Prototype-Driven Dual-Perspective Collaborative Contrastive Fusion Network for Rotating Machinery Fault Diagnosis

Recently, self-supervised learning frameworks based on contrastive learning have demonstrated superior performance in rotating machinery fault diagnosis with limited labeled data. However, existing methods still encounter several challenges in real applications. First, time–frequency features fail to interact effectively across domains, leading to semantic entanglement. Second, current contrastive frameworks offer weak supervisory signals, hindering the extraction of fine-grained discriminative features. Finally, insufficient consistency constraints make models susceptible to noise in low-label scenarios, thereby degrading diagnostic performance. To address these challenges, in this paper, we propose a self-supervised learning framework, called Prototype-driven Dual-perspective Collaborative Contrastive Fusion Network (PDC-CFN), for fault diagnosis. It integrates a multi‑scale semantic cognitive distiller and collaborative contrastive learning to efficiently extract features from large‑scale unlabeled data. Specifically, we design a multi-scale semantic cognitive distiller to achieve multi-level semantic alignment and adaptive fusion of time-frequency features at both local and global levels. In addition, a cross-domain interaction transformer (CDIT) is introduced to achieve semantic disentanglement between the time and frequency domains, effectively capturing cross-view correlated features. Furthermore, we design a collaborative contrastive learning strategy that integrates dual-perspective complementary contrastive learning (DPCCL) with adaptive prototype-driven contrastive learning (APDCL). It effectively enhances both feature discriminability and cluster compactness. Experimental results demonstrate that the proposed method achieves superior fault diagnosis performance, even in limited labeled data scenarios.