Intelligent Stiction Detection in Oscillatory Control Loops: A Machine Learning-Driven Approach

This paper introduces a machine learning technique aimed at evaluating the performance of control loops that exhibit oscillations, specifically focusing on the identification and quantification of stiction. The proposed method enhances the existing two-fold approach by integrating recent advancements in feature extraction and classification algorithms. Tailored for PID-based control loops with first order plus dead time (FOPDT) dynamics, this technique addresses the degradation of normal performance caused by stiction in the final control element (FCE). Utilizing support vector machines (SVM) and regression analysis, the method effectively differentiates stiction from other nonlinearities with varying behaviours and quantifies its severity. The classification model was evaluated using standard performance metrics, including precision, recall, and F1-score, and support. The results indicate a high overall classification accuracy of 92%, demonstrating the model’s effectiveness in distinguishing between stiction and non-stiction cases.