基于PSPNetM+网络的水轮机转轮叶片缺陷检测  

Defect Detection of Hydraulic Machinery RunnerBlade Based on PSPNetM+ Network

作　　者：李涛 刘成 田塘 胡桂川[1] 侯文赛 蒲小霞 LI Tao;LIU Cheng;TIAN Tang;HU Guichuan;HOU Wensai;PU Xiaoxia(School of Mechanical and Power Engineering,Chongqing University of Science and Technology,Chongqing 401331,China;School of Safety Engineering,Chongqing University of Science and Technology,Chongqing 401331,China)

机构地区：[1]重庆科技学院机械与动力工程学院,重庆401331 [2]重庆科技学院安全工程学院,重庆401331

出　　处：《重庆科技学院学报（自然科学版）》2022年第6期93-96,共4页Journal of Chongqing University of Science and Technology：Natural Sciences Edition

基　　金：重庆市人工智能技术创新重大主题专项重点研发项目“基于全生命周期数据的超高水头冲击式发电机组智能预警与诊断系统”(CSTC2017RGZN-ZDYFX0026);重庆科技学院研究生科技创新计划项目“基于轻量化网络模型的水电机组转轮缺陷检测模型研究”(YKJCX2120713),“基于协同深度学习的水电机组导轴承故障诊断及性能预测模型研究”(YKJCX2120727)。

摘　　要：与传统的水轮机转轮叶片缺陷检测方法不同,提出利用基于语义分割的PSPNetM+模型对叶片缺陷进行识别的方法。首先,以MobineNetV2作为主干提取网络;然后,在特征图融合之前降低通道数,以降低网络参数量,同时优化损失函数;最后,采用Labelme自生成转轮叶片缺陷样本集进行验证。实验结果表明,改进后的PSPNetM+网络参数量仅为14.78 MiB,平均像素准确率增加了3.22%,平均交并比提高了2.84%,GPU每秒可以处理41张图片,平均精确度提高了5.51%。Different from the traditional defect detection method for hydraulic turbine runner blades,a method based on semantic segmentation is proposed to identify blade defects.Firstly,MobineNetV2 is used as the trunk extraction network.Secondly,the number of channels is reduced before the feature graph fusion to reduce the number of network parameters and optimize the loss function.Finally,Labelme self-generated runner blade defect sample set is used for verification.Experimental results show that the weight of the improved PSPNetM+network parameters accounts for only 14.78 MiB,with an increase of 3.22%in mPA and 2.84%in mIOU.The GPU can process 41 images per second,with an average accuracy improvement of 5.51%.

关 键 词：水轮机转轮叶片 缺陷检测 语义分割 PSPNetM+ 

分 类 号：TV734[水利工程—水利水电工程]