Page 193 - Kaleidoscope Academic Conference Proceedings 2021

Connecting physical and virtual worlds

LOS/NLOS is above 99%, and the distinction of high and Table 5 – Accuracy of channel feature 1 based on BP
low multipath delay spread under NLOS is also above 99%. neural network under different SNRs
The recognition accuracy rate of different speeds is slightly
reduced, but they are all above 90%. Figure 10 shows the SNR/dB 20 15 10 5 0
confusion matrix obtained by using channel feature 2 when channel 99.38% 99.35% 99.31% 98.86% 97.11%
SNR=20dB. The recognition accuracy of LOS/NLOS is scenario 1
basically unchanged, but the distinction between high and channel 99.68% 99.52% 99.47% 99.31% 93.80%
low multipath delay extension under NLOS drops to 95%. scenario 2
The speed recognition accuracy rate also dropped to 86.6%. channel 98.53% 98.42% 98.25% 97.67% 92.49%
In contrast, the window length and the resolution of the scenario 3
multipath delay spread distribution will affect the accuracy channel 98.14% 98.10% 98.04% 97.85% 97.72%
of scenario recognition. Appropriately increasing the scenario 4
window length and the multipath resolution will improve channel 93.81% 93.80% 93.16% 93.04% 92.84%
the accuracy of recognition. Although the recognition scenario 5
accuracy of channel feature 2 is lower than that of channel channel 96.17% 96.04% 95.41% 93.90% 88.21%
feature 1, the length of the channel characteristic set of scenario 6
channel feature 2 is much shorter than that of channel channel 93.76% 92.70% 89.11% 84.14% 79.34%
feature 1, so the computing power of the neural network scenario 7
can be reduced to 1/3 of channel feature 1, which also channel 90.66% 86.81% 82.88% 77.80% 71.68%
reduces the complexity of hardware implementation. scenario 8
channel 96.83% 96.76% 95.45% 90.64% 81.93%
Confusion Matrix of Channel Feature 1 based on BP Neural Network scenario 9
1 99.38% 0.25% 0.00% 0.25% 0.12% 0.00% 0.00% 0.00% 0.00%
2 0.00% 99.68% 0.32% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
5. CONCLUSION
3 0.00% 1.47% 98.53% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
4 0.41% 0.00% 0.00% 98.14% 0.62% 0.41% 0.41% 0.00% 0.00% In this paper, we propose a wireless channel scenario
Actual class 5 0.17% 0.00% 0.00% 2.58% 93.81% 3.26% 0.00% 0.17% 0.00% recognition algorithm for 4G/5G dense urban areas and
0.00%
0.00%
6
0.00%
0.00%
0.00%
96.17%
0.13%
0.26%
3.44%
rural or suburban areas. The recognition algorithm is based
7 0.00% 0.00% 0.00% 0.05% 0.00% 0.00% 93.76% 4.56% 1.63%
on the neural network, and the wireless channel
8 0.00% 0.00% 0.00% 0.00% 0.17% 0.00% 8.05% 90.66% 1.11%
characteristics are extracted as the training data set, and the
9 0.00% 0.00% 0.00% 0.00% 0.00% 0.06% 2.00% 1.11% 96.83%
back propagation algorithm is used for training, and then
1 2 3 4 5 6 7 8 9
Predicted class the neural network model for wireless channel scenario
recognition is obtained. Through simulation verification, it
Figure 9 – Confusion matrix of channel feature 1 based on is proved that this method has a greater performance
neural network improvement than the traditional threshold algorithm.
Under high SNR, the minimum recognition accuracy can
reach 90.66%. But as the SNR decreases, the accuracy of
Confusion Matrix of Channel Feature 2 based on BP Neural Network
1 99.38% 0.00% 0.00% 0.49% 0.12% 0.00% 0.00% 0.00% 0.00%
channel scenario recognition will also decrease. In addition,
2 0.00% 99.68% 0.32% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
the composition structure of the wireless channel
characteristic set will also affect the recognition accuracy,
3 0.00% 1.34% 98.53% 0.00% 0.00% 0.12% 0.00% 0.00% 0.00%
4 0.52% 0.00% 0.00% 96.39% 0.82% 0.41% 1.75% 0.00% 0.10% such as the window length and resolution of the multipath
Actual class 5 0.17% 0.00% 0.00% 3.78% 86.60% 4.47% 2.23% 0.52% 2.23% delay spread distribution. The wireless channel
6
0.00%
0.00%
characteristics proposed in this paper are simple to calculate,
2.42%
93.37%
0.00%
0.26%
0.26%
0.00%
3.70%
easy to implement, and have high engineering application
7 0.00% 0.00% 0.00% 0.87% 0.49% 0.05% 93.27% 4.12% 1.19%
value.
8 0.00% 0.00% 0.00% 0.00% 1.03% 0.09% 9.43% 87.32% 2.14%
9 0.00% 0.00% 0.00% 0.00% 0.72% 1.00% 1.39% 1.56% 95.32%
1 2 3 4 5 6 7 8 9
Predicted class However, there are still some limitations in our work. For
example, the recognition accuracy under low SNR needs to
Figure 10 – Confusion matrix of channel feature 2 based be further improved; and a system simulation needs to be
on neural network constructed to verify the improvement of system
performance based on wireless channel scenario
Table 5 shows the recognition accuracy of channel feature 1 recognition. Nevertheless, we wish that our work can
under different SNRs. As the SNR decreases, the provide new insights and motivation for the study of
recognition accuracy of each channel scenario also wireless channel scenario recognition in 4G/5G commercial
decreases, especially when the multipath delay spread is systems.
relatively large under NLOS channel scenarios. The
recognition accuracy of different speeds is the worst with
71.68% for SNR=0dB.

– 131 –