基于代价敏感卷积神经网络的扣件缺陷检测算法

为解决扣件数据集不平衡问题,引入代价敏感策略对卷积神经网络算法进行改进,并以此检测断裂、丢失的缺陷扣件。该算法借鉴AdaBoost算法的思路,在训练过程中对整体误差函数中每个样本分配不同的权重,并依据先前模型的错误率不断地加以调整,使算法关注各个类别中的难学习样本,并对调整后的权重按类别进行归一化处理,以增大小类样本的关注度。分别在高速铁路无砟轨道和普速铁路有砟轨道2个扣件数据集上进行对照试验验证算法的有效性。引入G-mean作为评价指标平衡不同类别的召回率。结果表明:将改进后算法应用于高速铁路无砟和普速铁路有砟轨道的扣件数据集,改进后算法的G-mean值比原算法分别提高10%和25%以上;比传统的扣件识别方法分别提高13%和39%以上。

Fastener Defect Detection Algorithm Based on Cost-Sensitive Convolutional Neural Network

To solve the imbalance problem of fastener dataset,the cost-sensitive strategy was introduced to improve the convolutional neural network algorithm,and to detect fastener defects,such as fracture and loss.Based on the idea of AdaBoost,the algorithm assigned different weights to each sample in the overall error function during the training process,and constantly updated them according to the error rate of the previous models.The algorithm focused on the hard-to-learn samples in each category,and normalized the updated weights according to the category so as to increase the attention on the minor class samples.The effectiveness of the algorithm was verified by comparative experiments on two fastener data sets of the ballastless track of high-speed railway and the ballasted track of conventional speed railway.Meanwhile,G-mean was introduced as an evaluation index to balance the recall rates of different categories.Results show that:the improved algorithm is applied to the fastener data sets of the ballastless track of high-speed railway and ballast track of conventional speed railway,the G-mean values of the improved algorithm are increased by more than 10%and 25%respectively compared with the original algorithm,and 13%and 39%respectively higher than traditional fastener identification methods.