Fault diagnosis of rolling bearings under strong noise interference faces challenges such as the masking of fault size features and difficulty in discrimination. Traditional diagnostic methods often neglect effective interactions across different scales and lack cross-layer feature fusion strategies, resulting in limited diagnostic performance under noise interference. To address these challenges, this paper proposes a Multi-Scale Hierarchical Feature Attention Interaction (MSHFIA) method to enhance the model’s ability to identify multi-scale fault features and improve noise robustness. The method first employs a multi-granularity feature collaboration fusion module to efficiently extract and integrate features of different granularities. A dual-path residual weighting strategy is introduced to strengthen the model’s focus on critical channels and spatial regions. Lastly, a cross-scale attention interaction mechanism is used to enable cross-layer attention guidance between shallow and deep feature maps, enhancing the saliency of key features. Experimental results under various noise levels demonstrate that the proposed model maintains robust perception and discrimination capabilities for fault features of different sizes, even under strong noise interference.