基于多信息融合的铝合金脉冲GTAW过程焊接缺陷特征提取研究
发布时间：2018-08-20 18:59
【摘要】：智能化焊接是智能制造领域中最重要的研究课题之一。而传感技术及其信息处理则是实现焊接过程智能化及自动化的关键要素。近年来,具有小型化、无接触式及大传输量等特点的传感技术更多地被应用到焊接过程及质量实时控制中,如电弧传感、视觉传感、声音传感、光谱传感等,这些传感利用不同信息源获取了与焊接质量有关的大规模信息,但同时也不可避免地带来了焊接过程的“大数据”。因此,如何去除其中的大量噪声和冗余信息,更重要的,如何挖掘有效信息并将其及时反馈与利用在焊接质量的实时监测中是亟需解决的关键问题。本文以铝合金脉冲GTAW过程为研究对象,以实时预测识别焊接动态过程中的典型缺陷为目标,基于光谱传感、声音传感、电压传感及视觉传感技术,对焊接缺陷的特征提取、评价、选择以及多信息融合的预测识别方法开展了深入研究。搭建了一套脉冲GTAW焊接试验系统及多信息采集平台,可以实现对焊接过程的自动控制,对焊接电弧光谱、声音、电弧电压及焊缝图像信息的自动采集与存储。借助于多源同步信息,分析了多种典型焊接缺陷的产生机理,以及不同信号在时域-频域-时频域所表现出的奇异性,提出了针对各类传感的信号处理及特征提取方法。提出了一种基于感兴趣的光谱辐射区间soi(spectrumofintrests)的光谱统计特征提取及评价方法。根据最大奇异性原则选择了8段soi,从中所提取的均方根r、方差d及峰度k三个统计特征参数有效表征了焊接电弧soi辐射的平均强度、奇异程度及形态特征;利用小波包coief4小波函数5层分解的信号重构法有效去除了特征脉冲干扰。进一步,基于所提出的snr对数特征评价准则,量化了特征参数对焊接缺陷的敏感度。其次,基于所选波长为656.28nm的hi谱线和641.63nm的ari谱线,先后提出了谱峰面积、谱峰强度以及谱线方差之比等多个光谱特征,利于fisher评价准则定量评价了各特征值对焊缝氢致气孔缺陷的敏感度,基于特征std阈值线实现了对焊缝氢气孔的在线快速监测。针对声音信号分别在时域、频域及时频域开发了相应的特征提取算法。首先,提出了一种基于感兴趣的声音局部信号lsoi(localsoundofintrests)阈值法统计特征提取算法,研究了lsoi统计特征与未焊透及局部下榻缺陷的相关性;其次,提出了一种基于声音信号功率谱密度的频域分段注意sfsa(soundfrequencysegmentattetion)的统计特征提取方法,根据不同的注意机制,对weltch功率谱密度频域区间进行了分段及统计特征提取,分析了正常熔透及未焊透缺陷与声音psd频率的相关性。提出了一种声音小波包相对能量的特征提取及评价算法,根据所选db3小波基函数及3层分解方式计算得到了表征不同频域信号相对能量的特征集合e(j),进一步提出了最大类间标准差maximumstandarddeviationbetweenclass(msdbc)的特征评价准则,定量评价了e(j)对对未焊透、正常及焊漏焊缝三种不同的熔透状态的可分性,有效剔除了冗余特征。借助于小波包的多域交叉解析能力,发现7.5~10khz是一段非常重要的频率区间,其对应的时域、频域及时频域信号特征均对未焊透及焊漏缺陷表现出高度的相关性与敏感度。最后基于视觉注意机制提出的图像特征参数roi-1-countrate3、roi-2-countrate3及roi-3-countratio实现了对多种焊缝缺陷(焊漏、过熔透及表面氧化的同步检测。为了挖掘焊接过程“大数据”中隐藏的有效信息,选择有利于学习模型的最优特征组合,提出了一种数据驱动下的混合filter筛选器与wrapper封装器的hybridimprovedfisherfilterandsvm-cvwrapper(hifscw)特征选择器。首先,提出了一种自适应权重投票制改进fisher法(awvifc)的特征评价准则,作为特征筛选器。其中,根据统计的特征投票率,实现了特征样本权重的自适应更新,改进后的fisher准则保护了某些具有较小票数而较大fisher值的特征,实现了对特征的预筛选。其次,以支持向量机(svm)作为分类算法,结合10-fold交叉验证和网格搜索法参数寻优,构建了作为封装器的svm-cv分类模型。最后根据所得的分类准确率曲线定义了不充足特征子集区间、互补型特征区间、最优特征子集区间及冗余特征子集区间。构建了基于特征层融合的电压-声音-光谱svm-cv熔透状态预测识别模型,利用msdbc评价准则筛选得到的特征空间大大简化了分类融合模型,电压信息的融入弥补了声音信息特征识别未焊透与焊漏的不足。最终在成功预测缺陷的基础上,实现了对未焊透与焊漏缺陷的精确识别。缺陷的识别准确率从单一传感模型的74.19%提高到了多传感融合模型的94.31%。采用焊前打孔预埋氢化物的方式实现了对焊缝气孔、塌陷及氧化夹渣缺陷的定位可控制造,研究了缺陷产生机理及其对应信号特征的奇异性。利用hifscw特征选择器一方面实现了最佳特征组合的选择,另一方面利用其封装器中的svm-cv分类模型实现了对单一及耦合缺陷的预测和识别,在最佳特征空间区间内,该模型的分类识别准确率可达94.72%。与单一传感模型相比,融合模型的缺陷识别准确率有了较大提高,具有较高鲁棒性及稳定性。
[Abstract]:Intelligent welding is one of the most important research topics in the field of intelligent manufacturing. Sensing technology and its information processing are the key elements to realize the intelligent and automatic welding process. For example, arc sensing, visual sensing, sound sensing, spectral sensing and so on, these sensors use different information sources to obtain large-scale information related to welding quality, but also inevitably bring about the welding process "big data". Therefore, how to remove a large number of noise and redundant information, more important, how to mine effective trust. In this paper, the pulsed GTAW process of aluminum alloy is studied, and the typical defects in the dynamic process of welding are predicted and identified in real time. Based on spectrum sensing, sound sensing, voltage sensing and visual sensing technology, the welding defects are detected. A set of pulsed GTAW welding test system and multi-information acquisition platform were built to realize automatic control of welding process, automatic acquisition and storage of welding arc spectrum, sound, arc voltage and weld image information. Based on multi-source synchronization information, the generation mechanism of several typical welding defects and the singularity of different signals in time-frequency-time domain are analyzed. A method of signal processing and feature extraction for various sensors is proposed. A spectral statistical feature based on spectral radiation interval SOI (spectral of intrests) of interest is proposed. According to the principle of maximum singularity, eight SOI segments are selected, from which three statistical characteristic parameters, root mean square r, variance D and kurtosis k, are extracted to effectively characterize the