Transformer-enhanced meta-learning for few-shot fault diagnosis of electric submersible pump

Meta-learning represents a promising approach for few-shot fault diagnosis (FSFD). However, when directly applied to the fault diagnosis of electric submersible pumps (ESPs), its diagnostic performance requires enhancement due to insufficient consideration of the long-term dependencies among ESP variables. To address this issue, we propose a novel technique termed Transformer-Enhanced Meta-Learning (TEML), in which an encoder network based on the Transformer architecture serves as an embedding module integrated into a meta-learning framework that optimizes both intraclass and interclass relationships. Specifically, we develop a multivariate time-series tokenization strategy to preprocess ESP data, facilitating its effective input into the Transformer’s encoder network. The proposed TEML method not only adeptly captures the multivariate long-term dependency characteristics inherent in ESP data but also exhibits improved discrimination performance in scenarios characterized by limited sample availability. Extensive experiments conducted on practical defective datasets collected from Energy Development Co., Ltd., China, demonstrate that the proposed TEML approach outperforms state-of-the-art techniques while yielding favorable diagnostic results.