基于多时间特征融合网络的ADS-B实采信号分离

ADS-B Actual Signal Separation Based on Multi-Temporal Features Fusion Network

不同于以往单天线广播式自动相关监视(Automatic Dependent Surveillance-Broadcast,ADS-B)信号分离中利用仿真的ADS-B信号制作数据集,将单天线接收的真实飞机发射的ADS-B原始信号通过调整信号起始时间以及功率并人为增加噪声来制作数据集。为了提高信号分离的时域波形精度,提出一种多分辨率多时间特征融合重采样(Multi-Temporal fusion Resampling of Multi-Resolution Features,MTRM-RF)网络,通过卷积将信号转化成不同采样率的信号并分别使用多层堆叠逐渐膨胀的一维卷积提取不同时间间隔的特征,以获得更多的时间信息。对多种基于深度学习的语音分离网络进行比较发现,MTRM-RF网络能够有效地融合ADS-B信号的不同采样率、不同时间间隔采样点的特征进行训练。并且随着训练集数据量的增加,分离信号的平均解码正确率达到88.39%,证明该网络可有效分离单天线实采的ADS-B交织信号。

Different from previous single antenna Automatic Dependent Surveillance-Broadcast(ADS-B)signal separation papers that use simulated ADS-B signals to create datasets,this paper creates datasets by adjusting the signal start time or power and artificially adding noise to the original signals transmitted by real aircrafts with single antenna receivers.In order to improve the time-domain waveform accuracy of signal separation,this paper proposes a Multi-Temporal fusion Resampling of Multi-Resolution Features network(MTRM-RF),the network transforms the signal into signals with different sampling rates through convolution and extracts features of different time intervals using multi-layer stacked one-dimensional convolution that gradually expands,in order to obtain more temporal information.By comparing several deep learning-based speech separation networks,it is found that the MTRM-RF network can effectively fuse the features of different sampling rates and different time interval sampled points of ADS-B signals for training.Furthermore,as the amount of training data increases,the average decoding accuracy of separated signals achieves 88.39%,demonstrating the effectiveness of the network in separating actual single-antenna ADS-B overlapped signals.