Keyhole status prediction based on voting ensemble convolutional neural networks and visualization by Grad-CAM in PAW
Keyhole status prediction based on voting ensemble convolutional neural networks and visualization by Grad-CAM in PAW
复制标题
基于投票集成卷积神经网络的锁孔状态预测和 PAW 中 Grad-CAM 的可视化
DOI：
10.1016/j.jmapro.2022.06.034
复制
发表时间：
2022
影响因子：
6.2
通讯作者：
Chuansong Wu
中科院分区：
文献类型：
--
作者：
Fangzheng Zhou;Xinfeng Liu;Xue Zhang;Yang Liu;C. Jia;Chuansong Wu
K-PAW（小孔等离子弧焊）在焊接中厚金属工件方面有广泛的应用。然而，由于小孔周围的液态金属上脆弱的力平衡，焊接过程是脆弱的，并且可能导致烧穿或未熔合。为了优化焊接质量，大量的工作已经致力于通过视觉传感和深度学习算法来预测熔透/小孔状态。然而，单一的网络模型是具有挑战性的提取综合特征的弱区分力的熔池图像。本文重点研究了熔池表面图像与预测/小孔状态之间的对应关系。提出了一种新的预测方法，该方法基于八个经典的CNN训练、比较和融合。对于这些单一模型，预测精度高于95%，速度超过每秒10帧（FPS）。采用梯度CAM方法对模型进行可视化，清晰地显示了聚焦特征区域。虽然被认为是黑盒，这些预测模型被认为是强大的，当提取的特征与先验知识一致。基于所选择的三个鲁棒模型（即，InceptionNetV 3、InceptionResNetV 2和XceptionNet），并实现了96.62%的评估准确率。
K-PAW (Keyhole Plasma Arc Welding) has a wide range of applications in welding medium-thick metal workpieces. However, the welding processes are vulnerable due to the fragile force balance on the liquid metal around the keyhole, and burn-through or lack of fusion might be caused. In order to optimize the welding quality, considerable work has been devoted to penetration/keyhole status prediction by visual sensing and deep learning algorithms. However, a single network model is challenging to extract comprehensive features of weak-discrimination weld pool images. This paper focused on the correspondence between the topside weld pool images and the prediction/keyhole status. A novel prediction method has been proposed based on eight classical CNNs trained, compared, and fused. For these single models, the prediction accuracy is higher than 95 %, with the speed of faster than 10 frames per second (FPS). The model visualization by the Grad-CAM method was performed to show the focused feature regions clearly. Although regarded as black boxes, these prediction models are considered robust when the extracted features are consistent with the prior knowledge. KeyholeVot, a voting ensemble decision model, was established based on the selected three robust models (i.e., InceptionNetV3, InceptionResNetV2, and XceptionNet) and achieved 96.62 % evaluation accuracy.