<p>Fault diagnosis in complex industrial systems has witnessed significant progress through graph convolutional networks (GCNs). However, existing methods predominantly focus on binary relationships between samples or single-view feature extraction, failing to simultaneously model lower-order local features and higher-order group interactions, which limits further improvements in diagnostic performance. To address these limitations, this paper proposes a novel industrial intelligent diagnosis framework based on multi-granularity dual contrastive learning (MG-DCL). The proposed method employs Chebyshev graph convolutional network (ChebyGCN) to extract node features by constructing ordinary graphs that model lower-order pairwise relationships between samples. A multi-scale contrastive learning mechanism is designed to aggregate features from different neighborhood orders, enabling effective fusion of lower-order local features. Furthermore, hypergraphs are built to capture higher-order complex associations among samples, with hypergraph neural network (HGNN) subsequently extracting higher-order group-wise features. To ensure complementary integration of lower-order and higher-order information, a multi-view contrastive learning framework is implemented, aligning representations across topological granularities. Experimental results demonstrate that the proposed method outperforms baseline approaches across multiple datasets, achieving improved diagnostic accuracy and enhanced robustness. Specifically, the proposed method achieves the best performance for almost all fault categories across different datasets, and its average diagnostic accuracy is approximately 1% higher than that of the second-best method.</p>