First International Conference on Electronics, Communication and Signal Processing (ICECSP 2024)

COMPARATIVE ANALYSIS OF MACHINE

LEARNING ALGORITHMS FOR LOS/NLOS
IDENTIFICATION
2024 First International Conference on Electronics, Communication and Signal Processing (ICECSP) | 979-8-3503-6459-0/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICECSP61809.2024.10698216

Prachi Agarwal Rohit Kumar Dharmendra Kr. Jhariya Mahesh K. Singh

Dept. of Electronics & Dept. of Electronics & Dept. of Electronics & Dept. of Electronics &
Communication Engineering Communication Engineering Communication Engineering Communication Engineering
NIT Delhi NIT Delhi NIT Delhi NIT Delhi
[email protected] [email protected] [email protected] [email protected]

Abstract— Identification of Line of Sight (LOS) and Non movement of furniture in a room [5]. Wifi technology have
Line of Sight (NLOS) is a challenging task since majority centimeter level accuracy in large effective bandwidth but it
of physical effects that take place at distinct is not suitable for signals resulted due to multipath,
temporal/spatial conditions can affect the LOS interference etc [6]. Zigbee requires sensors of very less
propagation. Ultra Wide Band communication energy and also low cost technology but it offers certain
technology is the key technology in indoor positioning to disadvantage as well it is vulnerable to interference with
identify LOS and NLOS communication channels in signals same frequency band [7].RFID penetrates non-metal
different environmental conditions. Different machine object and solid object. Its positioning coverage is very
learning algorithms are used to identify on the basis of small and offers compatibility issue. Bluetooth is a lighter
various training feature which basically exploits several standard and highly ubiquitous [8], but it requires more
static and time-varying features of the channel impulse number of recieving cells which ultimately make it
response (CIR).In ML, the choice of training features expensive. All such technologies offer some advantages as
plays a crucial role. The paper addresses machine well as some disadvantages. [9].
learning techniques i.e. K-Nearest neighbour, Logistic Our work proposed UWB technology for indoor
Regression, Support Vector Machine and Random positioning. It has large transmission bandwidth. It can
Forest in terms of their performance on different achieve decimeter level positioning for obstacle penetration
parameters. which is the condition for NLOS propagation. UWB with
Keywords: LOS, NLOS, UWB, CIR, Machine Learning channel impulse response (CIR) is proposed in the paper.
The CIR provides information which is best suited when
INTRODUCTION there are obstacle and scattering of objects. The
environment becomes complex to analyse with conventional
The recognition of LOS and NLOS state is prerequisite for methods, therefore ML methods are used. Using ML
alleviating errors, designing a channel model and for methods, only relevant features are extracted. Therefore,
localization. LOS and NLOS are two states where LOS different ML techniques like Random Forest, K Nearest
stands for Line of Sight that is when there is no obstruction Neighbour, Logistic Regression and Support Vector
between transmitter and receiver. While NLOS stands for Machine are used and compared on the basis of Accuracy,
Non Line of Sight that is when signal arrives at receiver Precision, F1 score and Recall.
after going through many obstructions in between travelling The motivation for this research stems from the critical need
from transmitter [1]. to enhance wireless communication systems' reliability and
For localization of a signal, indoor and outdoor localization efficiency, particularly in scenarios where line-of-sight
technologies are two different approach in terms of scope of (LOS) and non-line-of-sight (NLOS) conditions
application and accuracy in identification of LOS/NLOS significantly impact signal propagation [10]. In various real-
state. Outdoor localization technologies include Global world applications such as indoor localization, asset
Navigation Satellite System (GLONASS), Global tracking, and autonomous navigation, accurately
Positioning System (GPS) are some examples [2]. To cover distinguishing between LOS and NLOS signals is pivotal
globally both are widely used localization system but both for ensuring robust performance. However, conventional
suffer a lot in indoor environment due to signal attenuation, methods often struggle to discern between these conditions
interference, object positions and human activity [3, 4]. effectively, leading to degraded system performance and
To tackle problems in indoor environment, different indoor increased error rates [1]. By utilizing the power of machine
positioning technologies are used like WLAN, Wifi, RFID, learning classification models and leveraging the wealth of
Zigbee, Bluetooth etc [4].WLAN covers more than one data collected from ultra-wideband (UWB) modules and
building and is not suitable when there is movement like transceivers, this research endeavor’s to develop advanced

