Abstract: In response to the adverse impacts of oil well failures on production efficiency and operational safety in oilfields, this study focuses on rapid and accurate diagnosis and effective mitigation of such failures in complex environments. An intelligent framework for oil well fault diagnosis and disposal based on large models is proposed. First, a multi-modal association rule mining algorithm (M–ARMA) is designed to extract association rules between abnormal data and fault types from multimodal data sources such as pressure, load, and indicator diagrams of oil wells. Second, the XLNet-DC model is proposed to automatically extract fault handling rules from oil well failure reports and maintenance logs. Finally, a knowledge graph for oil well fault diagnosis and handling is constructed, and inference rules for abnormal fault handling are established through entity matching. By integrating real-time faults detected by the oil well monitoring module, the framework enables intelligent fault diagnosis and disposal via knowledge-based reasoning. Experimental validation in an oilfield demonstrates that the proposed method achieves an F₁ of 93% in knowledge extraction and a fault diagnosis accuracy of 96%. The framework effectively supports the integrated matching of accurate fault identification with appropriate handling strategies, thereby enhancing the safety and stability of oilfield production.

Key words: large model, multimodal data, knowledge graph, knowledge reasoning, fault diagnosis and handling