基于GoogLeNet Inception V3模型的火电用钢金相组织智能识别  

Intelligent Recognition of Microstructure of Steel in Thermal Power Unit Based on GoogLeNet Inception V3 Model

作　　者：张艳飞 张永志[2] 公维炜 刘孝 卫志刚 ZHANG Yanfei;ZHANG Yongzhi;GONG Weiwei;LIU Xiao;WEI Zhigang(Inner Mongolia Power Research Institute Branch,Inner Mongolia Power(Group)Co.,Ltd.,Hohhot 010020,China;College of Mechanical and Electrical Engineering,Inner Mongolia Agricultural University,Hohhot 010018,China)

机构地区：[1]内蒙古电力(集团)有限责任公司内蒙古电力科学研究院分公司,呼和浩特010020 [2]内蒙古农业大学机电工程学院,呼和浩特010018

出　　处：《材料导报》2024年第17期247-253,共7页Materials Reports

基　　金：内蒙古电力科学研究院2022年自筹科技项目(2022-ZC-05);国家自然科学基金(52061037)。

摘　　要：火电机组用钢多为合金钢,金相组织复杂,且长期在高温高压环境下服役,存在组织老化等问题。金相检验人工识别易受主观因素影响,识别精度波动大,结果重复性差。利用金相检验图像,依据相关标准建立火电用钢金相组织数据集,并将其拆分为训练集、验证集和测试集。基于GoogLeNet Inception V3模型建立原始、迁移学习、微调迁移学习三个模型,以训练集训练模型,在验证集上泛化能力最佳的模型作为最终模型,测试集结合混淆矩阵对所建模型性能进行综合评价。三个模型对火电用钢金相组织的识别准确率分别为96.6%、93.0%和92.4%,均能够对复杂金相组织进行较为准确的识别。其中原始Inception V3模型性能最优,在测试集的准确率、精确度、灵敏度和特异度指标的均值分别为96.6%、96.6%、96.6%和99.2%。本研究为复杂金相组织智能识别提供了新方法。The steel used in thermal power units is mostly alloy steel,which have complex metallographic structures and have been used in high-temperature and high-pressure environments for a long time,resulting in structural aging and other phenomena.The manual recognition of metallographic organization is easily influenced by subjective factors of the testing personnel,with large fluctuations in recognition accuracy and poor repeatability of results.This study utilized metallographic organization images to establish a dataset of metallographic organization for thermal power steel based on relevant standards,and divided it into training,validation,and testing sets.Based on the GoogLeNet Inception V3 model,three models of original,transfer learning and fine tuning transfer learning were established,and the training set training model was used.The model with the best generalization ability on the verification set was taken as the final model.The test set was combined with the confusion matrix to comprehensively evaluate the performance of the built model.The results show that the recognition accuracy of the three models for the metallographic structure of steel used in thermal power is 96.6%,93.0%,and 92.4%,respectively,which can accurately identify complex metallographic structures.The original Inception V3 model has the best performance,with an average of 96.6%,96.6%,96.6%,and 99.2%for accuracy,positive predictive value,true positive rate,and true negative rate in the test set.This study provides a new method for intelligent recognition of complex metallographic structures.

关 键 词：金相组织 智能识别 深度学习 卷积神经网络 计算材料学 

分 类 号：TG142.15[一般工业技术—材料科学与工程]