average intensity, singularity and morphological characteristics of welding arc SOI radiation; signal reconstruction method based on wavelet packet coief 4 wavelet function 5-level decomposition is used to effectively remove the characteristic pulse. Furthermore, based on the proposed SNR logarithmic characteristic evaluation criterion, the sensitivity of characteristic parameters to weld defects is quantified. Secondly, based on the selected wavelength of 656.28 nm hi line and 641.63 nm ARI line, the spectral peak area, spectral peak intensity and the ratio of spectral line variance are proposed successively, which is beneficial to the determination of Fisher evaluation criterion. The sensitivity of each eigenvalue to weld hydrogen-induced porosity is evaluated quantitatively, and the on-line rapid monitoring of weld hydrogen-induced porosity is realized based on the characteristic STD threshold line. Corresponding feature extraction algorithms are developed for acoustic signals in time domain, frequency domain and time-frequency domain respectively. Firstly, a locally sound local signal LSOI (locally sound local signal of interest) is proposed. Ofintrests thresholding statistical feature extraction algorithm, studied the correlation between LSOI statistical characteristics and underpenetration and local defects; secondly, proposed a frequency domain segmentation attention sfsa (sound frequency segmentation attentions) statistical feature extraction method based on the power spectral density of sound signal, according to different attention mechanisms, the well power was extracted. The frequency domain of spectral density is segmented and extracted, and the correlation between normal penetration and non-penetration defects and PSD frequency is analyzed. A feature extraction and evaluation algorithm of relative energy of acoustic wavelet packet is proposed. The relative energy of signals in different frequency domains is calculated according to the selected db3 wavelet basis function and three-level decomposition method. The feature set E (j) of the quantity is further proposed, and the feature evaluation criterion of maximum standard deviation between classes betweenclass (msdbc) is proposed. The separability of E (j) for three different penetration states, i.e. impermeable, normal and leaky welds, is quantitatively evaluated, and the redundant features are effectively eliminated. It is found that 7.5-10 kHz is a very important frequency range, and its corresponding time-domain, frequency-domain and time-frequency domain signal characteristics show a high degree of correlation and sensitivity to impermeability and weld leakage defects. Finally, the image feature parameters roi-1-countrate3, roi-2-countrate3 and roi-3-countration based on visual attention mechanism are proposed to achieve a variety of features. Synchronous detection of weld defects (leakage, overpenetration and surface oxidation). In order to mine the effective information hidden in the "big data" of the welding process and select the optimal combination of features conducive to the learning model, a data-driven hybrid filter and wrapper packer hybridized Fisher filter dsvm-cvwrapper (hifscw) were proposed. Firstly, an adaptive weighted voting improved Fisher method (awvifc) is proposed as a feature filter. According to the statistical voting rate, the weights of feature samples are updated adaptively. The improved Fisher criterion protects some features with small votes but large Fisher values. Secondly, support vector machine (svm) is used as classification algorithm, and 10-fold cross-validation and grid search method are combined to construct svm-cv classification model as encapsulator. Finally, insufficient feature subset interval, complementary feature interval and optimal feature element are defined according to the classification accuracy curve. The voltage-sound-spectrum svm-cv penetration state prediction and recognition model based on feature layer fusion is constructed. The classification and fusion