Visual-Language contrastive learning for zero-shot compound fault diagnosis in sucker rod wells

• Published in: Measurement : journal of the International Measurement Confederation 2025-02, Vol.243, p.116320, Article 116320
• Main Authors: Wang, Xinyan, Zhang, Liming, Wang, Yunsong, Nie, Hao, Shen, Yaorui, Zhang, Kai
• Format: Article
• Language: English
• Subjects: Compound fault diagnosis; Deep Learning (DL); Dynamometer card; Oil and Gas; Sucker Rod Pump (SRP); Zero-shot learning
• ISSN: 0263-2241
• DOI: 10.1016/j.measurement.2024.116320

•The zero-shot framework diagnoses unknown compound faults using single fault samples.•Visual features and label semantics of dynamometer cards are aligned through contrastive learning.•A hierarchical diagnostic strategy is introduced to reduce misdiagnosis from unseen faults.•Achieved 86.96% F1 score in generalized zero-shot diagnosis with 23 fault types. Deep learning has achieved significant success in diagnosing sucker rod pump wells. However, the complex and variable working environments make it impractical to collect and label a comprehensive range of compound fault samples for traditional models. Therefore, this paper proposes a zero-shot diagnosis framework that identifies unseen compound faults training on single fault samples. The framework consists of two processes: Contrastive Language-Visual Pre-training (CLVP) and Hierarchical Diagnosis. Firstly, a fault-label semantic encoder and a visual feature extractor for the dynamometer cards are constructed. The fault semantics and visual features are aligned through contrastive learning. Subsequently, the image-text matching pattern is applied to the coarse-grained and fine-grained filtering process, mitigating the misdiagnosis from unseen faults in the general zero-shot diagnosis. Furthermore, an experiment involving 11 single faults and 12 compound faults demonstrates that the proposed method achieved an F1 score of 86.96%, effectively meeting the requirements in oil field development.