Abstract: To address the challenges posed by the spatiotemporal coupling of anomalous features and the difficulty in capturing collaborative correlations resulting from the coordinated operation of multiple devices in the complex industrial environment of a coal preparation plant, this study proposes a method for identifying video anomalies in coal preparation plants based on the ST−GCN algorithm. On−site video data is captured using industrial cameras. A hybrid Gaussian model is employed for background subtraction to extract feature maps of foreground objects. These feature maps are fed into the ST−GCN, incorporating an attention mechanism to weight the adjacency matrix, thereby fusing temporal continuity with spatial topological relationships. Subsequently, a spatiotemporal graph neural network is used to classify and score the confidence levels of anomaly events. Experimental results show that this method achieves recognition confidence scores above 96.5% for typical operational conditions such as abnormal equipment vibration, conveyor belt deviation, and coal flow blockages. The recognition latency is consistently within 3.5 ms, with inter−frame fluctuation rates below 0.2% and risk rates all below 1.5%. This method effectively eliminates environmental disturbances such as lighting and dust, enabling accurate, stable, and real−time identification of abnormal events across multiple operating conditions, thereby providing reliable technical support for intelligent management and safe production at coal preparation plants.