<p>As one of the key components of rotating equipment, rolling bearings have been of widespread concern. Accurate fault diagnosis of rolling bearings is significant for the safety and reliability of rotating machinery systems. Deep learning-based fault diagnosis approaches have greatly progressed in recent years, but most require extensive training data. Owing to the multiplicity of operating conditions and the sparseness of failure samples, there remains a gap between the current diagnosis approaches and real-world implementations. To solve this problem, a deep meta-transfer learning approach based on Relation Networks (DRTL-NGRU) is proposed, which combines the strength of deep transfer learning (DTL) and deep meta-learning (DML), aiming to effectively leverage prior knowledge from known faults and achieve rapid adaptation to new faults. In addition, inspired by the Nbeats model, a novel BiGRU architecture Nbeats BiGRU (NGRU) is proposed, which abandons the conventional cascade architecture and adopts a residual architecture to obtain the salient and subtle features in the fault feature vector by layer-by-layer decomposition, and applies feature fusion to obtain a more accurate representation of fault features. In this paper, two transfer scenarios for intelligent diagnosis of rolling bearings named conditional transfer and artificial-to-natural transfer are considered. Eight few-shot learning approaches are constructed for the few-shot diagnosis of two public datasets. Experimental results demonstrate the outperformance of the proposed approach in solving few-shot fault diagnosis under conditional transfer and artificial-to-natural transfer.</p>