model is greatly simplified by using the feature space screened by the msdbc evaluation criterion. The integration of voltage information makes up for the shortcomings of sound information feature recognition such as incomplete penetration and welding leakage. Based on the successful prediction of defects, the accurate identification of impermeable and leaky defects was realized. The accuracy of defect identification was improved from 74.19% of single sensor model to 94.31% of multi-sensor fusion model. The mechanism of defect generation and the singularity of corresponding signal features are studied. On the one hand, hifscw feature selector is used to select the best feature combination, on the other hand, the svm-cv classification model in its packer is used to predict and recognize single and coupled defects. In the optimal feature space interval, the model is classified and recognized. Compared with the single sensor model, the fusion model has better robustness and stability.
【学位授予单位】：上海交通大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TG441.7
，
本文编号：2194678
[Abstract]:Intelligent welding is one of the most important research topics in the field of intelligent manufacturing. Sensing technology and its information processing are the key elements to realize the intelligent and automatic welding process. For example, arc sensing, visual sensing, sound sensing, spectral sensing and so on, these sensors use different information sources to obtain large-scale information related to welding quality, but also inevitably bring about the welding process "big data". Therefore, how to remove a large number of noise and redundant information, more important, how to mine effective trust. In this paper, the pulsed GTAW process of aluminum alloy is studied, and the typical defects in the dynamic process of welding are predicted and identified in real time. Based on spectrum sensing, sound sensing, voltage sensing and visual sensing technology, the welding defects are detected. A set of pulsed GTAW welding test system and multi-information acquisition platform were built to realize automatic control of welding process, automatic acquisition and storage of welding arc spectrum, sound, arc voltage and weld image information. Based on multi-source synchronization information, the generation mechanism of several typical welding defects and the singularity of different signals in time-frequency-time domain are analyzed. A method of signal processing and feature extraction for various sensors is proposed. A spectral statistical feature based on spectral radiation interval SOI (spectral of intrests) of interest is proposed. According to the principle of maximum singularity, eight SOI segments are selected, from which three statistical characteristic parameters, root mean square r, variance D and kurtosis k, are extracted to effectively characterize the average intensity, singularity and morphological characteristics of welding arc SOI radiation; signal reconstruction method based on wavelet packet coief 4 wavelet function 5-level decomposition is used to effectively remove the characteristic pulse. Furthermore, based on the proposed SNR logarithmic characteristic evaluation criterion, the sensitivity of characteristic parameters to weld defects is quantified. Secondly, based on the selected wavelength of 656.28 nm hi line and 641.63 nm ARI line, the spectral peak area, spectral peak intensity and the ratio of spectral line variance are proposed successively, which is beneficial to the determination of Fisher evaluation criterion. The sensitivity of each eigenvalue to weld hydrogen-induced porosity is evaluated quantitatively, and the on-line rapid monitoring of weld hydrogen-induced porosity is realized based on the characteristic STD threshold line. Corresponding feature extraction algorithms are developed for acoustic signals in time domain, frequency domain and time-frequency domain respectively. Firstly, a locally sound local signal LSOI (locally sound local signal of interest) is