据红外技术网，于2023年04月报道，基于深度学习与域自适应的工件涡流热成像的缺陷检测。

        摘要: 机械设备运行过程中，标记的故障样本量小，导致建立的模型故障诊断准确率低，为此本文提出一种结合深度学习与域自适应的工件涡流热成像的缺陷检测方法。

        首先将注意力机制引入深度残差网络ResNet50中，加强模型的特征提取能力；然后将源域和目标域数据送入改进的ResNet50网络中提取深度特征，并且在网络的全连接层中引入局部最大均值差异，用于缩小两域特征间的分布差异，以此实现相关子域的分布对齐；最后在网络的Softmax分类器中实现对工件金属材料的缺陷检测。在公开的磁瓦数据集和本文实验采集的金属板涡流红外图像数据集上进行实验，结果表明，本文方法对涡流红外图像的裂纹缺陷检测识别准确率较高，通过t分布随机邻居嵌入方法对分析结果可视化，验证了本文方法的优越性。

        关键词: 涡流热成像 /  深度残差网络 /  注意力机制 /  域自适应 /  局部最大均值差异 

        Abstract: When operating mechanical equipment, the number of fault samples marked is small, which leads to low accuracy of the fault diagnosis of the established model. Therefore, this study proposes a defect detection method for eddy current thermal imaging of a workpiece that combines depth learning and domain adaptation. First, the attention mechanism is introduced into the deep residual network ResNet50 to enhance the feature extraction capability of the model. Then, the source and target domain data are sent into the improved ResNet50 network to extract the depth features. The local maximum mean difference is introduced into the full connection layer of the network to reduce the distribution difference between the two domain features to achieve the distribution alignment of related sub-domains. Finally, workpiece metal material defects were detected in the Softmax classifier of the network. The experiment was conducted on the open magnetic tile dataset and eddy current infrared image dataset of the metal plate collected during the experiment. The results show that the method proposed in this paper is highly accurate in detecting and recognizing crack defects in eddy current infrared images. The advantages of the method in this study were verified by visualizing the analysis results using the t-distribution random neighbor embedding method.

Key words: eddy current thermal imaging /  deep residual network /  attention mechanism /  domain adaptation /  local maximum mean discrepancy