of the main reasons for non‐linearity is control valve stiction. Stiction not only causes oscillations in the process variables but also shortens the life of the control valve, resulting in an economic loss for the process. In a process plant, a control engineer generally analyzes the time series plot of process value (PV), set point (SP), and controller output (OP) data and identifies stiction based on the typical shape pattern of PV/SP/OP plot. In this study, the same shape pattern methodology is adapted to identify stiction using convolutional neural network (CNN) technique. A one‐dimensional convolution neural network (Conv1D) algorithm is developed, which works directly on PV/SP/OP time series data for stiction detection. The proposed CNN algorithm is tested on both simulated and industrial control loop data. The suggested method provides promising results with a combined stiction prediction accuracy of 92% (92.2% in predicting non‐sticky and 91.53% in predicting sticky loops) for the industrial loops data studied.