Kolloquium - Details zum Vortrag
Sie werden über Vorträge rechtzeitig per E-Mail informiert werden, wenn Sie den Newsletter des kommunikationstechnischen Kolloquiums abonnieren.
Alle Interessierten sind herzlich eingeladen, eine Anmeldung ist nicht erforderlich.
Master-Vortrag: AI-Based Optimization of Pulse Shaped OTFS for Vehicular Communication
Vlerar Shala
Donnerstag, 16. September 2021
11:00 Uhr
virtueller Konferenzraum
Future vehicular communication systems require a reliable communication to enable connected and automated driving. Vehicular channels are considered to be truly doubly-dispersive channels, i.e., strongly varying in both time and frequency. However, the performance of the traditional wireless communication systems operating under doubly-dispersive channels degrades significantly. Orthogonal time frequency and space (OTFS) is a recently proposed modulation scheme designed to be robust against doubly-dispersive channels. OTFS is a new and novel two-dimensional (2D) pulse-shaped modulation scheme designed in the delay-Doppler (DD) domain and promises significant improvements on the physical layer. OTFS is a combination of classical Gabor signaling with a unique time-frequency (TF) spreading. Data symbols are located in the DD domain and spread over the entire TF domain using the symplectic finite Fourier transform (SFFT). The TF spreading accounts in a linear fashion for the doubly dispersive nature of time-variant multi-path channel. The TF spreading is also referred to as OTFS transform and the Gabor transform as the Heisenberg transform. However, OTFS requires proper channel information and the use of appropriated equalizers to exploit the full spreading gain.
Doubly-dispersive channels are characterized by their spreading region spanned by two times max, and max, corresponding to the maximum Doppler shift and delay spread. To mitigate the strong interference between different TF slots, TF grid and pulses should “match” the DD spreading region of the doubly-dispersive channel, typically determined on a longer time scale-like. This approach is known as pulse and grid matching rule. In this thesis, we use the so-called mobility modes with distinct grid and pulses to follow the pulse and grid matching rule. However, mobility modes control the interference on a coarse level because they cannot match the pulse and grid matching rule completely. The remaining interference is accounted by the implemented mean square error (MMSE) linear equalizer, which is tuned for each frame. We implement the OTFS transceiver based on a polyphase implementation for orthogonalized Gaussian pulses and evaluate mobility modes with doubly-dispersive vehicular channels generated by the QuaDRiGa channel simulator.
For practical implementation, the performance of mobility modes should be predicted and then the mobility mode with the best performance is selected to send the next frame. However, the prediction of mobility modes requires the prediction of the channel itself. In this thesis, we propose a novel method to predict the channel by predicting the spreading function which represents the channel in the DD domain. We propose to predict the spreading function using a new variant of Long Short-Term Memory (LSTM) neural networks known as Convolutional LSTM. The architecture of Convolutional LSTM is both recurrent and convolutional which enables it to capture the spatiotemporal correlation of the data and utilize it for the prediction.
