基于DCGAN数据集增强的ECT系统流型识别技术改进  

Improvement of flow pattern identification technology for ECT system based on DCGAN dataset enhancement

作　　者：胡红利[1] 王茂杰 杨姝楠 杨海潮 卢家宇 HU Hongli;WANG Maojie;YANG Shunan;YANG Haichao;LU Jiayu(State Key Laboratory of Electrical Insulation and Power Equipment,Xi’an Jiaotong University,Xi’an 710049,China;Beijing Research Institute of Mechanical Equipment,Beijing 100854,China)

机构地区：[1]西安交通大学电力设备电气绝缘国家重点实验室,陕西西安710049 [2]北京机械设备研究所,北京100854

出　　处：《西北大学学报（自然科学版）》2023年第4期554-562,共9页Journal of Northwest University（Natural Science Edition）

基　　金：国家自然科学基金(52177009);陕西省科技攻关项目(2016GY001)。

摘　　要：基于电容层析成像(electrical capacitance tomography,ECT)系统的流型识别技术具有非侵入性和可视化的优点,被广泛应用于各种工业过程的动态测量中。为提高ECT系统的流型识别准确率,基于神经网络模型的识别算法被广泛应用,但ECT系统难以获得大量实测样本,从而导致识别准确率低、模型泛化性差。针对这一问题,该文采用了基于深度卷积生成对抗网络(DCGAN)的数据集扩增方法,利用经过预处理的LSSVM方法重建所得的图像,搭建基于DCGAN的流型图像生成模型进行实验,并利用GAN-train和GAN-test两个客观指标对生成图像进行评价,之后使用扩增后的数据集进行基于YOLOv3的流型识别。实验结果表明,利用DCGAN方法能够有效扩增原有数据集,生成图像准确度高、丰富性强,配合YOLOv3的目标检测算法,可以对ECT系统的流型识别准确率提升有较大程度的改进。The flow pattern recognition technology based on capacitive tomography(ECT)system is widely used in dynamic measurement of various industrial processes due to its non-invasive and visual advantages.To improve the flow pattern identification accuracy of ECT system,recognition algorithm based on neural network model is widely used.But the ECT system can’t get a lot of ECT system measured samples,resulting in low accuracy,low model generalization.To solve this problem,this paper adopted the dataset amplification method based on deep convolution generative adversarial network(DCGAN),and used the images reconstructed by the preprocessed LSSVM method to build a flow pattern image generation model based on DCGAN for experiments.Gan-train and GAN-test were used to evaluate the generated images.After that,the amplified dataset is used for flow pattern identification based on YOLOv3.The experimental results show that the DCGAN method can effectively amplify the original data set and generate images with high accuracy and richness.Together with the target detection algorithm of YOLOv3,the flow pattern recognition accuracy of ECT system can be improved to a great extent.

关 键 词：流型识别 数据集扩增 机器学习 深度卷积生成对抗网络 YOLOv3 

分 类 号：TP-391[自动化与计算机技术]