Abstract: In petroleum exploration, fault identification serves as a critical component of seismic interpretation, with its accuracy directly determining exploration results. Manual interpretation, constrained by mass data processing and complex geological structures, is plagued by low efficiency, inadequate precision, and strong subjectivity, and thus fails to meet the dual requirements for high efficiency and high accuracy in modern petroleum exploration. Deep learning has propelled intelligent fault identification, but it still faces three major bottlenecks: loss of small-scale fault information, difficulty in multi-scale fault extraction, and noise interference. To address these challenges, this paper proposes a fault identification neural network based on a multiscale hybrid attention mechanism, namely multiscale spatial channel UNET (MSC-UNET). A multiscale feature fusion module is introduced into the encoder, so that fine-grained edge details and coarse-grained structural semantics of faults can be simultaneously extracted, significantly enhancing cross-scale feature representation capability. Then, to mitigate noise interference, a parallel spatial–channel attention mechanism is incorporated: the spatial branch focuses on spatial fault distribution, and the channel branch suppresses noise responses, effectively enhancing discriminative feature extraction. Furthermore, a residual feature refinement module is designed to bridge the gap between shallow and deep features, accelerating model convergence and reducing small fault information loss. Finally, the MSC-UNET model is tested on both theoretical and field datasets, and evaluated against the commonly used UNET and VNET architectures as well as the ant-tracking algorithm. Experimental results demonstrate that the proposed MSC-UNET-based fault identification method outperforms other approaches in accuracy, fault continuity, and noise immunity, and it is promising for future application.