一种基于深度生成模型的ADS-B信号增强和目标识别方法

自动相关监视广播数据(automatic dependent surveillance-broadcast, ADS-B)信号在航空领域通信中占据非常重要的地位，其检测、分析对航空运输安全保障意义重大。ADS-B信号常常带有噪声或干扰，这使得直接解码的准确性受到影响。为了更好地捕捉ADS-B信号的信息提升其准确性，提出了EASTR深度学习模型。所提模型首先使用基于非因果扩张卷积和残差网络结构的方法，对原始含噪ADS-B信号进行降噪与增强；随后，经过降噪处理的信号被转换为星群图像，再利用多层感知机进行分类识别。收集了5 000条来自不同飞机的ADS-B信号数据，在此数据集上将EASTR与其他同类模型进行比较。实验结果表明：不同信噪比下EASTR均在准确率上优于其他模型。通过消融实验验证了数据增强模块的效能。

ADS-B Signal Enhancement and Target Recognition Based on Deep Learning

ADS-B signal plays a very important role in the area of communication in aviation, and it is of great significance for the detection and analysis of ADS-B signal in aviation communication. However, it is so much difficult to directly decode ADS-B signal with noise or interference. In order to better capture information contained in ADS-B signal, this paper proposed a method called EASTR. Firstly, ADS-B signal is de-noised by non-causal dilated convolutions and a residual network which are utilized to enhance the original noisy ADS-B signal to reduce noise. Then, the de-noised signal is transformed into Contour Stellar image and then a multi-level perceptron is leveraged for classification and recognition. In this paper, 5000 ADS-B signal data from different aircrafts are collected. With this dataset, EASTR is compared with multiple baselines under different SNRS and the experimental results show that EASTR outperform all baselines in all cases. The significance of signal enhancement module is also analyzed with ablation experiments.