Fault diagnosis plays a crucial role in ensuring the safe and stable operation of complex industrial systems. Among various existing methods, graph convolutional network (GCN)-based methods can handle graph-structured data commonly found in industrial systems, due to their capability to capture complex relationships within datasets through association graphs. However, conventional GCN models exclusively consider positive association relationship between data, but do not account for the negative one, resulting in embedding vectors that exhibit clustering tendencies and the over-smoothing issue. To address these challenges, a positive–negative GCN-based fault diagnosis method is proposed in this paper. In the method, the conventional association graph is treated as a positive graph, and a corresponding negative one is constructed to capture negative relationships within the dataset. Furthermore, a graph-level data aggregation strategy is introduced to aggregate the information from both association graphs, and incorporate them into the forward propagation process. Fault diagnosis is achieved using a iteratively trained GCN model. Experimental results on two cases demonstrate that the proposed method outperforms existing fault diagnosis methods in terms of diagnostic performance.