TBM岩渣数据集

全断面岩石掘进机在道路掘进过程中，刀盘挤压切削岩体容易产生刀盘磨损及损坏，从而产生经济损失，因此需要检测刀盘磨损的理论和技术。岩渣是掘进过程的直接产物，携带丰富的信息，能够反映当前的施工状况，因此可以利用这些信息，通过岩渣识别间接实现刀盘的监测。本文提出了一种基于卷积神经网络的岩渣识别算法，在岩渣数据集上实现了96.5%的分类准确率。随后为了便于FPGA硬件部署，提出一种网络压缩方法，将网络规模压缩到原始网络的2.28%，同时总准确度相比原网络仅下降了0.9%。最后将算法使用OpenCL技术在Intel Arria 10 GX1150上实现了部署，达到了224.54GOP/s的吞吐率以及11.23GOP/s/W的能效比。

During road excavation with a full-face rock tunnel boring machine (TBM), the cutterhead's extrusion and cutting of rock mass easily cause wear and damage to the cutterhead, resulting in economic losses. Therefore, theories and technologies for detecting cutterhead wear are urgently needed. Muck is the direct byproduct of the excavation process, which carries abundant information reflecting the current construction status. Thus, indirect cutterhead monitoring can be realized via muck recognition by utilizing this information. This paper proposes a muck recognition algorithm based on convolutional neural network (CNN), which achieves a classification accuracy of 96.5% on the muck dataset. Subsequently, to facilitate FPGA hardware deployment, a network compression method is proposed, which reduces the network scale to 2.28% of the original network, while the overall accuracy only decreases by 0.9% compared to the original network. Finally, the algorithm is deployed on the Intel Arria 10 GX1150 using OpenCL technology, achieving a throughput of 224.54 GOP/s and an energy efficiency ratio of 11.23 GOP/s/W.

搜集汇总

背景与挑战

背景概述

TBM岩渣数据集是一个用于全断面岩石掘进机刀盘磨损监测的数据集，包含岩渣识别算法和硬件部署方法，数据量为226.87MB，文件格式为zip。数据集通过卷积神经网络实现了96.5%的分类准确率，并优化了网络规模以适应FPGA硬件部署。