Page 186 - Kaleidoscope Academic Conference Proceedings 2020
P. 186
2020 ITU Kaleidoscope Academic Conference
predetermined parameter. Similarly, the parameters w 2 is to the problem of handover.
chosen according to the degree of negative impact of one
handover. By using the Dijkstra’s shortest path algorithm, we Intuitively, the RSRP series in the LEO network has strong
can find the longest path from the first time slot to the last regularity, so a relatively simple neural network structure
time slot, which is actually the optimal handover strategy for should be chosen to reduce the training time and prevent
this UE. overfitting. The LeNet-5 [12] was firstly designed for
character recognition and is a relatively simple modern CNN
structure. The default size of the input of LeNet-5 is 32 × 32.
3.2 A CNN based optimization for handover
However, in the LEO network model presented in Subsection
RSRP is defined as the linear average over the power of 2.1, the number of detectable beams for one UE is generally
the resource elements that carry some predefined reference smaller than 32. Therefore the size of the input data needs
signals. Assume a UE can predict the RSRP of different to be reduced. Actually, in the simulation the number of
beams for a long period, then the method in Subsection considered beams in each time slot is set to be 10. The length
3.1.2 can be used to find the optimal handover strategy. of a time slot is set to be 0.5s and the RSRP values in the
However, in most cases a UE only knows its historical RSRP. previous 10 time slots are used to form the input. Then the
A standard 5G UE needs to measure the RSRPs of detectable input of the CNN is a matrix of size 10 × 10. In LeNet-5, two
cells, and handover to the strongest cell if its RSRP minus convolutional layers are used. The two convolution kernels
a predetermined threshold is larger than the serving RSRP. both have size 5 × 5. Besides, two pooling layers are used
In this way the information hidden in the historical RSRP is to reduce the number of trained parameters. Because of
ignored. Actually, at least in the LEO scenario, the historical the reduced input size, some layers in LeNet-5 need to be
RSRP is able to help UEs to make suboptimal handover customized. First, one convolution kernel is reduced to have
decisions. The series of historical RSRPs of the strongest size 3 × 3. Then the pooling layers are deleted since the
K beams in each time slot forms a two-dimensional matrix. number of parameters is not large. The structure of the
A customized CNN is used to optimize the handover decision resulting CNN is presented in Figure 5. The output of size 10
based on the matrix in this subsection. is corresponding to the 10 kinds of handover decisions, i.e.,
which one of the 10 strongest beams the UE will connect in
CNN is an effective tool to elicit information from the next time slot.
two-dimensional data. It has been widely used to extract
features from images. A classical CNN consists of The data preprocessing and the training procedure consist of
one or more convolutional layers, pooling layers, and four steps as follows.
fully-connected layers. The features of the input data are 1. For every UE, generate the RSRP values of different
extracted layer by layer, and are summarized in the last satellites in every time slot. If one satellite is invisible
fully-connected layer to generate the final output. Compared or its signal is too weak to detect, then the RSRP values
to the fully-connected layer, the convolutional layer takes are regarded as 0.
advantage of the strong local spatial correlation in natural
images and only has a few parameters to be trained. It 2. Compute the best handover strategy for every UE
is worth mentioning that the matrix of RSRP also has the based on the proposed directed graph based method in
“local spatial correlation”, i.e., the cooperation of the RSRP Subsection 3.1.2.
values in adjacent time slots and the RSRP values of the
nearest 3 or 4 beams are more likely to contain information 3. For every UE in every time slot, the previous 10 RSRP
for handover decisions. Therefore it is suitable to apply CNN values of the 10 strongest beams are used to form a
Figure 5 – The CNN structure for handover optimization
– 128 –