(200c) Neural Network Based Noise-Suppressor and Predictor for Quantifying Valve Stiction in Oscillatory Control Loops

Valve stiction-induced oscillations in chemical processing systems adversely affects control loop performance and can degrade the quality of end products. Estimating the degree of stiction in a valve is a crucial step in compensating for the effect. This work proposes a neural network approach to quantify the degree of stiction in a valve once the phenomenon has been detected. Several degrees of stiction are simulated in a closed loop control system by specifying the magnitude of static (fs) and dynamic (fd) friction in a physical valve model. Each simulation generates  controller output OP(t) and process variable PV(t) time series data. A feed-forward neural network (the predictor) is trained to model the relationship between a given OP and PV pattern, and the stiction parameters. To test the models predictive capability, a separate set of stiction patterns are generated with and without added process noise. An inverse neural network-based nonlinear principal component analysis (INLPCA) noise-suppressor effectively extracts the underlying stiction behaviour from the noise-corrupted OP(t) and PV(t) stiction patterns. In the noiseless test patterns, the predictor is shown to estimate fs and fd with a 0.65% average error. In the case of the noisy test patterns, the average error achieved was 1.85%. Since the predictor is developed offline, the use of computationally intensive real-time search/optimization routines to quantify stiction is avoided. The neural networks proved to be easily implementable, highly flexible models for extracting stiction behavior from control loops and accurately quantifying stiction, as long as an adequate first-principles description of the process dynamics can be developed.