Di Tian

As shale gas wells enter the middle and late stages of development, abnormal wellbore conditions such as liquid loading, tubing blockage, and tubing perforation occur with increasing frequency, severely constraining production capacity. Accurate and timely diagnosis of daily production monitoring data is therefore essential to ensure the efficient development of shale gas resources. Conventional diagnostic methods for abnormal wellbore conditions are limited by their strong dependence on labeled data, inadequate capability for multimodal information integration, and the continued requirement for manual post-interpretation of results. To address these challenges, an intelligent diagnostic approach based on a vision–language model (VLM) is proposed. In this method, historical production time-series data are normalized, segmented by sliding windows, and encoded into two-dimensional image representations, which are then combined with text-based expert knowledge prompts to construct structured training samples. The VLM is fine-tuned using task-specific instructions and structured training data, enabling anomaly type classification, start–end time identification, diagnostic interpretation, and drainage strategy recommendation. In addition, structured output templates are introduced to ensure that the generated results are not only accurate but also consistent and interpretable. The proposed method was validated on 500 expert-annotated shale gas well cases and compared with traditional machine learning methods and time-series diagnostic approaches based on large language models. The results demonstrated that, under limited labeled data conditions, the VLM maintained high diagnostic accuracy and, in particular, significantly outperformed existing methods in reducing false alarm rates. The model was further shown to automatically generate operator-oriented textual explanations of abnormal conditions, thereby reducing the need for manual interpretation and enhancing both human–machine collaboration and real-time performance in the diagnostic process. Additional analysis revealed that the VLM exhibits clear advantages in handling long-term production monitoring data from shale gas wells: by preserving both local fluctuations and global trends in the time series through image representations, the model can capture short-term anomalies as well as long-term cumulative effects, thereby improving diagnostic robustness under complex operating conditions. Moreover, the incorporation of textual prompts effectively compensates for the lack of domain knowledge in purely data-driven models, ensuring that diagnostic outputs align more closely with field operational logic and provide directly actionable recommendations. In summary, the proposed VLM-based diagnostic framework for shale gas well anomalies not only enhances the accuracy and reliability of anomaly diagnosis while reducing reliance on manual intervention, but also offers a scalable and low-maintenance solution for real-time diagnosis and response under complex operating conditions. This work opens a new avenue for intelligent anomaly diagnosis and supports the advancement of digital and automated production management in the oil and gas industry.

Co-author/s:

Liang Xue, Associate Dean, China University of Petroleum.

Dr. Haiyang Chen, Student, China University of Petroleum.

Shengdon Zhang, Student, Computer Network Information Center，Chinese Academy of Sciences.