چكيده به لاتين
In low Earth orbit satellites, the received information from the satellite has frequency shift in the receiver, which is called the Doppler effect. Various methods have been proposed to compensate for this shift. With the advancement of machine learning and artificial intelligence as well as the development of powerful hardware, some researchers have gradually started to investigate the problem of estimating Doppler shift using machine learning methods and its subfields, such as deep learning.
In this thesis, some methods for estimating Doppler shift are introduced. A common method for estimating Doppler shift based on fast Fourier transform simulation, and a method for compensating for Doppler shift using deep learning are also introduced. Then, these two methods are compared in terms of speed and accuracy. To compare in terms of accuracy, the bit error rate curve is plotted for these two methods, and based on the plotted curves, the deep learning method has higher accuracy than the fast Fourier transform method. To compare in terms of speed, these two methods are simulated on a hardware. The fast Fourier transform-based method takes 2 milliseconds, and the deep learning-based method takes 0.1 milliseconds. Also, if the deep learning-based method is implemented on a tensor processing unit, it takes 3 microseconds. In the deep learning method, the Doppler shift value considered for network training ranges from -80 to 80 kHz, with a spacing of 10 Hz. After training the network, the bit error rate curve is plotted to display the network accuracy. Finally, the deep learning-based method is evaluated in the real world with Doppler values of 10, 20, 30, 40, 50, 60, 70, 80 kHz.
