Page 192 - Kaleidoscope Academic Conference Proceedings 2021

2021 ITU Kaleidoscope Academic Conference

CDF:angle of channel correlation value in time domain each channel scenario to evaluate the recognition
1
0.9 performance of the trained model. The recognition
performance given in this section is the verification
0.8
0.7 LOS,Low-Speed performance for test set.
LOS,Mid-Speed
0.6 LOS,High-Speed
NLOS,Low-DS,Low-Speed
Probabilty 0.5 NLOS,Low-DS,Mid-Speed Confusion Matrix of Threshold Synthetic Decision
NLOS,Low-DS,High-Speed
0.4 NLOS,High-DS,Low-Speed 1 100.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
NLOS,High-DS,Mid-Speed
0.3 NLOS,High-DS,High-Speed 2 0.00% 99.90% 0.00% 0.00% 0.10% 0.00% 0.00% 0.00% 0.00%
0.2
3 0.00% 10.10% 89.90% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
0.1
4 5.90% 0.00% 0.00% 94.10% 0.00% 0.00% 0.00% 0.00% 0.00%
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
normalized abs value Actual class 5 0.40% 0.00% 0.00% 35.90% 42.60% 3.40% 7.20% 9.70% 0.80%
0.00%
3.30%
Figure 7 – CDF of the angle ( ) for channel time 6 7 3.50% 13.20% 9.00% 17.20% 17.90% 53.10% 78.40% 0.00% 0.00%
0.00%
4.00%
0.00%
0.00%
0.00%
0.40%
0.00%
correlation value 8 0.00% 0.00% 0.00% 7.00% 10.70% 0.20% 32.90% 47.90% 1.30%
9 0.00% 0.00% 0.00% 0.20% 11.90% 34.10% 0.20% 13.20% 40.40%
Figure 6 and Figure 7 show the channel time correlation 1 2 3 4 Predicted class 6 7 8 9
5
normalization factor and its phase. Along with the speed
increases, the correlation value gradually becomes smaller Figure 8 – Confusion matrix of threshold comprehensive
and the phase gradually becomes larger, which can decision
distinguish different speeds to a certain extent.
The simulation in this section uses the wireless channel
4.2 The results of threshold comprehensive decision characteristic set χ = [δ, P, γ, β, φ, y] proposed in Section 2,
where the frequency domain fading factor δ , multipath
This section presents the classification results of the power delay distribution P , and channel power peak
threshold comprehensive decision algorithm. The CIR h(t) response ratio γ in the time domain are calculated on two
is obtained when SNR is 20dB. The threshold used in the receiving antennas and two pilot symbols. The time
algorithm is shown in Table 3. correlation value β and φ is derived from the two pilot
symbols, and is averaged on the two antennas. In this
Table 3 – Threshold value for threshold comprehensive section, we take the CIR of 256 subcarriers in the frequency
decision algorithm domain and do a 256-point IFFT to change it to a time
points are taken before the
Threshold Value domain. Assuming that the L 1
peak and the L points are taken after the peak, then the
2
η δ 0.24 window length is L +L +1. Here the window length of the
2
1
η P 0.10 multipath power delay distribution has two explorations: (1)
channel feature 1 is L =18 and L =36, and the window
2
1
η β 0.65 length is 55. (2) channel feature 2 is L =9 and L =26, and
2
1
η γ 0.96, 0.89 the window length is 36, as in Table 4. For channel feature
η φ 0.003, 0.01 2, we have done further processing: every 4 points are used
as an interval, so there are nine intervals for the 36 points,
and the energy ratio between the total energy of each
Figure 8 shows the confusion matrix of the threshold interval and the energy sum of all 36 points is calculated;
comprehensive decision algorithm. Categories 1-9 finally we can obtain 9-point multipath delay energy
correspond to the nine wireless channel scenarios in Table 1. distribution. So, the total length of the channel
It can be seen that the recognition accuracy of the first 4 characteristic set in channel feature 2 is 46, as in Table 4.
channel scenarios is fair, above 89%, but the recognition
accuracy of other channel scenarios is very low. Different Table 4 – The length of multipath power delay distribution
multipath delay spreads and different speeds under NLOS
channels are almost indistinguishable. The threshold Channel The length of The length The length of
comprehensive decision algorithm has low implementation Feature multipath of window channel
complexity and is easy to implement in hardware, but the power delay L +L +1 characteristicsχ
classification performance is not satisfactory. distribution 1 2
1 L =18, L =36 55 226
4.3 The results of neural network classifier 1 2
2 L =9, L =26 36 46
2
1
A neural network is a supervised classifier, and it needs a
training data set to train the network before it can be used Figure 9 shows the confusion matrix obtained by using
for classification. In this paper, 10000 samples are obtained channel feature 1 when SNR=20dB. The recognition
respectively for the nine wireless channel scenarios in accuracy of each channel scenario is above 90%, which is
Table 1, and the total number of training data is 90000 significantly improved compared to the recognition
samples. In addition, 5000 test samples are prepared for performance of the threshold comprehensive decision
algorithm in Section 4.2. The recognition accuracy rate of

– 130 –