基于原型学习的裂纹方向细粒度分类方法

打开文本图片集

关键词：原型学习；超声反射波；细粒度分类；裂纹方向；自适应类特征学习；小样本学习方法中图分类号：TN911.73-34；TP391.4 文献标识码：A 文章编号：1004-373X（2026）09-0114-08

Prototype learning based fine-grained classification method for crack direction

ShiWeiyan，LuYingchun，LuYongfeng，ZhaoJunda （College ofElectricalandPower Engineering，Taiyuan Universityof Technology，Taiyuan O3O024，China）

Abstract：The prototype networksconsider thesamplesasequallimportantand ignore theirfine-grained class features， whichresultsinlowaccuracyinsamplerecognition，sothispaperproposesaprototypelearningbasedfine-grainedclasification method forcrack direction.This methodconsistsoftwoparalelsubnets.Oneis the backbone network，which focuses on the subtlelocalclassfeaturesofrackdirectioanddynamicallgeneratefinegrainedprototypfeaturesduringthetrainingprocess; theotherisanadaptiveclassfeaturelearningnetwork，whichcanadaptivelyconvergeprototypefeaturestothecenterofnormal samplesandoptimizetheoverallrepresentationperformanceoftheprototypeforcategoriesbyfusionoperationswiththe backbonenetwork.Thismethoddynamicalloptimizestheprototypelearningprocessinbothlocalandglobalperspectives， changing theprototypenetwork'sinherentmechanismof takingtheaverage valuedirectlytogenerate prototypes.Theexperiments show thatthe average classification accuracyrateof the method for crack direction onthe field test dataset is 91.71% ，which indicates the method can be used for evaluating crack direction inpractice.

Keywords：prototypelearning;ultrasonicreflectionwave;fine-grainedclasification;crack direction;adaptiveclass feature learning;small sample learning method

0 引言

裂纹是金属结构中常见的危害性最大的缺陷，裂纹会沿着开裂方向扩展，造成结构的断裂。（剩余11564字）