Heatmap-based temporal analysis for valve stiction detection in industrial control loops

Control valve stiction is a frequent nonlinear fault in process industries that causes stick-slip motion, inducing sustained oscillations in process variables. This negatively impacts plant performance, reducing product quality and lowering plant efficiency. Many existing techniques lack robustness across varying operating conditions, highlighting the need for both accurate and computationally fast solutions. In this work, we propose a model-free data-driven method for stiction detection by analyzing the dynamic interaction between the controller output (OP) and process variable (PV). The proposed method utilizes time-augmented signal representations to construct heatmaps that emphasize slow-varying and stagnation-prone regions in PV relative to OP. A physically interpretable PV/OP heat ratio is then extracted as a compact feature for threshold-based classification using ROC analysis. The classifier trained on a dataset of 58 industrial loops achieved an optimal threshold that, when evaluated on a test set of 20 loops, resulted in 95% accuracy. In addition, cross-validation results (78-81%) and sensitivity analysis confirm consistent performance, robustness, and reasonable generalization under different dataset partitions and parameter variations. Overall, the proposed framework provides a practical and interpretable solution for near real-time stiction detection in industrial control systems.