State Key Laboratory for Manufacturing Systems Engineering, Xi'an Jiaotong University, Xi'an 710000, China; International Joint Laboratory for Micro/Nano Manufacturing and Measurement Technology, Xi'an Jiaotong University, Xi'an 710000, China.
State Key Laboratory for Manufacturing Systems Engineering, Xi'an Jiaotong University, Xi'an 710000, China; International Joint Laboratory for Micro/Nano Manufacturing and Measurement Technology, Xi'an Jiaotong University, Xi'an 710000, China.
Ultrasonics. 2021 Apr;112:106344. doi: 10.1016/j.ultras.2020.106344. Epub 2021 Jan 2.
High precision classification of ultrasonic signals is helpful to improve the identification and evaluation accuracy for detecting defects. In the previous research, the deep neural network (DNN) has been used to classify the signal with obvious differences. But for different defects of the same depth, or when the defect position is close, the ultrasonic A-scan signal curve is very similar, causing the classification accuracy not high enough. In this paper, an optimized softmax classifier is proposed based on the traditional softmax classifier, and the convolution neural network (CNN) framework is built, which can achieve the accurate classification of signals with similar curves. Through a comparative experiment, the performance of the proposed classifier is evaluated from the loss curve decline rate, classification accuracy and feature visualization. The results show that the classifier has high classification accuracy and strong robustness.
超声信号的高精度分类有助于提高检测缺陷的识别和评估精度。在之前的研究中，已经使用深度神经网络 (DNN) 对信号进行分类，这些信号具有明显的差异。但是，对于同一深度的不同缺陷，或者缺陷位置较近时，超声 A 扫描信号曲线非常相似，导致分类精度不够高。在本文中，提出了一种基于传统 softmax 分类器的优化 softmax 分类器，并构建了卷积神经网络 (CNN) 框架，可实现相似曲线信号的精确分类。通过对比实验，从损失曲线下降率、分类准确率和特征可视化三个方面评估了所提出的分类器的性能。结果表明，该分类器具有较高的分类准确率和较强的鲁棒性。