proposed. Ofintrests thresholding statistical feature extraction algorithm, studied the correlation between LSOI statistical characteristics and underpenetration and local defects; secondly, proposed a frequency domain segmentation attention sfsa (sound frequency segmentation attentions) statistical feature extraction method based on the power spectral density of sound signal, according to different attention mechanisms, the well power was extracted. The frequency domain of spectral density is segmented and extracted, and the correlation between normal penetration and non-penetration defects and PSD frequency is analyzed. A feature extraction and evaluation algorithm of relative energy of acoustic wavelet packet is proposed. The relative energy of signals in different frequency domains is calculated according to the selected db3 wavelet basis function and three-level decomposition method. The feature set E (j) of the quantity is further proposed, and the feature evaluation criterion of maximum standard deviation between classes betweenclass (msdbc) is proposed. The separability of E (j) for three different penetration states, i.e. impermeable, normal and leaky welds, is quantitatively evaluated, and the redundant features are effectively eliminated. It is found that 7.5-10 kHz is a very important frequency range, and its corresponding time-domain, frequency-domain and time-frequency domain signal characteristics show a high degree of correlation and sensitivity to impermeability and weld leakage defects. Finally, the image feature parameters roi-1-countrate3, roi-2-countrate3 and roi-3-countration based on visual attention mechanism are proposed to achieve a variety of features. Synchronous detection of weld defects (leakage, overpenetration and surface oxidation). In order to mine the effective information hidden in the "big data" of the welding process and select the optimal combination of features conducive to the learning model, a data-driven hybrid filter and wrapper packer hybridized Fisher filter dsvm-cvwrapper (hifscw) were proposed. Firstly, an adaptive weighted voting improved Fisher method (awvifc) is proposed as a feature filter. According to the statistical voting rate, the weights of feature samples are updated adaptively. The improved Fisher criterion protects some features with small votes but large Fisher values. Secondly, support vector machine (svm) is used as classification algorithm, and 10-fold cross-validation and grid search method are combined to construct svm-cv classification model as encapsulator. Finally, insufficient feature subset interval, complementary feature interval and optimal feature element are defined according to the classification accuracy curve. The voltage-sound-spectrum svm-cv penetration state prediction and recognition model based on feature layer fusion is constructed. The classification and fusion model is greatly simplified by using the feature space screened by the msdbc evaluation criterion. The integration of voltage information makes up for the shortcomings of sound information feature recognition such as incomplete penetration and welding leakage. Based on the successful prediction of defects, the accurate identification of impermeable and leaky defects was realized. The accuracy of defect identification was improved from 74.19% of single sensor model to 94.31% of multi-sensor fusion model. The mechanism of defect generation and the singularity of corresponding signal features are studied. On the one hand, hifscw feature selector is used to select the best feature combination, on the other hand, the svm-cv classification model in its packer is used to predict and recognize single and coupled defects. In the optimal feature space interval, the model is classified and recognized. Compared with the single sensor model, the fusion model has better robustness and stability.
【学位授予单位】：上海交通大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TG441.7
，
本文编号：2194678
本文链接：https://www.wllwen.com/kejilunwen/jinshugongy/2194678.html
教材专著
