1 Department of Materials Science and Engineering, Southern University of Science and Technology, Shenzhen 518055, Guangdong, China
2 Songshan Lake Materials Laboratory, Dongguan 523830, Guangdong, China
3 Institute of Physics, Chinese Academy of Sciences, Beijing 100190, China
4 Jiaxing Research Institute, Southern University of Science and Technology, Shenzhen 518055, Guangdong, China
5 High Performance Computing Department, National Supercomputing Center, Shenzhen 518055, Guangdong, China
Ti-6Al-4V is a benchmark Ti alloy. Laser wire additive manufacturing (LWAM) offers advanced manufacturing capability to the alloy for applications possibly including exploration of outer space. As a typical multiple-variable process, LWAM is complex, which, however, can be analyzed, predicated or even optimized by artificial intelligence (AI) methods such as machine learning (ML). In this study, printing parameters of the Ti-6Al-4V is firstly optimized using single-track-single-layer experiments, and then single-track-multiple-layer samples are printed, whose properties in terms of hardness and compressive strength are analyzed subsequently by both experiments and ML. The two ML approaches, artificial neural network (ANN) and support vector machine (SVM), are employed to predict the experimental results, whose coefficients of determination R2 show good values. Further optimized properties are realized by adopting genetic algorithm (GA) and simulated annealing (SA) approaches, which contribute to high mechanical properties achieved, for instance, an engineering compressive strength of about 1694 MPa. The results here indicate that important mechanical properties of the LWAM-prepared Ti alloys can be well predicted and enhanced using suitable ML approaches.
laser technique laser wire additive manufacturing (LWAM) Ti-6Al-4V machine learning mechanical properties support vector machine (SVM) artificial neural network (ANN)
1 海军工程大学, 武汉 430000
2 中国人民解放军91715部队, 广州 510000
3 陆军勤务学院, 重庆 401000
为解决低信噪比条件下传统雷达辐射源识别准确性低、时效性差、稳健性不强的问题, 提出了一种基于随机森林的雷达辐射源型号识别算法。算法以载频、脉宽、脉冲重复周期为识别特征向量, 首先从先验样本集中随机抽取得到多个训练集, 然后使用训练集构建多个决策树分类器, 最后通过多个决策树分类器对新识别特征向量进行识别并投票得到最终识别结果。仿真实验表明, 该算法在低信噪比条件下依然具有较好的稳健性与时效性, 能够有效解决战场雷达辐射源识别的问题。
雷达辐射源 识别算法 随机森林 radar emitter identification algorithm random forest
火箭军工程大学 导弹工程学院,陕西 西安 710025
针对红外图像相比于RGB图像纹理较少的特性,提出一种使用Lp归一化权重的红外目标检测网络压缩方法,旨在改进基于卷积神经网络的目标检测方法对红外图像场景的适应性,在压缩网络规模的同时提升其泛化能力。首先阐述了Lp归一化权重的稀疏性可以通过调节p进行精确控制这一现象。基于该现象,提出了一种目标检测网络稀疏化训练方法。该方法分别使用Lp球面梯度下降与经典梯度下降训练主干网络和检测器,以平衡网络规模与拟合精度。仿真红外数据集测试结果表明,其在网络规模和目标检测精度方面均优于稠密模型:在网络规模上,稀疏化方法将Faster R-CNN、(Single Shot multibox Detector,SSD)与YOLOv3的有效参数分别减少了52%、78%和66%;在检测精度上,稀疏化方法将Faster R-CNN、SSD和YOLOv3的(mean Average Precision, mAP)分别提高了0.1%、0.3%和0.2%,验证了所提出方法的有效性。
红外目标检测 稀疏神经网络 Lp归一化 受约束梯度下降 infrared object detection sparse neural network Lp normalization constrained gradient descent 红外与激光工程
2021, 50(8): 20200510
火箭军工程大学 导弹工程学院,陕西 西安 710025
针对红外数据集规模小,标记样本少的特点,提出了一种红外目标检测网络的半监督迁移学习方法,主要用于提高目标检测网络在小样本红外数据集上的训练效率和泛化能力,提高深度学习模型在训练样本较少的红外目标检测等场景当中的适应性。文中首先阐述了在标注样本较少时无标注样本对提高模型泛化能力、抑制过拟合方面的作用。然后提出了红外目标检测网络的半监督迁移学习流程:在大量的RGB图像数据集中训练预训练模型,后使用少量的有标注红外图像和无标注红外图像对网络进行半监督学习调优。另外,文中提出了一种特征相似度加权的伪监督损失函数,使用同一批次样本的预测结果相互作为标注,以充分利用无标注图像内相似目标的特征分布信息;为降低半监督训练的计算量,在伪监督损失函数的计算中,各目标仅将其特征向量邻域范围内的预测目标作为伪标注。实验结果表明,文中方法所训练的目标检测网络的测试准确率高于监督迁移学习所获得的网络,其在Faster R-CNN上实现了1.1%的提升,而在YOLO-v3上实现了4.8%的显著提升,验证了所提出方法的有效性。
红外目标检测 半监督学习 迁移学习 伪监督损失函数 infrared object detection semi-supervised learning transfer learning pseudo-supervised loss function 红外与激光工程
2021, 50(3): 20200511