基于融合特征的CNN-Transformer墙体瓷砖粘贴空鼓检测算法

Fused feature based CNN⁃Transformer algorithm for empty drum detection of pasting tile in exterior wall

建筑墙体瓷砖粘贴空鼓的敲击检查方法是目前无损检测中应用最多的检测方法。为实现对复杂敲击位置下的识别和智能化检测,使用敲击法获取空鼓声音信号,提取连续小波变换(CWT)时频图和梅尔倒谱系数(MFCC)等时序特征。设计轻量化注意力CNN-Transformer双分支网络GATRNet,提出一种基于门控机制的特征融合模块,对CWT时频图和融合时序特征分别提取深度特征并进行融合。试验结果表明,所提方法测试精度可达99.10%,特征融合模块能够充分融合多种特征;相较于机器学习和神经网络识别方法,GATRNet在面对复杂敲击位置的声音时,多样性评价指标明显较优异。

At present,the percussion inspection method for empty drum detection of the pasting tile in exterior wall is the most widely used non-destructive testing method.In order to recognition and intelligent detection under complex tapping positions,the tapping method is used to obtain empty drum sound signals,extract time-frequency maps of continuous wavelet transform(CWT)and time-series features such as Mel frequency cepstral coefficients(MFCC).A lightweight attention CNN-Transformer two branch network GATRNet is designed,a feature fusion module based on gating mechanism is proposed,and deep features are extracted from CWT time-frequency maps and fused temporal features separately,and perform fusion.The testing results show that the testing accuracy of the proposed method can reach 99.10%,and the feature fusion module can fully integrate multiple features.In comparison with the machine learning and neural network recognition methods,GATRNet has significantly better diversity evaluation indicators when facing complex tapping positions of sound.