Equipment anomaly detection is a critical task in the industrial domain. However, traditional detection approaches often falter due to the extreme imbalance of industrial datasets and the irregular, unexpected behavior of anomalous events. In this paper, we propose a feature enhancement framework, termed the clustering-based autocorrelated graph network (CAGN). The method begins by partitioning the dataset via an unsupervised clustering algorithm and constructing graphs at the cluster level to capture latent relationships among samples. Subsequently, an autocorrelation network and a graph neural network are employed to extract and reinforce structural and contextual features within the cluster-generation graph, thereby improving the separability between normal and anomalous instances. The proposed method has been comprehensively evaluated on multiple publicly available benchmark datasets as well as a practical dataset from injection moulding machines, demonstrating its robustness and performance advantages. The source code and model are available at https://github.com/napoleonkuroba/CAGN .