Effect of image size on performance of a plastic gear crack detection system based convolutional neural networks: an experimental study
Effect of image size on performance of a plastic gear crack detection system based convolutional neural networks: an experimental study
复制标题
DOI：
10.1117/12.2557977
复制
发表时间：
2020-04
期刊：
影响因子：
--
通讯作者：
Huy Kien Bui;D. Iba;Y. Tsutsui;A. Kajihata;Yu Lei;N. Miura;Takashi Iizuka;A. Masuda;A. Sone;I. Moriwaki
中科院分区：
文献类型：
--
作者：
Huy Kien Bui;D. Iba;Y. Tsutsui;A. Kajihata;Yu Lei;N. Miura;Takashi Iizuka;A. Masuda;A. Sone;I. Moriwaki
如今，深度学习（DL）已成为一种快速发展的技术，并为处理和分析大机械数据提供了有用的工具。许多研究项目使用卷积神经网络（CNN）从机械数据中进行故障分类取得了成功，这是DL最广泛的研究方面之一。在这种趋势下，我们使用深度卷积神经网络（DCNN）构建了POM（聚甲醛）齿轮的裂纹检测系统。在我们的工作中，从塑料齿轮上采集的振动数据被可视化并标记为裂纹数据或非裂纹图像。基于预训练的VGG16的DCNN，首先从ImageNet的数据中预学习，然后从标记的图像中重新学习，用于分类塑料齿轮的裂纹或非裂纹情况。在本研究中，数据集的图像质量失真（如模糊、噪声或对比度）是稳定的，不会影响DCNN的性能。然而，图像的尺寸，这保持了至关重要的作用，以达到高性能的检测系统，一直是未知的。因此，本文揭示了从振动数据创建的图像的优化大小，以实现高精度的学习。
Nowadays, deep learning (DL) has become a rapidly growing and provides useful tools for processing and analyzing big machinery data. Many research projects achieved success in failure classification from machinery data using convolutional neural networks (CNNs), one of the most extensive study aspects of DL. On this trend, we constructed a crack detection system of POM (Polyoxymethylene) gears using a deep convolutional neural network (DCNN). In our work, vibration data collected from plastic gears was visualized and labelled as crack data or non-crack images. A DCNN based on pre-trained VGG16, which firstly pre-learned from ImageNet’s data and then re-learned from the labelled images, is utilized to classify crack or non-crack situations of plastic gears. In this case of study, the image quality distortions of the dataset such as blur, noise or contrast are stable and do not affect the performance of the DCNN. However, the image size, which keep a vital role to reach high performance of the detection system, has been unknown. Hence, this paper reveals an optimized size of images created from vibration data for high-accuracy of learning.