<p>Due to the swift advancement of artificial intelligence technology, semantic segmentation has emerged as a critical method for identifying rotational flaws in conveyor belts of pipe belt conveyors. Nonetheless, current segmentation algorithm models typically exhibit issues of excessive model parameters and sluggish inference speed, making it challenging to simultaneously satisfy the accuracy demands of conveyor belt edge segmentation and real-time requirements, thereby complicating their adaptation to the real-time monitoring needs of actual production lines. This research offers an efficient backbone network, D2Net, for real-time semantic segmentation of conveyor belt edges to address this issue. The network comprises a dual-branch Deep Dual-Resolution Network that facilitates information exchange across the branches via several bilateral fusion processes. A Multi-Scale Attention Aggregation Module (MSAAM) is developed to effectively broaden the network’s sensory field and improve segmentation performance via multi-scale contextual fusion of low-resolution feature maps. This research constructs a dataset of pipe belt conveyor operations at a steel factory across three locations to validate the model’s performance, and the suggested network is trained and tested using this dataset. The experimental findings indicate that the network attains an average intersection over union (mIoU) of 75.34<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2024_1614_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> on a dataset with a resolution of 512 <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11554_2024_1614_Article_IEq2.gif" Format="GIF" Height="13" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$\times $</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> 512, while the inference speed reaches 169.5 frames per second (fps), and the model comprises merely 22.9 MB of parameters. Compared to existing methods, D2Net successfully achieves a balance between accuracy and speed, making it well suited for real-time applications in practical industrial environments. This research demonstrates that the D2Net achieves an effective equilibrium between speed and accuracy in the belt edge segmentation of pipe belt conveyors. It exhibits optimal equilibrium and delivers robust technical assistance for real-time semantic segmentation in edge computing environments. To further advance research in this field, we commit to making the dataset used in this study publicly available upon acceptance of the paper. By sharing this data resource, we aim to support subsequent studies and foster continuous progress in the development of rotational detection technology for pipe belt conveyors.</p>

D2net: a dual-branch lightweight network for conveyor belt rotation detection in pipe belt conveyors

期刊论文
发表日期：

2025年1月3日

查看全文

Xingyu Wang,
Nini Hao,
Yu Yun,
Mengchao Zhang,
Yuan Zhang,
Zeqing Zhong

摘要

Due to the swift advancement of artificial intelligence technology, semantic segmentation has emerged as a critical method for identifying rotational flaws in conveyor belts of pipe belt conveyors. Nonetheless, current segmentation algorithm models typically exhibit issues of excessive model parameters and sluggish inference speed, making it challenging to simultaneously satisfy the accuracy demands of conveyor belt edge segmentation and real-time requirements, thereby complicating their adaptation to the real-time monitoring needs of actual production lines. This research offers an efficient backbone network, D2Net, for real-time semantic segmentation of conveyor belt edges to address this issue. The network comprises a dual-branch Deep Dual-Resolution Network that facilitates information exchange across the branches via several bilateral fusion processes. A Multi-Scale Attention Aggregation Module (MSAAM) is developed to effectively broaden the network’s sensory field and improve segmentation performance via multi-scale contextual fusion of low-resolution feature maps. This research constructs a dataset of pipe belt conveyor operations at a steel factory across three locations to validate the model’s performance, and the suggested network is trained and tested using this dataset. The experimental findings indicate that the network attains an average intersection over union (mIoU) of 75.34 $\%$ $%$ on a dataset with a resolution of 512 $\times $ $\times$ 512, while the inference speed reaches 169.5 frames per second (fps), and the model comprises merely 22.9 MB of parameters. Compared to existing methods, D2Net successfully achieves a balance between accuracy and speed, making it well suited for real-time applications in practical industrial environments. This research demonstrates that the D2Net achieves an effective equilibrium between speed and accuracy in the belt edge segmentation of pipe belt conveyors. It exhibits optimal equilibrium and delivers robust technical assistance for real-time semantic segmentation in edge computing environments. To further advance research in this field, we commit to making the dataset used in this study publicly available upon acceptance of the paper. By sharing this data resource, we aim to support subsequent studies and foster continuous progress in the development of rotational detection technology for pipe belt conveyors.