基于多态数据驱动的铝合金脉冲GTAW焊接缺陷在线预测及定量评估

On-line weld defect detection is one of the key problems for aerospace component in Welding Intelligent Manufacturing Engineering(WIME). In this project, a new method of on-line weld defect prediction and quantitative evaluation for aluminum alloy pulse Gas Tungsten Arc Welding (GTAW) in aerospace industry will be explored based on multi-state data-driven technology derived from the welding arc spectrum, sound, arc voltage and image information. .Main contents include: weld defect feature extraction, evaluation and feature selection of the multisource polymorphism information. Building the complete knowledge database of polymorphism characteristics for aluminum alloy TIG welding defects and preliminarily exploring the dynamic mechanism and evolution rules of typical welding defects. Then, taking advantage of the characteristics of overlapping and complementary for multi-source sensors system, the characterization for defects status with deep fusion of multi-source information will be researched. The feature level-based fusion model for aluminum alloy welding defects prediction will be built using the improved support vector machine (SVM). On the basis of the local decision of polymorphic sensors, the fusion method in decision-making level will be proposed with the application of Dempster-Shafer(DS) evidence theory and fuzzy integral method for on-line aluminum alloy welding defects identification. The preliminary standard system will be constructed for on-line quantitative weld defect evaluation. The universal method of on-line welding defect quantitative evaluation for aluminum alloy pulse GTAW will be established based on the multi-information driven fusion model. The proposed methodology has the ability to evaluate and predict the weld defects from the multi-dimensional and multi-perspective view. The research achievement will promote intelligent development of aerospace lightweight welding manufacturing technology.

焊接缺陷在线检测是航空航天构件智能化焊接制造工程所关注的重点问题之一。本项目尝试通过焊接过程电弧光谱、声音、电压及图像多源多态数据为驱动，探索形成一种针对航空航天用铝合金脉冲GTAW焊接缺陷的在线预测及定量评估新方法。主要内容包括：研究多源多态海量信息的缺陷特征提取、评价及选择方法，构建完备焊接缺陷多态特征知识库，初步探索典型焊接缺陷产生的动态机理和演化规律；其次，利用多源传感系统的交叠互补特性，研究多源信息深度融合下的缺陷状态表征方法，构建基于改进支持向量机的缺陷预测特征层融合模型；以多态传感的局部判识结果为基础，研究应用DS证据理论和模糊积分方法形成焊接缺陷在线预测识别的决策层融合方法；初步构建焊接缺陷在线定量评估标准体系，形成具有普适性的多信息融合模型驱动下的铝合金脉冲GTAW焊接缺陷在线定量评估方法，实现对焊接缺陷多维多视角评判和预测，推动航空航天等轻量化焊接制造技术的智能化发展。

针对国防装备焊接制造效率低及无损检测周期长，开展机器人焊接过程缺陷在线检测研究。搭建了电弧光谱、声音、图像、电流及电压多传感信号同步获取及机器人焊接实验系统；研究焊接缺陷多态信息的特征提取-评价-选择；提出了多模型驱动下的机器人焊接缺陷在线评估，包括电弧光谱深度信任网络DBN多缺陷识别模型，随机森林RF缺陷识别模；电弧声音SVM-GSCV，时频图像TF-CNN熔透状态识别模型；熔池图像CNN焊透状态识别模型；研究了焊接气孔缺陷的定量表征及在线检测。.以航空航天用铝合金薄板脉冲GTAW为研究对象，提取了PCA主分量系数SOI-MgI，SOI-AlI及SOI-MgI的面积特征，无需进行脉冲干扰去除，可克服峰值像素点漂移问题。提出了OOB-Gini光谱特征综合评价指标，分析了光谱重要特征的规律，发现氩谱线比金属谱更利于识别建模； Fe I(407.84 nm)谱线与铝合金脉冲GTAW过程送丝状态具有唯一的强相关性；与MgI(383.83nm)及AlI(369.15nm)有负相关关系；发现PCA主成分有效挖掘出了电弧声特征不同的响应规律，电弧声音信号在频域冗余度为68% ~95%。.利用不同焊接电流和工件形状的焊缝对SVM-GSCV熔透识别模型算法对进行了验证，熔透识别精度由81.52%到98.46%。电弧声音时频图像TF-CNN熔透识别模型自动提取特征与机理分析一致，提高了平均识别率98.2%；基于熔池图像的CNN焊透状态识别模型充分利用了弧光干扰，自动提取特征为熔池不同区域，不同弧光强度以不同形式的组合，呈现互补特性，识别精度可达99.7%, 熔透识别率进一步得到提高。基于6维显著特征的随机森林缺陷识别模型相对于52维全特征模型，计算速度提高了72.52%；利用SEM,EDS多种微观表征手段，发现铝合金脉冲GTAW焊接中会产生氢气孔及镁蒸气气孔。在熔池动态变化过程中，镁元素高度参与了气孔的产生及长大，镁蒸气气孔的内部镁平均含量是外缘的两倍。建立了氢谱线PCA特征向量系数与内部气孔之间的强相关关系，所提出的氢谱线PCA 系数PCoefHS可有效地检测一定长度焊缝的气孔缺陷。研究可提高无损检测精度及效率，指导工艺优化，提高机器人焊接质量稳定性；缩短国防装备制造周期，提高综合国防实力。