Beamweight Prediction Using Complex Valued Convolutional Neural Networks
Publicerad
Författare
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
The advanced antenna systems used for telecommunication today utilize antenna
array structures. These structures consist of grids of antenna elements divided into
smaller subarrays of elements where each subarray is dynamically controllable, independently
from the others. Each subarray is typically mapped to a baseband port,
and the amplitude and phase of the signal on each of this baseband port can be
controlled independently, which is called excitation vectors or beamweights.
The main aim of this thesis was to investigate and develop a machine learning model
that can predict beamweights for unseen antenna arrays such that the radiated far
field of the antenna follows a specific pattern. The model was evaluated and compared
to a more common method, stochastic optimization. Furthermore, the model
was tested to see how stable the outputs are against noise and if the model is scalable
to antenna arrays of different sizes.
The main conclusion from the thesis is that it is possible to train a complex valued
convolutional neural network (CV-CNN), a type of machine learning model,
to predict the desired beamweights. The results also show that the beamweights
generated from the two different methods result in far fields which favour different
performance metrics. An important difference in the models is the time it takes to
generate beamweights. A pre-trained CV-CNN generates beamweights in a fraction
of the time it takes to generate beamweights using the more common method. The
conclusion from the comparison is that the method should be chosen case by case.
The stability analysis shows that the model is more sensitive to noise in the phase
than noise in the amplitude of the input. The reason for this might come from
an approximation made while processing the data but a possible solution to this is
presented.
Finally, the results show promise of scalability of the model. The model must
however be optimized for each size of antenna array in order to get satisfactory
results.
Beskrivning
Ämne/nyckelord
Advanced Antenna Systems, Antenna Arrays, Beamforming, Beamweights, Machine Learning, Artificial Neural Networks, Complex Valued Convolutional Neural Networks.