基于人工免疫和支持向量机的焊接缺陷分类方法

Defects Classification Method of Welding Joints Based on Artificial Immune and Support Vector Machine

焊接缺陷的分类属于不平衡样本多分类问题,在不平衡样本中重要的特征子集通常也不相同,需要通过特征选择获得更具差异性的特征,提高稀有类样本的识别率。使用支持向量机作为分类器评价特征子集,人工免疫系统算法寻找可靠的特征,并且利用人工免疫算法优化支持向量机的参数,目的是同时完成参数优化和特征子集的选择。将此算法应用于低碳钢对接焊、低碳钢角接焊、低碳钢T型焊、低碳钢搭接焊、不锈钢对接焊、不锈钢角接焊、不锈钢T型焊、不锈钢搭接焊8类焊接缺陷数据集上进行常见焊接缺陷的气孔、夹渣、裂纹、未熔合、未焊透和伪缺陷的分类识别,并与未进行特征选择直接进行分类的结果进行比较及不同特征选择和分类算法下进行比较。结果表明,采用本文算法,焊接缺陷的气孔、夹渣、裂纹、未熔合、未焊透和伪缺陷的的平均分类准确率达到了(96.21±0.67)%,平均敏感度值达到了(85.43±1.65)%,比传统的基于相关性的特征选择算法(CFS)、最小冗余最大相关性算法(m RMR)、粗糙集条件互信息算法(RCMI)特征选择法和贝叶斯(Bayes)、分类回归树(CART)分类方法的组合具有明显的提高。因此,本文算法优于传统分类方法,利用较少属性的同时提高焊接接头缺陷分类准确率,保证稀有类的识别率,并能够提供不同缺陷的最优特征参数。

Classifications of welding defects are a multi-classification problem of unbalanced samples and the feature distribution of unbalanced samples also vary. Feature selection is needed to acquire more discriminating features and improve the recognition rate of rare samples. In order to complete the parameter optimization and selection of feature subset at the same time, the support vector machine was selected as the classifier to evaluate feature subset, and the artificial immune system algorithm was used to search reliable features and optimize the parameters of support vector machine. This algorithm was applied to classify and identify common welding defects such as pore, slag inclusion, crack, incomplete fusion, the lack of penetration and pseudo-defect on eight types of weld defect dataset, including mild steel butt welding, mild steel fillet welding,mild steel T-shape welding, mild steel lap welding, stainless steel butt welding, stainless steel fillet welding, stainless steel T-shape welding, and stainless steel lap welding, and was compared with direct classification result with no feature selection and under different feature selections and classification algorithms. Results indicated that the average classification accuracy rate of welding defects of pore, slag inclusion, crack, incomplete fusion, lack of penetration and pseudo-defect was（96.21±0.67）%, average sensitive value was（85.43±1.65）%, and had an obvious increase compared with traditional algorithms of correlation-based feature selection algorithm（CFS）, minimum redundancy maximum relevance algorithm（m RMR）, rough condition mutual information algorithm（RCMI）, and the combination of Bayes and classification and regression tree（CART） on the basis of pertinence. Therefore, the proposed algorithm is superior to traditional classification methods.