A self-supervised multi-task framework for downhole fault diagnosis in oil fields: Enhancing temporal feature learning under limited labels

In newly developed oil fields, the accumulation of historical labeled data is relatively scarce, making it challenging to obtain large volumes of annotated data. To address this challenge, this paper proposes a downhole fault diagnosis framework based on a complex self-supervised multi-task collaborative training strategy. The model integrates three temporal feature-driven self-supervised tasks: a contrastive module based on differential temporal features and noise perturbations, a neighborhood-aware completion module, and a temporal consistency-aware relational module. These tasks jointly update a shared backbone network with an adaptive multi-task loss that dynamically balances each module’s contribution, enhancing feature representation quality and ensuring stable training. With only limited fine-tuning data, the proposed method achieves superior diagnostic performance compared with baselines. Experiments on downhole datasets show accuracy of 0.9575, precision of 0.9709, recall of 0.9430, and F1-score of 0.9567. Additional evaluations on the UCR datasets further demonstrate its generalization ability, where it consistently outperforms other approaches under varying labeled data proportions. The study highlights the effectiveness of self-supervised pre-training combined with adaptive loss weighting in enhancing temporal feature learning, providing a promising solution for intelligent fault diagnosis in data-scarce industrial scenarios.