基于深度学习的火电机组耐热钢显微组织识别

Recognition of Microstructure of Heat-resistant Steel in Thermal Power Unit Based on Deep Learning

火电机组所用的耐热钢是复杂的多元高合金钢,金相检验结果由人工判别,容易由于主观因素导致识别结果的不确定。在经典AlexNet模型的基础上构建新的卷积神经网络模型,利用火电机组耐热钢金相检验图像建立样本数据集,经训练与验证,模型对火电机组耐热钢显微组织识别准确率达到89%,比AlexNet模型提高了11%。进一步利用批归一化方法优化所建模型,识别准确率提升到94%,已接近人类对自然图像的识别精度。将深度学习技术应用于火电机组耐热钢显微组织识别,为火电机组金相检验提供了新方法。

The heat-resistant steel of thermal power unit is a complex multi-element high-alloy steel. The results of metallographic examination are judged manually, which is easy to lead to the uncertainty of identification results due to subjective factors. A new CNN(convolutional neural networks) model was built based on the classic AlexNet model and it was trained and tested on the sample dataset of metallographic images of heat resisting steel for thermal-electric generator. The results show that the new model has an 89% accuracy in classification and recognition of steel microstructures, which has improved by 11% compared with AlexNet. Furthermore, the model was optimized with Batch Normalization algorithm and the accuracy could reach 94%, which approximates the accuracy of human recognition for natural images. Deep learning technology is applied to identify the microstructure of heat-resistant steel for thermal power units, which can provide a new method for metallographic inspection of thermal power units.