基于迁移学习的正温度系数热敏电阻表面损伤分类算法研究

Research on Surface Damage Classification Algorithm of Positive Temperature Coefficient Thermistor Based on Transfer Learning

正温度系数热敏电阻(Positive Temperature Coefficient Resistor,PTCR)在生产转运过程中易遭受表面损伤,人们通常在出厂前对其甄别损伤,实施筛选。然而现在常采用人工手动分拣,效率低,容易漏选,因此使用机器视觉技术来进行智能化分类将极大地提升其分拣效率。论文提出了一种基于迁移学习的PTCR表面损伤分类算法,此算法将获取的数据集进行数据增强后在三种卷积神经网络模型Inception-V3、ResNet50、Xception分别提取特征向量,然后将提取到特征向量进行合并训练,使用SoftMax进行分类,最后获得了准确率最高的OurNet(自命名)模型。该模型可以自动识别出边缘破损、完全破裂、表面刮擦等三种表面损伤,从而保障产品质量,提高生产效率。

Positive temperature coefficient resistors(PTCR)are susceptible to surface damage during transfer and transporta-tion in the manufacturing process,so they are usually screened for damage before they leave the factory.However,existing methods often use manual sorting,which is inefficient and easy to miss,so it is especially important to use machine vision to perform intelli-gent sorting.In this paper,a migration learning-based algorithm for classifying PTCR surface damage is proposed.This algorithm ex-tracts feature vectors in three convolutional neural network models Inception-V3,ResNet50,and Xception after data enhancement of the acquired dataset,and then combines the extracted feature vectors for training and uses SoftMax for classification,and finally obtains OurNet(self-named)model with the highest accuracy.The model can automatically identify three kinds of surface damage,namely,edge breakage,complete rupture and surface scratching,so as to guarantee product quality and improve production effi-ciency.