Feature extraction and dimensionality reduction of arc sound under typical penetration status in metal inert gas welding
Feature extraction and dimensionality reduction of arc sound under typical penetration status in metal inert gas welding
复制标题
DOI：
10.3901/cjme.2012.02.293
复制
发表时间：
2012-02
影响因子：
4.2
通讯作者：
Lijun Liu;Hu Lan;Hongyan Zheng;Xiaoxia Jian
中科院分区：
文献类型：
--
作者：
Lijun Liu;Hu Lan;Hongyan Zheng;Xiaoxia Jian
众所周知，电弧声是监测和控制焊缝熔深状态的潜在和可用资源，这对于焊接过程质量控制非常重要，因此电弧声与焊接参数之间的关系受到了人们的关注。借助神经网络建立了一些将电弧声音与焊接参数相关联的非线性映射模型。然而，利用电弧声监测和诊断焊接过程的研究还处于起步阶段。采用自制的实时传感系统，对喷射过渡金属惰性气体(MIG)平缝焊在部分熔透、不稳定熔透、完全熔透和过度熔透等典型熔透状态下的电弧声进行了研究。采用小波去噪和短时加窗技术对电弧声进行预处理，提取表征焊缝熔深状态的时域、频域、倒谱域和几何域特征。随后，利用主主成分分析（PCA）的概念构建高维特征向量并成功融合特征级参数。最终将60维特征向量替换为8维向量的合成，实现了特征空间的压缩，为未来MIG焊接中借助电弧声对典型熔深状态的模式分类提供了技术支撑。
Arc sound is well known as the potential and available resource for monitoring and controlling of the weld penetration status, which is very important to the welding process quality control, so any attentions have been paid to the relationships between the arc sound and welding parameters. Some non-linear mapping models correlating the arc sound to welding parameters have been established with the help of neural networks. However, the research of utilizing arc sound to monitor and diagnose welding process is still in its infancy. A self-made real-time sensing system is applied to make a study of arc sound under typical penetration status, including partial penetration, unstable penetration, full penetration and excessive penetration, in metal inert-gas (MIG) flat tailored welding with spray transfer. Arc sound is pretreated by using wavelet de-noising and short-time windowing technologies, and its characteristics, characterizing weld penetration status, of time-domain, frequency-domain, cepstrum-domain and geometric-domain are extracted. Subsequently, high-dimensional eigenvector is constructed and feature-level parameters are successfully fused utilizing the concept of primary principal component analysis (PCA). Ultimately, 60-demensional eigenvector is replaced by the synthesis of 8-demensional vector, which achieves compression for feature space and provides technical supports for pattern classification of typical penetration status with the help of arc sound in MIG welding in the future.