Deep encoder-decoder networks for belt longitudinal tear detection

Abstract

The belt conveyor is susceptible to longitudinal tearing, which poses a serious threat to the safety of coal mines. Traditional methods for detecting longitudinal tears have limitations such as poor image quality, limited applicability, and high hardware costs. An improved encoder-decoder network was proposed to solve the longitudinal tear detection problem. This method utilizes a line structured light system for image acquisition. The input images are downscaled using a sorting algorithm to extract the information of pixels with high grayscale values as the input feature map for the neural network. The reduced-dimensional encoder-decoder network then semantically segments the input feature map, and the resulting pixel segmentation is mapped to the location of the longitudinal tear. Finally, the position and length of the tear are calculated by back-projecting the semantic segmentation result to the world coordinate system. Experimental results demonstrate that this method effectively reduces hardware resource consumption and improves detection speed. The DICE and MIOU scores for the improved network are 97.69% and 95.47%, respectively, while the recall and precision for improved detection are 96.60% and 95.67%, respectively. Therefore, this method can successfully monitor longitudinal tear failures and ensure the safety of transportation.