Multi-Scale Spatial-Temporal Graph Model for Unsupervised Anomaly Detection in the Wheat Flour Transportation Process

Wheat flour transportation involves extended transit periods, which present considerable challenges for safety risk oversight. Therefore, it is essential to develop an efficient anomaly detection method to support risk assessment during this stage. However, existing anomaly detection methods often neglect the coupling effects across different time scales and the spatial clustering of hazard factors. To address this limitation, we propose a multi-scale spatial-temporal graph model-based unsupervised anomaly detection framework (MSTUAD), which can simultaneously capture the spatial-temporal correlations between importance and hazard factors across multiple time scales. Specifically, feature maps are first constructed for each hazard factor within a given time window to represent coupling effects. Secondly, a multi-scale spatial-temporal graph model is designed to extract spatial-temporal characteristics from these feature maps. Finally, a reconstruction model based on a variational autoencoder learns latent representations of spatial-temporal features of hazard factors, thereby capturing the intrinsic characteristics of normal wheat flour. Experimental validation on the wheat flour transportation hazard factor dataset and three public industrial datasets demonstrates that MSTUAD significantly outperforms state-of-the-art anomaly detection methods, achieving an average F1-score greater than 84.55%. This approach provides valuable decision support and technical guidance for relevant regulatory authorities.