An ensemble ML model design to classify LOS/NLOS propagation channel, Toward positioning accuracy enhancement in 5G-NR InF scenarios
Talebian, Hamed (2023)
2023
Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. Only for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2023-12-27
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202401031050
https://urn.fi/URN:NBN:fi:tuni-202401031050
Tiivistelmä
Line of sight (LOS) and non-line of sight (NLOS) channel classification has numerous applications in wireless communication engineering. It can be exploited for target identification and tracking in sensing-dependent applications such as autonomous driving or surveillance systems. Additionally, it can enhance the positioning accuracy of location-based services since employing a LOS radio channel is required for highly accurate location estimation. Although there are several statistical and data-driven methods for this purpose, apparently, the application of heterogeneous parallel ensemble learning (EL) is not thoroughly investigated for 5G New Radio (5G-NR) signals and indoor factory (InF) channel propagation scenarios. This thesis addresses this research gap by simulating the essential transmission and reception signal and implementing a stacking ensemble learning to predict LOS/NLOS received signals.
To reach the above achievement, firstly, the positioning reference signal (PRS) is generated based on 5G NR technical specification (TS)s to be used as the transmitted signal. These TSs include instructions for generating an array of pseudo-random sequence/noise (PNS) and mapping them into physical resource block (RB)s in time and frequency domain to create the final PRS. Secondly, channel effects are simulated by using a deterministic raytracing dataset, provided for 5G-NR DH-InF channel scenario, which is a 3D Cartesian space with dense clutter density and high transmission point (TP)s, elevated above the clutters. The received signals for all reception point (RP)s are generated accordingly, and a collection of statistical features such as average power are extracted to be regarded as the input data for EL.
Thirdly, a collection of level-one pipelines is selected, including preprocessing and classification functions. By execution of an exhaustive search, the pipelines are ranked according to multiple performance metrics and nine top pipelines are selected. Afterward, their critical hyperparameters are tuned. The best level-two classifier is chosen by searching for the best classifier among level-one classifiers. Finally, the best structure is used for investigating the effects of physical radio parameters such as SNR on LOS/NLOS prediction accuracy. To attain this analysis, a range of variation is defined for five parameters, and EL model performance is analyzed to understand the robustness of the model in variable conditions.
The results indicate that two-level heterogeneous parallel EL outperforms the single-level classification even when sequential or homogeneous EL (boosting) classifiers are used, both in terms of model performance metrics and prediction accuracy since the worst True LOS and NLOS label prediction is more than 90% and 80% respectively. In addition, EL provides a straightforward solution to classify imbalanced binary samples by using an internal cross-validation (CV) algorithm to evaluate the model. Thus, it can be considered as a novel ML-assisted 5G New Radio (5G-NR) positioning accuracy enhancement (PAE) method.
To reach the above achievement, firstly, the positioning reference signal (PRS) is generated based on 5G NR technical specification (TS)s to be used as the transmitted signal. These TSs include instructions for generating an array of pseudo-random sequence/noise (PNS) and mapping them into physical resource block (RB)s in time and frequency domain to create the final PRS. Secondly, channel effects are simulated by using a deterministic raytracing dataset, provided for 5G-NR DH-InF channel scenario, which is a 3D Cartesian space with dense clutter density and high transmission point (TP)s, elevated above the clutters. The received signals for all reception point (RP)s are generated accordingly, and a collection of statistical features such as average power are extracted to be regarded as the input data for EL.
Thirdly, a collection of level-one pipelines is selected, including preprocessing and classification functions. By execution of an exhaustive search, the pipelines are ranked according to multiple performance metrics and nine top pipelines are selected. Afterward, their critical hyperparameters are tuned. The best level-two classifier is chosen by searching for the best classifier among level-one classifiers. Finally, the best structure is used for investigating the effects of physical radio parameters such as SNR on LOS/NLOS prediction accuracy. To attain this analysis, a range of variation is defined for five parameters, and EL model performance is analyzed to understand the robustness of the model in variable conditions.
The results indicate that two-level heterogeneous parallel EL outperforms the single-level classification even when sequential or homogeneous EL (boosting) classifiers are used, both in terms of model performance metrics and prediction accuracy since the worst True LOS and NLOS label prediction is more than 90% and 80% respectively. In addition, EL provides a straightforward solution to classify imbalanced binary samples by using an internal cross-validation (CV) algorithm to evaluate the model. Thus, it can be considered as a novel ML-assisted 5G New Radio (5G-NR) positioning accuracy enhancement (PAE) method.