梅尔频率倒谱耦合神经网络的焊接缺陷检测

Welding defect detection algorithm based on Mel-frequency cepstral coupling neural networks

当前焊接图像缺陷检测技术因依赖焊接几何特征缺陷,对微小缺陷中黑暗边缘的噪声较为敏感,导致其定位精度不佳,为此提出一种梅尔频率倒谱耦合神经网络特征匹配的焊接缺陷检测算法。利用DCT(discrete cosine transform)与Zigzag机制,将焊接图像排列成1D信号数组;将1D信号分割为多个帧,构造窗口函数,增强相邻帧之间的连续性,引入倒谱技术,查询1D信号的稳定特性,提取其梅尔频率倒谱系数;定义两个正交多项式,建立多项式系数计算模型,提取多项式系数。基于神经网络训练,对提取特征与数据库特征进行匹配,完成缺陷检测。实验结果表明,与当前焊接缺陷检测技术相比,该算法的定位精度高达90%,鲁棒性更强,不受噪声影响。

To solve the problem of low location accuracy of weld defects caused by depending on welding geometrical characteris-tic flaw and sensitivity to the noise of the dark edge in the micro-defect of current welding defect detection algorithm,the welding defect detection algorithm based on Mel-frequency cepstral coupling neural networks feature matching was proposed.The welding image was transformed into 1 D signal array using the DCT and Zigzag mechanism.The window function was constructed to enhance the continuity between the adjacent frames by dividing the 1D signal into multiple frames,and also the Mel-frequency coefficients were extracted by introducing cepstral technology to query the stability of 1 D signal;the polynomial coefficient model was established by defining two orthogonal polynomials to extract polynomial coefficients.The detection of the defects was accomplished based on the training of neural network to match feature extraction feature and database feature.Experimental results show that,compared with the current welding defect detection technology,this algorithm has higher positioning accuracy that reaches 98.62%,and its robustness is stronger without affected by the noise.