摘要
针对具有小样本、非线性、高维模式识别特点的冷轧带钢表面缺陷,且部分缺陷分布零散、不相连而导致后期识别数量增加、识别率低的情况,提出聚类与优化支持向量机相结合的改进分类算法。利用矩形框将缺陷进行标记,实现缺陷聚类合并,减少后期缺陷识别分类个数,便于后期正确识别判断;利用粒子群优化算法结合交叉验证自动选取最优参数,确定支持向量机结构,并结合实际生产线上出现频率较高的5类带钢缺陷进行分类研究。实验结果表明,相较于改进BP神经网络和网格优化的支持向量机,聚类与优化支持向量机相结合的改进分类算法不仅解决了位置接近的同种缺陷重复分类的问题,而且耗时短、缺陷正确识别率可达98%,符合实际生产线需求。
Considering small sample,nonlinear and high dimensional pattern recognition characteristics of cold rolled steel strip surface defects,and the distributions of some defects are scattered,which leads to high identification number and low recognition rate,an improved classification algorithm based on clustering and optimization of support vector machine（SVM）was proposed.By using rectangular box marked on the defects,the defect clustering was merged to decrease the number of defect recognition classification for correct recognition;particle swarm optimization（PSO）algorithm combined with cross validation was used to automatically select the optimal parameters,determine the structure of support vector machine and classify high frequency of five types of strip defects combined with the actual production line.The experimental results show that,compared with the improved BP neural network and grid optimization SVM,the improved classification algorithm based on clustering and optimization of SVM not only solves the same defects classification problem,but also takes shorter time and improves the defect recognition correct rate to 98%.The method meets the demand of practical production line.
出处
《塑性工程学报》
CAS
CSCD
北大核心
2016年第5期83-88,共6页
Journal of Plasticity Engineering
基金
国家自然科学基金资助项目(61104213)
中央高校基本科研业务费专项资金资助项目(JUSRP11008)
关键词
冷轧带钢
聚类
支持向量机
识别分类
cold rolled strip
clustering
support vector machine（SVM）
recognition classification
