Localization of Conveyor Belt Damages Using a Deep Neural Network and a Hybrid Method for 1D Sequential Data Augmentation

The article describes an innovative method for detecting conveyor belt damage using a strain gauge system and deep learning techniques. The strain gauge system records measurement data from conveyor belts. A major challenge encountered during the research was the insufficient amount of measurement data for effectively training deep neural networks. To address this issue, the authors implemented a hybrid data augmentation method that combines generative artificial intelligence techniques and signal analysis. The TimeGAN model, based on Generative Adversarial Networks (GANs), was used to augment data from undamaged belts. Meanwhile, the superposition of one-dimensional observation sequences was applied to generate data representing damages by combining signals from randomly selected undamaged runs with strain gauge system responses to damage, effectively increasing the number of samples while accurately replicating defect conditions. For damage diagnosis, a Long Short-Term Memory Network with an attention mechanism (LSTM-AM) was employed, enabling anomaly detection in strain gauge signals. The application of the LSTM-AM algorithm allows for real-time monitoring of conveyor operation and facilitates precise localization and estimation of damage size through data synchronization.