Research on the Application of Machine Learning Algorithms for Steel Defect Classification from the Perspective of Smart Infrastructure

本文基于美国国家标准与技术研究所和钢铁工业协会的UCI带钢缺陷公开数据库，对智慧基建中的钢材缺陷分类问题展开研究。样本来源于北美3家钢厂2018~2021年的1941个样本，共有27类特征。选择6种算法并采取数据预处理、特征工程以及超参数调整策略来建立高效的精准钢材缺陷智能分类方案。创新点包括：搭建多种算法融合模型；设计特征分类筛选与调优方案；采用SMOTE解决不均衡样本问题；设置完整的试验评价系统。结果表明，LightGBM和神经网络的精确率和召回率均超过96%。消融实验与参数敏感性分析证明了这些方法对于特征选取和超参数的重要性。后续的研究将扩大收集的样本数量，并尝试结合深度学习和计算机视觉等新技术，使模型更具有普适性、鲁棒性和更高的检测精度，促进智慧基建更广泛的智能化发展。This paper investigates the problem of steel defects classification in smart infrastructure based on the UCI Strip Steel Defects Public Database of the National Institute of Standards and Technology and the Steel Industry Association. The samples are derived from 1941 samples from 3 North American steel mills from 2018~2021, with a total of 27 types of features. Six algorithms are selected and data preprocessing, feature engineering, and hyper-parameter tuning strategies are adopted to establish an efficient and accurate intelligent classification scheme for steel defects. The innovations include: building a fusion model of multiple algorithms; designing a feature classification screening and tuning scheme; adopting SMOTE to solve the problem of unbalanced samples; and setting up a complete experimental evaluation system. The results show that the precision and recall of LightGBM and neural network are more than 96%. Ablation experiments and parameter sensitivity analysis demonstrate the importance of these methods for feature selection and hyperparameterization. Subsequent research will expand the number of collected samples and try to combine new technologies such as deep learning and computer vision to make the model more pervasive, robust, and higher detection accuracy, and promote the intelligent development of smart infrastructure more widely.