基于改进GoogLeNet的玻璃安瓿瓶微孔检测  

Microhole detection of glass ampoule based on improved GoogLeNet

作　　者：曹林杰 任德均[1] 任秋霖 闫宗一 李鑫[1] 唐洪 CAO Lin-Jie;REN De-Jun;REN Qiu-Lin;YAN Zong-Yi;LI Xin;TANG Hong(School of Mechanic Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学机械工程学院,成都610065

出　　处：《四川大学学报（自然科学版）》2022年第5期47-53,共7页Journal of Sichuan University(Natural Science Edition)

摘　　要：在玻璃安瓿瓶包装完整性检测领域,常用高压放电法对微米级漏孔缺陷进行检测,针对现有方法存在的难以找到合适滤波方式、判别阈值依赖人工设计、检测准确率较低的问题,提出一种基于改进的GoogLeNet的微孔检测方法.对于原始放电电流数据,利用小波变换(WT),以广义Morse小波函数(GMW)为基小波,将一维的电流时间序列转换为二维的时频索引图以呈现数据完整的细节信息.在GoogLeNet原型基础上引入Relu激活函数以减少过拟合,将输入端卷积缩减至1层,然后进行了三种不同层次的Inception模块裁剪,对比分析发现只用前6个Inception模块并调高Inception(4d)的大尺寸卷积核占比时,模型能在参数量更少的情况下同样达到很好的微孔判别效果.在生产现场工控机中用训练好的模型替换原有算法,进行1000个正负样本的验证测试,结果表明该算法的准确率达到99.15%,阳性样品漏检率仅0.8%,优于现有方法的96.45%准确率和5.3%漏检率,具有较好实用价值.In the field of glass ampoule packaging integrity detection, high voltage discharge method is commonly used to detect micron-level leaky hole defects. In view of the existing methods, it is difficult to find appropriate filtering mode, discrimination thresholds depend on manual design, and detection accuracy is low, a microhole detection method based on improved GoogLeNet is proposed. For the original discharge current data, through the wavelet transform(WT),and using the generalized Morse wavelet function(GMW) as the basic wavelet, transform the one-dimensional current time series into a two-dimensional time-frequency index graph to present the complete details of the data. On the basis of GoogLeNet prototype, Relu activation function is introduced to reduce overfitting, the input convolution is reduced to 1 layer, and then Inception module cutting at three different levels is carried out. Comparative analysis shows that when only the first 6 Inception modules are used and the proportion of large-size convolution kernels is increased for Inception(4 D),the model can also achieve a better effect of microhole discrimination with fewer parameters.In the industrial computer of production site, the trained model was used to replace the original algorithm, and 1000 positive and negative samples were tested. The results show that the accuracy of the algorithm is 99.15%,and the positive sample missing rate is only 0.8%,which is better than the 96.45% accuracy rate and 5.3% missing rate of the existing method.

关 键 词：玻璃安瓿 高压检漏 时间序列分类 GoogLeNet 小波变换 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]