Control Valve Stiction Detection using Learning Vector Quantization Neural Network
Seshu Kumar Damarla, Biao Huang
Abstract
The performance of a process control loop can be limited when nonlinear problems like deadband, hysteresis, backlash, stiction, etc. exist in control valve. Stiction occurs more frequently than the other valve problems and has potential to cause adverse oscillations in the control loop, resulting in poor quality products, excessive use of raw materials and energy, and an environmental footprint. Timely detection of sticky control valves can help control engineers to take appropriate actions (retuning the controller or using stiction compensation methods) to prevent further degradation of the performance of the control loop. In connection with the aforesaid fact, this work proposes a novel stiction detection method founded on learning vector quantization neural network (LVQNN). Simulated database is generated and used to train the LVQNN with the training algorithm: LVQ2.1. To further enhance the performance of the method, transfer learning is adopted to retrain the pre-trained LVQNN model by using industrial data. The retrained LVQNN is tested on practical data obtained from a wide variety of industries. Results highlight that the proposed method can outperform the existing methods.