基于深度神经网络和随机森林集成模型的ADS-B辐射源个体识别

针对辐射源个体识别中单一神经网络随着辐射源个体数量的增加，模型的识别准确率显著下降的问题，提出一种基于深度神经网络和随机森林集成模型的广播式自动相关监视(ADS-B)辐射源个体识别方法。该方法利用多种深度神经网络模型和随机森林对增强数据集进行训练，然后利用集成学习方法中的硬投票方法对各网络模型和随机森林识别得到的结果进行集成表决，使得识别结果更具有说服力，同时在在辐射源个体数量增加的情况下依旧保持较高的识别率。实验结果表明，在融合了DRSN、VGG、ResNet、GoogleNet、DenseNet 5类神经网络和随机森林后，相比于单一的神经网络，识别准确率能够提升了3%～20%,且在辐射源个体数量增加的情况下依然能保持较高的识别准确率。

Arc Extraction Analysis for Circularity Measurement of Small Cylindrical Parts by the Segmenting-stitching Method

To reduce the stitching error of circularity measurement of small cylindrical workpieces (Diameter less than 3 mm) by the segmenting-stitching method, arc contour extraction is analyzed in this paper. The coordinates of cross-sectional circle of a small cylindrical part are segmented into several equal arcs to be obtained by a two-dimensional coordinate measuring machine. The circularity contour of the small cylindrical part can be formed by stitching a series of arc contours which are calculated by the obtained arc coordinate data. Due to the different measuring pressure angles of different measuring positions, the accuracy of obtained arc coordinate points is different. The bigger the pressure angle is, the accurate the obtained arc coordinate data are. The ex-periments show that the accuracy of two ends of the arc data is not as good as the central part. Therefore, the two ends of the obtained arc data are appropriately to be cut off, namely, only the central part of the arc data are extracted to be used for the stitching. As a result, the mean value of the matching coefficient is enhanced by 12%, the deviation between the overlap part of the neighbouring arc contour is reduced by 26%, and the average curvature of the arc contours is improved with the extraction method. Thus, the accuracy of the stitched cir-cularity contour can be improved by this extraction procedure in the segmenting-stitching method for the cir-cularity measurement of the small cylindrical parts.