基于轻量化网络和知识蒸馏的回转窑工况识别

回转窑烧成带图像蕴含丰富火焰信息，准确识别燃烧状态是回转窑优化控制的前提，基于卷积神经网络的方法可以快速识别回转窑工况，提高回转窑自动化水平，但存在网络规模大，所需计算资源多的问题，为此提出了一种基于轻量化网络和知识蒸馏的回转窑工况识别方法，在网络卷积层后引入协方差池化层改进教师模型和学生模型，以改进的轻量化网络MobilenetV2作为学生模型的主干网络，以改进的Resnet50作为教师模型的主干网络，通过构建混合蒸馏损失函数，将教师模型蕴含的丰富分类标签信息迁移到学生模型中，并将蒸馏训练得到的学生模型作为回转窑工况识别模型，以提高网络对高相似火焰图像的识别精度。实验结果表明，经蒸馏后的学生模型总体识别准确率相较于原始模型提高了3.33%,对测试集中3种工况的识别率分别达到了93%、99%、90%,准确率和模型规模均优于目前其他主流网络模型，满足实际生产中实时、低成本要求。

Working condition recognition based on lightweight network and knowledge distillation for rotary kilns

The firing zone images of rotary kilns contain rich flame information and accurate combustion state recognition are the premiseof optimal control for the rotary kilns. The working conditions can be quickly recognized, and the automation level of the rotary kilns canbe improved through the convolutional neural network-based methods, but there are problems with large network size and highcomputational resources required. Therefore, a working condition recognition method based on lightweight network and knowledgedistillation was proposed in this paper. The teacher model and the student model were improved by introducing the collaboratordifferential layer after the convolutional layer of the network. The improved lightweight network MobilenetV2 was used as the backbonenetwork of the student model while the improved Resnet50 was used as the backbone network of the teacher model. The rich classificationlabel information contained in the teacher model was transferred to the student model by constructing the mixed distillation loss functionand the student model obtained by distillation training was used as the working condition recognition model to improve the recognitionaccuracy of the high similar rotary kiln flame images. The experimental results show that the overall recognition accuracy of the improvedstudent model is increased by 3. 33% compared with the original model, and the recognition accuracy of the three working conditions inthe test set reaches 93%, 99%, 90% respectively. The accuracy and network size are better than those of other mainstream networks,and the requirements of real time and low cost in actual production process are met.