【作者】 王涛；

【导师】 陈波；

【作者基本信息】 哈尔滨工业大学 ， 材料与化工（专业学位）， 2023， 硕士

【摘要】 AlSi10Mg由于密度小,耐腐蚀性优良,比强度高等特性被广泛应用于制造航空航天、汽车制造等领域的结构件。随着技术进步,铝合金结构件的形状造型和结构越发复杂,传统切削加工的方式在加工这些结构件时效率偏低而且会产生大量材料浪费。激光定向能量沉积技术可以较为快速的一次性制造出完整结构件,同时具有材料利用率高、结构件加工周期短等优势。但在AlSi10Mg激光定向能量沉积过程常常出现尺寸精度低,表面成形不良、结构件内部出现气孔缺陷等问题。对AlSi10Mg激光定向能量沉积过程进行质量监测对实现沉积过程的精细化控制,改善成形、减少气孔缺陷具有重要意义。本文利用同轴CMOS相机采集沉积过程实时熔池图像,采用基于深度学习的方法对AlSi10Mg沉积层的成形和气孔缺陷进行预测。主要工作内容如下:(1)搭建了自行设计的可调节特征提取层深度的深度决策回归网络对AlSi10Mg单层单道沉积高度、宽度和熔深进行预测,研究发现卷积层数分别为14、10、16时分别对沉积高度、宽度和熔深预测误差最小。(2)在单层沉积层成形预测模型基础上,结合多层沉积时前层成形影响后层的特性搭建了多输入深度回归网络用于AlSi10Mg单道多层沉积高度预测,优化后模型均方根误差相较于原始模型降低21.3%。(3)基于经典深度学习分类网络对AlSi10Mg单层单道沉积层气孔预测进行了系统测试,并对模型结构和损失函数进行改进;改进模型气孔分类准确率达到82.61%,相较于Resnet34预测准确率提高了1.83%,并且计算速度比Resnet34快2.2倍。(4)将普通的二值交叉熵损失函数替换为三元组损失函数,三元组损失函数阈值为0.2时,验证集准确率最高达到85.84%,相较于原始模型准确率提高了3.23%,优化后模型召回率达到91.84%,表明模型对气孔的检出效果达到较高的水平。(5)结合单道多层沉积高度预测模型、单层单道气孔预测模型、单道多层沉积过程熔深预测模型的结果,建立了全新的多输入深度分类网络,将AlSi10Mg单道多层沉积过程气孔预测问题分解成对重熔区和堆积区气孔的预测两个任务,多输入深度分类网络在重熔区验证气孔数据集上准确度为76.84%,召回率为82.76%。在堆积区气孔验证数据集上准确度为79.80%,召回率为87.18%。本文通过深度学习方法对AlSi10Mg沉积成形和气孔缺陷进行监测,对多层沉积高度预测模型进行针对性改进并对成形进行控制,创新性的提出通过多模型协同对多层沉积过程气孔进行监测的方案并相较于传统模型有较好的气孔识别效果。更多还原

【Abstract】 Due to its low density,excellent corrosion resistance and high specific strength,Al Si10 Mg is widely used in aerospace,automotive and other structural parts.And with the progress of technology,the shape and structure of aluminum alloy structural parts are more and more complex,and the traditional cutting way in the processing of these structural parts is low efficiency and will produce a lot of material waste.Laser-based directed energy deposition technology can produce complete structural parts quickly,and has the advantages of high material utilization rate and short processing cycle of structural parts.However,in the process of Al Si10 Mg laser directed energy deposition,there are often problems such as low dimensional accuracy,poor surface forming,and porosity defects inside the structure.The quality monitoring of Al Si10 Mg laser directed energy deposition process is of great significance to realize the fine control of the deposition process,improve the forming and reduce the porosity defects.In this study,a coaxial CMOS camera was utilized to collect real-time molten pool images during the deposition process,and the deep learn-based method was adopted to predict the formation of Al Si10 Mg sediments and whether pores were generated.The main work contents are as follows:(1)A self-designed depth decision regression network was built to predict the deposition height,width and penetration depth of Al Si10 Mg monolayer single channel under different process parameters.It was found that when the convolutional layers were 14,10 and 16,the prediction error of deposition height,width and penetration depth was minimum.The minimum relative errors were 10.48%,6.12%and 11.44%,respectively.(2)Based on the prediction CNN model of single-layer and single-channel deposition,a multi-input deep regression network was established to predict the height of Al Si10 Mg single-channel multi-channel deposition combined with the characteristics of multi-layer deposition.The test model found that the prediction error was minimum when the weight ratio of the adjacent molten pool image was set as 1:1:0.5.The root mean square error of the optimized model is reduced by 21.3% compared with that of the original model with high prediction directly.(3)Based on the classical deep learning classification network,the pore prediction of Al Si10 Mg single-layer and singlechannel sedimentary layer was systematically tested,and the model structure and loss function of the classical network were improved.With the improved basic block A,the highest porosity classification accuracy reaches 82.61%,which is 1.83% higher than that of Resnet34,and the calculation speed is 2.2 times faster than that of Resnet34.(4)When the threshold value of triplet loss function is 0.2,the accuracy of validation set reaches85.84%,which is 3.23% higher than that of the original model,and the recall rate of the optimized model reaches 91.84%,indicating that the model has a relatively high porosity detection effect.(5)Combined with the results of the prediction model of single-channel multilayer deposition height,single-layer single channel porosity prediction model and single-channel multilayer deposition depth prediction model,a new multi-input depth classification network was established,and the prediction of porosity in Al Si10 Mg singlechannel multilayer deposition process was divided into two tasks: prediction of porosity in remelting zone and accumulation zone.The accuracy of the trained multi-input deep classification network is 76.84% and the recall rate is 82.76% in the validation of stomatal data set in the remelted zone.The accuracy of stomatal validation data set in accumulation area is 79.80%,and the recall rate is 87.18%.In this paper,the deep learning method is used to monitor the formation and porosity defects of Al Si10 Mg deposition,improve the prediction model of multilayer deposition height and control the formation,and innovatively propose a scheme of monitoring porosity in the process of multilayer deposition through the cooperation of multiple models,which has a better porosity recognition effect compared with the traditional model.更多还原