SUMMARY
Modern chemical plants consist of a large number of process units that have hundreds of control loops. As one of the most important elements of a control loop, control valves are essential assets to the plant because they ensure the high quality of products, as well as the safety of personnel and equipment. Maintaining their performance is usually very time- consuming but necessary because of the increasing environmental, societal and competitive demands. Unfortunately, control valves tend to suffer from the issue of stiction nonlinearity that results in oscillations in important process variables which are highly undesirable. However, current methods for the detection of control valve stiction are highly application-specific and have poor generalisation to other processes. In the present work, an automatic stiction detection framework has been developed to diagnose control loops based on the use of global recurrence plots and texture analysis. Texture features are extracted from distance matrices derived from control-loop variable data generated from a valve stiction model. A neural network model is then trained based on the extracted features. The optimised classification model can be readily applied in industrial control loops to identify the presence of stiction. The results from the simulated case study and 78 benchmark industrial loops show a 15% improvement over the best available method in the literature.