基于Ultra-ViT的金属缺陷检测模型  

Detection Model for Defect Signal Spectrograms Based on Ultra-ViT

作　　者：唐东林[1] 李恒辉 丁超 赵云亮 TANG Dong-lin;LI Heng-hui;DING Chao;ZHAO Yun-liang(School of Mechanical and Electrical Engineering,Southwest Petroleum University,Chengdu 610500;School of Intelligent Manufacturing,Chengdu Technological University,Chengdu 611730)

机构地区：[1]西南石油大学机电工程学院,四川成都610500 [2]成都工业学院智能制造学院,四川成都611730

出　　处：《制造业自动化》2025年第3期134-141,共8页Manufacturing Automation

基　　金：四川省市场监督管理局科技计划(SCSJZ2022007);中国职业安全健康协会创新创业项目(CXCY-2021-22);成都市技术创新研发项目(2018-YF05-00201-GX)。

摘　　要：为了避免金属设备在超声缺陷检测过程中,人为主观因素及单一时频域分析对检测精度造成影响,提出利用时频分析获取超声信号谱图信息,结合视觉Transformer模型(Ultra-ViT)实现对超声A扫描缺陷检测信号识别分类。首先利用超声A扫描获取人工制作的金属缺陷板的缺陷信号,制作缺陷信号数据集,提取缺陷信号谱图特征。以谱图特征作为训练样本,提取谱图图像块并对其进行位置编码,利用Ultra-ViT检测模型进行训练,实现对金属设备的安全检测。实验表明,使用Ultra-ViT模型对超声谱图特征的识别准确率达97.73%,与经典CNN模型相比,模型能够以较高的检测速度实现更高的检测精度,在超声谱图检测上具有优异性。To avoid the influence of human subjective factors and a single time-domain and frequency-domain analysis on the detection accuracy of metal equipment in the process of ultrasonic defect detection,this paper proposes to use time-frequency analysis to obtain ultrasonic signal spectrum information,and combine it with visual Transformer model(Ultra-ViT)to realize the recognition and classification of ultrasonic A-scan defect detection signals.Firstly,ultrasonic A-scan is used to obtain the defect signal of the artificial metal defect plate,make the defect signal data set and extract the spectrum characteristics of the defect signal.Using spectral features as training samples,spectral image blocks are extracted,their positions are encoded,and the Ultra-ViT detection model is utilized for training to realize the safety detection of metal equipment.The Experiments show that the recognition accuracy of ultrasonic spectrogram features using Ultra-ViT model is 97.73%.Compared with the classical CNN model,this model can achieve higher detection accuracy with higher detection speed.It shows the superiority of this method in ultrasonic spectrum detection.

关 键 词：金属缺陷 梅尔谱图 TRANSFORMER 数据增强 

分 类 号：X947[环境科学与工程—安全科学]