StictionGPT: Detecting valve stiction in process control loops using large vision language model

Stiction detection in control valves is a critical challenge in control loop performance assessment and fault diagnosis within the process industry. Existing stiction detection methods often require determining a threshold or rely on a large amount of image representation data to train deep neural networks. However, they face challenges such as difficulty in threshold determination, poor transferability, and lack of interpretability. Recent advancements in large language models (LLMs) and large vision-language models (LVLMs) offer new possibilities for improving the generalization of detection models by leveraging their multimodal understanding and reasoning capabilities. We propose StictionGPT, an LVLM-based agent for valve stiction detection. To overcome limitations of traditional methods, we leverage LVLMs to mimic human decision-making, combining textual semantics with visual shape features to determine the presence of stiction. First, we transform time-series data into images that contain shape features. These images are time-series plot, PV-OP plot, OP-$\Delta$PV plot, and CRD-PV plot. Next, we construct a multimodal dataset based on the semantics of these shapes for image–text alignment, and apply low-rank adaptation (LoRA) to foundation LVLMs to enable efficient few-shot generalization to the stiction detection task. Finally, we test the model on the ISDB benchmark and another real-world plant dataset. It turns out that StictionGPT achieves the highest accuracy on the ISDB benchmark and demonstrates excellent performance on the plant dataset.