<p>In the field of intelligent fault diagnosis, graph neural networks (GNNs) can create a richer fault feature space by modeling dependencies between sensor signals and embedding them in a structural attribute graph. However, existing GNNs models typically use monitoring signals to directly construct graphical datasets, which suffer from poor representation and edge redundancy. To handle fault diagnosis tasks in noisy environments, the feature representation of graph data and the adaptability of GNNs models need improvement. This paper proposes a hybrid graph neural network based on an adaptive feature fusion network (AFFN-HGNN) for rotating machinery fault diagnosis to address these challenges. First, we introduce an approach to construct multiple knowledge-based graphs, enhancing the exploration of fault information from various perspectives and levels. Next, an adaptive feature fusion network is developed to a) adaptively fuse the time-spatial dependencies from the multiple knowledge-based graphs and hidden correlations between nodes and b) dynamically adjust the fusion weights of each feature layer to improve learning efficiency while minimizing computational overhead. Experimental results on three mechanical systems in different noisy environments show that the proposed method offers improved robustness and accuracy compared to recent GNNs-based fault diagnosis models.</p>