样本标签污染条件下的雷达辐射源个体识别技术

针对辐射源个体识别(Specific Emitter Identification,SEI)中由于数据集存在错误标签导致识别率下降的问题,提出了 1种有监督和无监督融合的错误标签识别和纠正方法。首先采用无监督密度峰值聚类方法将数据集中出现的标签错误样本找出,再使用 K折交叉实验对这些标签异常的样本进行预测投票,将得票数多的标签作为错误标签纠正的结果。经过清洗的数据集再通过卷积神经网络进行训练,得到 1个较为理想的辐射源个体识别的网络模型,保证了在样本污染条件下,辐射源个体识别网络仍能具有较好的识别率。文章所提方法的识别率相比未经处理的数据集的识别率在标签错误率小于 30%时平均提高 3.3%；在标签错误率大于 30%时,也能使个体识别率达到 90%左右,验证了文章所提方法在对错误标签的识别和纠正上可以取得较好的效果。

Individual Identification Technology of Radar Radiation Sources Under Sample Label Pollution Conditions

In response to the problem of reduced recognition accuracy in Specific Emitter Identification (SEI) due to wrong label in the dataset,a supervised and unsupervised fusion method for mislabel recognition and correction is pro-posed. Firstly, the unsupervised density peak clustering method is used to identify samples with label errors in the dataset, and then K-fold crossover experiments are used to predict and vote on these samples with abnormal labels, using the label with a large number of votes as the result of correcting incorrect labels. The cleaned data set is trained by convolutional neural network to obtain an ideal network model for emitter individual recognition, which ensures that the emitter individ-ual recognition network can still have a good recognition accuracy under the condition of sample pollution. The recogni-tion accuracy of the proposed method is improved by an average of 3.3% when the label error rate is less than 30% com-pared to the unprocessed dataset. When the label error rate is greater than 30%, the individual recognition accuracy can al-so reach around 90%, verifying that the proposed method can achieve good results in identifying and correcting incorrect labels.