基于USRP与YOLOv3算法的信号采集与识别设计

Signal acquisition and recognition based on USRP and YOLOv3 algorithm

为了加快射频信号的采集和分析,采用NI USRP-2901硬件设备和LabVIEW软件实现一个射频信号采集处理平台。射频信号经过USRP-2901设备被放大、混频、滤波,滤波后的信号经过AD9361捷变收发器进行模数转换。通过FPGA将数字信号下变频到基带同相I信号和正交相Q信号。分别对I/Q信号加入汉明窗、汉宁窗、矩形窗,之后进行FFT得到频谱图数据。通过YOLOv3算法对标注好的频谱图数据集进行训练,训练完成后得到权重文件。调用权重文件对频谱图进行识别,识别时域信号中添加的是哪种窗函数。实验结果表明,采用YOLOv3算法比其他基于RPN算法的目标检测算法要好,YOLOv3算法泛化能力强,对于背景物体的识别有更高的准确率。

To speed up the acquisition and analysis of radio frequency signals,NI USRP-2901 hardware device and LabVIEW software are used in this paper to realize an RF signal acquisition and processing platform. The analog-to-digital conversion of RF signal is conducted by the AD9361 agile transceiver after the RF signal is amplified,mixed,filtered in the USRP-2901 device. The digital signal is converted to the baseband in-phase I signal and the quadrature phase Q signal by FPGA. The Hamming window,the Hanning window and the rectangle window are added respectively to the I/Q signal,and then FFT is performed to obtain the spectrogram data. The labeled spectrogram data set is trained with the YOLOv3 algorithm to obtain the weight file. The weight file is called to identify the spectrogram to know which window function is added in the timedomain signal. The experimental results show that the method using YOLOv3 algorithm is better than other target detection methods based on RPN algorithm. YOLOv3 algorithm has strong generalization ability and high accuracy rate for background object recognition.