Distributed fiber optic acoustic sensing reservoir fluid production signal recognition based on ST-FMA

Abstract

To address the challenges of massive data redundancy, severe noise interference, and insufficient in-distribution model robustness when recognizing reservoir fluid production signals via Distributed Acoustic Sensing (DAS) in extreme environments, this paper proposes a novel Spatio-Temporal Feature Fusion and MixStyle Attention Network (ST-FMA). The core innovations of this model are threefold: (1) An ST-Fusion Block that integrates Graph Convolutional Networks (GCN) and Bidirectional Long Short-Term Memory networks to simultaneously capture spatial correlations across wellbore depths and the nonlinear dynamic evolution of signals over time. (2) A MixStyle module introduced to enhance the model’s in-distribution robustness under complex well conditions in the target well by regularizing the feature space distribution. (3) A Temporal Attention Block that automatically isolates key transient features while reducing computational redundancy. Comparative experiments demonstrate that the ST-FMA model achieves a 94.0% recognition accuracy on field datasets. Furthermore, ablation studies confirm that incorporating the MixStyle and Attention modules maintains recall rates for challenging production layer signals above 88.2% and 94.0%, respectively, significantly outperforming baseline models. Comparative experiments on computational complexity and inference time further demonstrate that ST-FMA significantly reduces model complexity while maintaining high inference speed, confirming its strong feasibility for practical engineering applications. This research provides a robust theoretical framework for the deep characterization of DAS signals and holds substantial engineering value for the intelligent monitoring of oil and gas wells.