测控技术2025，Vol.44Issue(12)：13-21,29,10.DOI:10.19708/j.ckjs.2025.05.232
基于融合特征增强与高效重建机制的耐张线夹无监督异常检测方法
Unsupervised Anomaly Detection Method Based on Fusion Feature Enhancement and Efficient Reconstruction Mechanism for Strain Clamp
摘要
Abstract
An unsupervised anomaly detection method based on fusion feature enhancement and efficient recon-struction mechanism is proposed to address the imbalance of sample distribution and the difficulty in construc-ting abnormal samples in digital radiography(DR)images of strain clamps used in power transmission lines.The method combines a feature enhancement module based on adaptive histogram equalization and Laplace transform(AHELt)and a detection module using efficient memory feature reconstruction for anomaly detection(EMFR-AD).AHELt enhances local contrast and edge information of the image by introducing adaptive histo-gram equalization and Laplace transform,improving feature compatibility between DR images and unsupervised models.EMFR-AD integrates an encoder-decoder architecture with knowledge distillation to build a compact memory matrix that stores normal features and identifies anomalies by comparing input and reconstructed ima-ges.Experimental results show that AHELt significantly improves detection performance on a self-constructed DR dataset.EMFR-AD algorithm achieves 89.98％in receiver operating characteristic-area under curve(ROC-AUC)and a detection speed of approximately 26 f/s on this dataset.It also reaches an average detection accu-racy of 89.53％on the public MVTec anomaly detection(MVTec AD)dataset.关键词
耐张线夹/无监督/知识蒸馏/无损检测/异常检测Key words
strain clamps/unsupervised/knowledge distillation/nondestructive testing/anomaly detection分类
信息技术与安全科学引用本文复制引用
李鸿,郑皓亮,丁龙,贾智伟,李灵..基于融合特征增强与高效重建机制的耐张线夹无监督异常检测方法[J].测控技术,2025,44(12):13-21,29,10.基金项目
湖南省教育厅科学研究项目(23A0255,22B0329) （23A0255,22B0329）
国家自然科学基金青年基金项目(62103063) （62103063）
