Research on a Spatiotemporal Fusion Model–Based Method for Predicting the Torsional State of a Conveyor Belt

Abstract Aiming at the problem that the belt twisting fault of pipe belt conveyors is prone to cause shutdown and production halt, this paper proposes a belt twisting state prediction method based on a spatiotemporal fusion model. Three types of spatiotemporal features—transient dynamics, short-term evolution, and long-term cumulative effects—are extracted from image sequences. A Hybrid Convolutional-Recurrent Spatiotemporal Network (HybridConvLSTM) is designed to realize the joint modeling of spatial geometric features and temporal dynamic evolution laws. A dataset containing 3 million image frames is constructed based on industrial scenarios. Experimental verification shows that the model achieves an accuracy of 89.7% and an F1-score of 83.2% on the test set, which are 7.5 and 17.6 percentage points higher than those of the ViViT and LSTM+Attention models, respectively. After TensorRT optimization, the inference speed reaches 28 FPS, meeting the real-time requirements of industrial applications. This method provides an efficient solution for the early warning of belt twisting faults and has excellent engineering application value.