电弧焊熔透状态视觉检测模型研究

发布时间：2017-12-28 08:21

本文关键词：电弧焊熔透状态视觉检测模型研究　出处：《广东工业大学》2015年硕士论文　论文类型：学位论文

【摘要】：焊接是制造业的基础。随着工业制造要求的提高,焊接质量的要求也变得越来越重要。电弧焊由于自身技术发展十分成熟,且适用的范围也十分广泛,目前应用在诸多领域。因此,人们对电弧焊的焊接质量提出了更高的要求。在越来越注重效率的今天,电弧焊焊接实时监测显得愈发重要,为此论文研究基于机器视觉传感的熔透状态监测模型。在实时监测焊接质量的过程中,必须通过当前的焊接状态与焊接质量之间的关系建立数学模型以获取熔透实时监测状态。许多研究表明,焊接过程中形成的熔池与焊缝背部的熔透有着密切的联系,而熔透与最终的焊接质量联系紧密,因此通过提取的熔池特征能与当前的焊接质量建立对应关系。通过视觉传感器拍摄焊接过程的熔池图像,利用相关的图像处理技术从中提取出最能代表熔池信息的特征参数,再与此时的焊接质量对应起来,建立熔池特征与熔透状态之间的数学模型。通过多组焊接试验条件验证模型,证明建立的数学模型能够有效地预测出当前的熔透状态,最终得到当前的焊接质量。对熔透当前的状态进行实时监测可以看作是模式识别的一类,而神经网络则在模式识别这一方面有着较高的预测精度,因此论文根据神经网络的特点设计了两个不同训练算法的神经网络：BP算法(Back Propagation Algorithm)和ICA算法(Imperialist Competitive Algorithm),利用神经网络模型对熔透状态进行准确预测。其中,ICA算法是通过模拟帝国主义竞争的过程来得到优化问题最优解的一种全局优化搜索算法。首先,从熔池图像中提取出3个特征参数：熔池面积、熔池熔宽、熔池半长。其次,采用试凑法确定神经网络隐含层神经元个数,建立一个预测精度较高的神经网络模型,最后用两个算法训练神经网络,获得两个模型对熔透各自的识别率。试验结果表明选用的BP神经网络最优识别模型比ICA神经网络的最优模型对不同焊接条件的熔透状态有着很高的预测精度。针对BP神经网络的识别精度高但过于依赖初始权值阈值的特点和ICA算法在全局搜索的优点,论文将ICA算法与BP神经网络结合,利用ICA算法选取BP神经网络初始权值及阈值,提出一种ICA-BP神经网络的熔透识别模型。首先将模型的权值阈值作为优化问题要求的解,然后根据熔透状态输出的实际误差与期望误差的差的大小及时调整权值及阈值。其次利用ICA得到神经网络的初始权值、阈值,通过BP算法对神经网络进行训练,得到使得模型识别精度最高的一组权值和阈值。最终利用不同试验条件寻找对熔透的识别效果最好的神经网络模型。试验结果证明ICA-BP神经网络对熔透状态的预测精度优于BP神经网络的预测精度。
[Abstract]:Welding is the foundation of the manufacturing industry. With the improvement of industrial manufacturing requirements, the requirement of welding quality is becoming more and more important. Arc welding is widely used in many fields because of its mature technology and wide range of application. Therefore, people have put forward higher requirements for the welding quality of arc welding. Nowadays, more and more attention is paid to efficiency. The real-time monitoring of arc welding is becoming more and more important. Therefore, a penetration monitoring model based on machine vision sensing is studied in this paper. In real-time monitoring of welding quality, we must establish a mathematical model through the relationship between the current welding state and the welding quality, so as to get the real-time monitoring state of penetration. Many studies have shown that the weld pool formed during the welding process is closely related to the penetration of the back of the weld, and the penetration is closely related to the final welding quality. Therefore, the characteristics of the molten pool can be used to establish the corresponding relationship with the current welding quality. The image of molten pool is captured by visual sensor, and the characteristic parameters that represent the information of molten pool are extracted from related image processing technology. Then the mathematical model between molten pool characteristics and penetration state is established based on the corresponding welding quality. Through multi group welding test condition verification model, it is proved that the established mathematical model can effectively predict the current penetration state, and finally get the current welding quality. The current penetration state monitoring can be regarded as a kind of pattern recognition, pattern recognition and neural network in this area has a higher prediction accuracy, so this paper according to the characteristics of the neural network to design a neural network with two different training algorithms: BP algorithm (Back Propagation Algorithm) and ICA algorithm (Imperialist Competitive, Algorithm) by using the neural network model to forecast the penetration state. Among them, ICA algorithm is a global optimization search algorithm to optimize the optimal solution by simulating the process of imperialist competition. First, extract 3 feature parameters from the image of weld pool in molten pool area, pool and pool half length Rongkuan. Secondly, we use the trial and error method to determine the number of neurons in the hidden layer of the neural network, and establish a neural network model with high prediction accuracy. Finally, we use two algorithms to train the neural network, and get the recognition rates of two models for penetration. The experimental results show that the optimal identification model of BP neural network is higher than that of ICA neural network, and has high prediction accuracy for penetration condition of different welding conditions. According to the characteristics and advantages of the ICA search algorithm of BP neural network with high accuracy but is too dependent on the initial weights in the global threshold, the ICA algorithm combined with BP neural network, using ICA algorithm to select BP neural network initial weights and threshold, proposes a ICA-BP neural network fusion recognition model. First, the weight threshold of the model is used as the solution of the optimization problem, and then the weights and thresholds are adjusted according to the difference between the actual error and the expected error of the penetration state output. Secondly, we use ICA to get the initial weights and thresholds of neural network, and train the neural network through BP algorithm, and get a set of weights and thresholds that make the model identification accuracy the highest. Finally, different experimental conditions are used to find the neural network model which has the best recognition effect on penetration. The experimental results show that the prediction accuracy of ICA-BP neural network is better than that of BP neural network.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TG441.7;TP391.41

【相似文献】