A Novel Toeplitz Matrix and CNN-LSTM Based Method for Identifying Control Valve Stiction

Control valve stiction is a prevalent issue in industrial process control, often leading to oscillations that degrade system performance, reduce product quality, and increase operational costs. Existing stiction detection methods, particularly machine learning (ML)-based approaches, often fail to generalize effectively to real-world industrial data due to their reliance on simulated data sets that lack real-world complexities. To address this challenge, this study proposes a novel hybrid deep learning framework that integrates Toeplitz matrix-based image encoding with a CNN-LSTM architecture for accurate and computationally efficient stiction detection. In the proposed method, time-series control loop data, comprising process variable (PV) and controller output (OP) signals, are transformed into structured images using Toeplitz matrices, preserving temporal dependencies while enabling efficient feature extraction. A hybrid convolutional neural network long short-term memory (CNN-LSTM) model is then employed, leveraging CNN’s spatial feature extraction and LSTM’s sequential pattern recognition capabilities. To enhance generalization, a transfer learning strategy is applied by fine-tuning the model with industrial data sets from the International Stiction Database (ISDB). An accuracy of 90.47% underscores the model’s strong predictive performance and reliable classification ability, while its streamlined architecture effectively reduces the complexity and computational demands of earlier approaches.