摘要
针对扣件状态检测算法适应性弱、误检率高的问题,通过平滑样本标签缓解卷积神经网络训练的过拟合问题,提出基于层次语义多项式DS融合的扣件状态分布学习。首先,以弱监督方式将图像子块卷积特征描述为高斯混合模型,通过高斯混合模型计算样本语义多项式(semantic multinomial, SMN);然后,为提高SMN对扣件样本的描述能力,对来自不同层次特征的SMN进行DS融合,获得样本状态分布,分布反映了不同标签的描述程度,实现了对单一标签的平滑。实验结果表明,将单一标签替换为状态分布进行网络训练,缩小了训练精度和验证精度的差距,误检率为1.9%,漏检率为2.3%,误检率相比于传统单标签网络降低了54%。所提算法能够缓解过拟合现象,提高网络泛化性能,实现鲁棒性的扣件状态检测。
Aiming at the problems of weak adaption and high false alarm in established algorithms for fastener condition detection, sample labels are smoothed to alleviate the over-fitting problem during the process of convolutional neural network training, and a fastener condition distribution learning algorithm is proposed based on Dempster Shafer(DS) fusion of hierarchical semantic multinomial(SMN). Firstly, the convolutional features of image sub-block are represented as Gaussian mixture models in a weakly supervised manner, and consequently sample SMNs are computed according to the Gaussian mixtures. Then, in order to improve the description ability of SMNs for fastener samples, DS fusion is conducted on the SMNs derived from multi-level features, which leads to the conditional distribution of each sample. Such distributions reflect the description degrees of different labels, and realize the smoothing of single labels. Experimental results shown that, by replacing the single labels with the condition distribution for CNN training, the gap between training and validation accuracies is reduced, and the proposed algorithm yields the false alarm rate of 1.9% and the missing rate of 2.3%, the rate of false alarm is reduced by 54% compared to conventional single label networks. The proposed algorithm can alleviate the over-fitting symptom, improve the generalization ability of networks, and realize robust detection of fastener conditions.
作者
黄翰鹏
罗建桥
李柏林
HUANG Hanpeng;LUO Jianqiao;LI Bailing(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China)
出处
《铁道标准设计》
北大核心
2022年第7期48-52,共5页
Railway Standard Design
基金
四川省重大科技专项课题(2018GZDZX0031)。
关键词
铁路扣件
状态检测
卷积神经网络
标签分布学习
语义多项式
railway fastener
condition detection
convolutional neural network
label distribution learning
semantic multinomial
