基于深度迁移学习的池沸腾瞬时状态快速识别方法  

Fast recognition of instantaneous states of pool boiling based on deep transfer learning

作　　者：张轩[1] 洪敏[1] 古江杭 莫冬传 衡益 ZHANG Xuan;HONG Min;GU JiangHang;MO DongChuan;HENG Yi(School of Computer Science and Engineering,Sun Yat-sen University,Guangzhou 510006,China;National Supercomputing Center in Guangzhou(NSCC-GZ),Sun Yat-sen University,Guangzhou 510006,China;Guangdong Province Key Laboratory of Computational Science,Sun Yat-sen University,Guangzhou 510006,China;School of Materials,Sun Yat-sen University,Shenzhen 518107,China;Guangdong Engineering Technology Research Centre for Advanced Thermal Control Material and System Integration(ATCMSI),Guangzhou 510275,China)

机构地区：[1]中山大学计算机学院,广州510006 [2]中山大学国家超级计算广州中心,广州510006 [3]中山大学广东省计算科学重点实验室,广州510006 [4]中山大学材料学院,深圳518107 [5]广东省先进热控材料及系统集成工程技术研究中心,广州510275

出　　处：《中国科学：技术科学》2024年第6期1057-1068,共12页Scientia Sinica(Technologica)

基　　金：广东省重点领域研发计划(编号:2021B0101190003);广东省自然科学基金面上项目(编号:2022A1515011514)资助。

摘　　要：沸腾过程高效传热的同时也面临着沸腾危机,而沸腾状态的瞬时识别与预测是维持“高功率密度电子器件”高效散热与延长生命周期的关键问题之一.得益于人工智能的快速发展,智能监测与快速识别技术逐渐被用于开发高效的沸腾状态检测器.传统机器学习方法作为纯数据驱动方法难以摆脱对数据的过度依赖,而在实际沸腾应用场景下,模型会因为工况改变与数据稀缺面临泛化瓶颈.基于此,本文提出了一种基于深度迁移学习的池沸腾瞬时状态快速识别方法.首先基于卷积神经网络构建状态识别器OrigCNN,并利用自行开展的池沸腾实验获取沸腾瞬时状态数据集Dataset A训练源模型,测试准确率为100%;针对源模型的泛化瓶颈,本文以自有数据集Dataset A为源域,公开数据集Dataset B和Dataset C为目标域,在源模型OrigCNN中引入迁移学习技术构建深度迁移模型TLCNN,并分别使用目标域数据集的10%,5%,2.5%和1%构建小样本数据集用于迁移训练.测试结果表明,迁移训练样本数据量越多,TLCNN的测试准确率越高,以使用5%Dataset B(132张)进行迁移训练为例,TLCNN测试准确率达99.83%,临界状态检测假阴性率(false negative rate,FNR)为0.38%,证明该模型在实际场景切换时迁移应用的有效性与可靠性.此外,本文提出的深度迁移学习模型的识别效率在单机设备上即可达到毫秒级,对开发实时沸腾瞬态状态识别器及相应的数字孪生工具软件有重要的意义.The boiling process is an efficient heat transfer method that creates a boiling crisis.Instantaneous recognition and prediction of the boiling state are necessary to maintain efficient heat dissipation and prolong the life cycle of high-power-density electronic devices.Benefiting from the rapid advancements in artificial intelligence,intelligent monitoring and rapid recognition technologies have been gradually applied to develop efficient boiling-state detectors.Traditional machine learning methods,such as pure data-driven methods,encounter difficulties in avoiding their excessive dependence on data.However,in practice,models frequently experience generalization bottlenecks due to changes in working conditions and data scarcity.To resolve these issues,an instantaneous boilingstate recognition method based on deep transfer learning is proposed in this paper.First,a state recognizer OrigCNN is constructed based on the convolutional neural network and subsequently trained using Dataset A,obtained from our pool boiling experiments.The accuracy of the test is up to 100%.Considering the generalization bottleneck of the source model OrigCNN,the transfer learning technology is applied to further improve OrigCNN by constructing a deep transfer model TLCNN,with“Dataset A”as the source domain and the publicly available“Dataset B”and“Dataset C”as the target domain.10%,5%,2.5%,and 1%of Dataset B and Dataset C are used to construct small sample datasets for transfer training.The test results show that the amount of sample data for transfer learning positively correlates with the prediction accuracy of TLCNN.The TLCNN test accuracy reached 99.83%when 5%Dataset B(132 photos)was used for transfer training,and the false negative rate of critical state detection was 0.38%,demonstrating the effectiveness and reliability of the TLCNN models in actual scene switching.Furthermore,the deep transfer learning method TLCNN proposed in this research exhibits high identification efficiency on the millisecond scale using a single

关 键 词：沸腾危机 状态识别 实时检测 卷积神经网络 迁移学习 

分 类 号：TK124[动力工程及工程热物理—工程热物理] TP18[动力工程及工程热物理—热能工程]