一种新联合损失函数优化的迁移学习神经网络磨粒识别研究

Research of Abrasive Particle Identification Based on Transfer Learning Neural Network with a New Joint Loss Function

为了提高齿轮齿条在不同工作情况下磨损状况识别的准确度,考虑到大模数齿轮齿条实际有效故障数据缺乏和数据标记缺失的特点,提出基于一种新联合损失函数优化的迁移学习神经网络(LCNNE)对大模数齿轮齿条磨粒识别方法。将LCNNE模型作为特征提取器提取特征,利用外部分类器SVM进行分类,验证了该方法在磨粒数据集上识别率达到99%左右,并且该方法的集成模型提取的特征输入SVM分类的识别率比VGG19和GoogleNet提取的联合特征高2%~3%。利用t-SNE技术,对DCNN、VGG19、GoogleNet和LCNNE模型的最后一个隐含层的提取特征进行可视化,证明了LCNNE模型的特征表达能力更强、识别效果更好。

In order to improve the accuracy of gear rack wear recognition under various operating conditions,considering the characteristics of lack of actual effective fault data and missing data mark of large modulus gear rack,a method of large-module pinion and rack abrasive particle identification based on transfer learning neural with a new joint loss(LCNNE)was proposed.The LCNNE model was used as a feature extractor to extract features and an external classifier SVM was used for classification.It is found that the recognition rate of the LCNNE model on the abrasive data set reaches about 99%,and the recognition rate of the feature input SVM extracted by the integrated model of the method is 2%and 3%higher than that of the joint feature extracted by VGG19 and GoogleNet.t-SNE technology was used to visualize the features extracted from the last hidden layer of DCNN,VGG19,GoogleNet and LCNNE models,it was proved that LCNNE model has stronger feature expression ability and better recognition effect.