强弧光环境下焊缝识别模型及实现方法研究

发布时间：2018-05-10 11:07

本文选题：电弧焊接 + 焊缝检测　；参考：《广东工业大学》2015年博士论文

【摘要】：现代焊接自动化水平日益发展和计算机视觉技术的进步以及对效率和质量的高要求,基于视觉传感的焊缝图像处理和自动识别技术已成了焊缝自动跟踪的一个重要研究方向。由于在实际焊接过程中,存在强烈的弧光辐射、高温、工件热变形、图像传感误差等噪声的干扰,难以获得清晰稳定的焊缝区图像,对焊缝区图像的摄取,需要通过特定光谱的滤光系统,由图像采集卡采集图像并存储于工控机,再经图像预处理、图像识别等过程提取出焊缝的准确位置。焊缝跟踪技术是自动电弧焊接的一个重要研究领域,实现精确的焊缝跟踪对于提高焊接质量具有非常重要的作用。而要实现精确的焊缝跟踪,焊缝偏差(即焊缝中心与电弧的偏差)检测技术是一个关键。本研究设计安装了一套用于实验的焊缝跟踪系统装置,采用工控机控制,无辅助光源的视觉检测系统,来完成焊缝的自动跟踪控制。视觉传感式焊缝跟踪的一般方法是利用CCD (Charge Couple Device)传感器实时获取焊接图像,再利用图像处理技术提取熔池中心与熔池前端的焊缝位置信息,两者之间的差值就反映了当前时刻电弧与焊缝的偏差情况。对包括图像滤波、图像增强、图像二值化、边缘检测、焊缝位置的提取等基于图像处理的焊缝处理技术过程和方法用数学方法进行分析和研究。通过对焊缝位置中心提取方法的研究对比,着重研究了基于图像质心的图像处理方法,并基于图像质心研究了焊缝位置偏差信息提取的新技术。不同于传统的通过图像处理技术来直接获取焊缝偏差信息,而是选取特定熔池图像处理区域,并将熔池图像质心作为分析焊缝偏差的参量,研究利用熔池特性图像质心参数来建立焊缝偏差视觉模型的方法。在此基础上,针对强电弧干扰所存在的视觉成像不清晰等问题,提出采用分区多尺度统计的方法,获取表征焊接过程偏离状态的高维度特征空间,并以此为依据建立网络预测模型。首先,将焊接图像分离为电弧区、熔合区及焊缝区。其次对分区图像进行横纵坐标重构,并计算其相关性。再提取三个尺度的重构信号的时频特征,构建高维度特征空间。采用主成分分析法,实现对高维特征空间的维度控制,并以此为依据构成网络预测模型的输入向量。分别对强干扰、电弧畸变、聚焦偏离、光圈反射、传感器偏移等问题进行数据采集,获取模型训练及测试样本。实验结果表明,采用BP网络的建模精度高于Elman网络的预测精度。特别的,采用多尺度特征空间的模型明显优于单尺度特征模型。本文还提出构建分区多尺度统计特征BP网络预测模型,经过实验能够避免图像预处理效果对焊缝偏离测量的影响。特别的,该方法对于强电弧干扰下的识别能力极强,无需测量识别对象的几何特征,也无需对图像进行预处理,即可实现对焊接偏离状态的有效识别,采用分区统计方法,可以间接获取焊接偏离状态。实验结果表明该方法所建立的网络预测模型稳定性高,泛化能力强。
[Abstract]:With the development of modern welding automation and the progress of computer vision technology and the high demand for efficiency and quality, the technology of weld image processing and automatic recognition based on visual sensing has become an important research direction of automatic seam tracking. It is difficult to get a clear and stable image of weld zone by the interference of distortion, image sensor error and so on. It is necessary to collect images by a specific spectral filter system, collect images from the image acquisition card and store it in the industrial control machine, and then extract the accurate position of the weld by the process of image preprocessing, image recognition and so on. It is an important research field in automatic arc welding. To realize accurate seam tracking is very important for improving the quality of welding. In order to achieve accurate weld seam tracking, weld deviation (the deviation of weld center and arc) is a key point. A set of welding seam tracking system used in this study is designed and installed in this study. The system uses the industrial control machine and the visual inspection system without auxiliary light source to complete the automatic tracking control of the weld. The general method of visual sensing weld seam tracking is to use the CCD (Charge Couple Device) sensor to obtain the welding image in real time, and then use the image processing technique to extract the weld position information of the weld pool center and the front of the weld pool. The difference between the two is a reflection of the current deviation between the arc and the weld. The process and method of the welding seam processing based on image processing, including image filtering, image enhancement, image two value, edge detection and weld position extraction, are analyzed and studied by mathematical methods. In contrast, this paper focuses on the image processing method based on image centroid, and studies the new technology of extracting weld position deviation information based on image centroid. It is different from the traditional image processing technology to obtain the weld deviation information directly, but select the specific pool image processing area, and take the molten pool image centroid as the analysis welding. The parameter of the seam deviation is used to study the method of establishing the visual model of the weld deviation by using the image centroid parameters of the weld pool characteristics. On this basis, the method of multi-scale statistics is proposed to obtain the high dimension characteristic space which characterizes the deviation state of the welding process by using the method of multi scale statistics for the strong arc interference. Firstly, the welding images are separated into arc region, fusion zone and weld zone. Secondly, the cross longitudinal coordinates are reconstructed and their correlation is calculated. Then the time frequency characteristics of the reconstructed signals of three scales are extracted and the high dimension feature space is constructed. The high dimension feature space is realized by using the principal component analysis method. The input vector of the network prediction model is formed on the basis of dimension control, and the data acquisition of strong interference, arc distortion, focusing deviation, aperture reflection, sensor migration and so on are collected respectively, and model training and test samples are obtained. The experimental results show that the modeling accuracy of BP network is higher than that of the Elman network. The multi scale feature space model is obviously superior to the single scale feature model. This paper also proposes the construction of the multi scale statistical feature BP network prediction model, which can avoid the influence of the image preprocessing effect on the weld deviation measurement. The geometric features of the object and the image are not pre processed, the effective identification of the welding deviation state can be realized. The welding deviation state can be obtained indirectly by the method of zoning statistics. The experimental results show that the network prediction model established by this method has high stability and strong generalization ability.

【学位授予单位】：广东工业大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TG409

【相似文献】