文献类型：期刊文献

中文题名：基于多模型融合的数据驱动连铸铸坯缺陷检测方法

英文题名：Data?driven slab defect detection method based on multi?model fusion in continuous casting

作者：金翔[1];孙丽华[1];姜庆超[1]

机构：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237

年份：2025

卷号：49

期号：6

起止页码：113

中文期刊名：冶金自动化

外文期刊名：Metallurgical Industry Automation

基金：国家自然科学基金优秀青年基金项目(62322309)。

语种：中文

中文关键词：连铸过程;高斯混合模型;自编码器;贝叶斯推理;异常检测

外文关键词：continuous casting process;Gaussian mixture model;autoencoder;Bayesian inference;anomaly detection

摘要：在钢铁连铸过程中,由于钢种、铸速和温度等生产条件的动态变化,单一模型在铸坯表面缺陷检测中的适应性较差且往往无法灵活应对不同生产环境下的变化,容易在复杂的实际操作中失去精准性,从而导致误报或漏报问题频发,严重影响铸坯的质量控制。针对连铸过程上述特点,本文提出了一种基于多模型融合的数据驱动铸坯缺陷检测方法。首先利用高斯混合模型对工业过程数据进行聚类,将不同分布下的样本有效区分。接着在各个分布下的样本运用自编码器建立局部异常检测子模型,并确定各个子模型的控制限与重构误差。最后通过贝叶斯推理将各局部子模型的检测结果进行融合,实现对复杂多条件环境下的全局异常检测。利用某钢铁制造企业的连铸过程铸坯数据进行方法验证,通过实验对比分析不同模型检测性能,结果表明本文方法给出的检测效果最优,有效降低了漏报率和误报率。本文方法可以推广到其他场景的工业数据分析与建模,对于利用工业数据提升产品质量具有重要的参考价值。
In the steel continuous casting process,due to dynamic changes in production conditions,such as steel grade,casting speed,and temperature,single models exhibit poor adaptability in detec?ting surface defects of continuous casting slabs,and often fail to flexibly adapt to changes in different production environments,and tend to lose accuracy in complex practical operations,thereby leading to frequent false positives and missed detections,severely affecting the quality control of continuous cast?ing slabs.To address these challenges,this study proposes a data?driven slab defect detection method based on multi?model fusion.First,a Gaussian Mixture Model was employed to cluster industrial process data,effectively distinguishing samples from different distributions.Next,local anomaly detec?tion sub?models were constructed using autoencoders for each distribution.Control limits and recon?struction errors are determined for each sub?model.Finally,Bayesian inference was used to fuse the detection results of the local sub?models,enabling global anomaly detection in complex multi?condition environments.Using data from the continuous casting process of a certain steel manufacturing compa?ny,the proposed method is validated through comparative analyses of the detection performance of va?rious models.The results demonstrate that the proposed approach achieves superior performance,ef?fectively reducing both false negative and false positive rates.This method can be extended to other in?dustrial data analysis and modeling scenarios,offering significant reference value for improving product quality through industrial data utilization.