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

Authorized licensed use limited to: Chang'an University. Downloaded on March 18,2025 at 13:42:36 UTC from IEEE Xplore. Restrictions apply.
techniques capable of reliably identifying LOS and NLOS RXPACC Received RX preamble symbols
signals [11,12]. Through this endeavour, we aim to pave the CH Channel number
way for improved wireless communication systems with FRAME_LEN Length of frame
heightened resilience to challenging propagation PREAM_LEN Preamble length
environments, ultimately facilitating the seamless operation BITRATE Bitrate
of a wide array of emerging applications [13]. PRFR Pulse repetition frequency rate (in
MHz)
DATA COLLECTION CIR0-CIR15 Absolute values of channel impulse
response (1016 samples each)
Measurements were conducted across seven distinct indoor
locations to ensure the diversity and representativeness of
the collected data. These locations include Office1, Office2 2) TRAINING SET
with Small Apartment, Small Workshop, Kitchen, Boiler
Room and Bed Room, Living Room. Each location yielded As a transceiver DWM1000 module was used which is
a substantial dataset, comprising 3000 samples for both based on Ultra-Wideband (UWB) transceiver IC. The
Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) signal integration of all Radio Frequency circuits, power
conditions. This distribution ensures a balanced distribution, antennas are performed by the module & it
representation of both LOS and NLOS scenarios across supports positioning accuracy of 10cm and data transfer rate
varied indoor environments. The total dataset consists of up to 6.8Mbps.
42,000 samples, equally divided between LOS and NLOS The SNPN-UWB board offers robust communication
conditions, to enable robust model training and evaluation. capabilities through IEEE 802.15.4-UWB short-range low-
Randomization techniques were employed during data power ultra-wideband communication channels, utilizing the
collection to mitigate the risk of model overfitting to Deca Wave DWM1000 UWB module. The versatile board
specific locations, enhancing the generalizability of the supports multiple communication interfaces, including
developed models [13]. Throughout the whole UART, USB, I2C, and SPI, facilitating seamless integration
measurements, two UWB nodes were used: one as an into diverse systems. Its adaptable power supply options,
anchor and the other as a tag, capturing LOS and NLOS including external 5V, external 3.3V, battery, or USB
channel traces without attributing to specific positioning, connector power, provide flexibility and convenience in
thereby focusing solely on the characterization of signal various deployment scenarios [13].
propagation conditions rather than localization evaluation.
DEVICES & SOFTWARES USED
1) DATASET
The experimental setup employed for data collection and
The dataset contains samples with headers indicating analysis comprises several key components, each
various parameters. Each sample includes information on contributing to the successful execution of the project. The
LOS/NLOS status, time of flight, first path index and computational backbone of this endeavour is an Intel i5 10th
amplitudes, noise characteristics, channel details, frame generation laptop, chosen for its robust processing
properties, bitrate, and pulse repetition frequency. capabilities and efficient handling of machine learning tasks.
Additionally, it incorporates 1016 columns representing the The Python programming language was utilized to develop
absolute values of the channel impulse response (CIR), and execute the classification algorithms essential for
ranging from CIR0 to CIR15, each with nanosecond analysing the collected data. Facilitating the collection of
resolution. These parameters collectively provide a signal propagation data is the SNPN-UWB board, equipped
comprehensive overview of signal propagation conditions, with the DecaWave DWM1000 UWB radio module. The
vital for analysing and classifying LOS and NLOS scenarios board serves as the interface between the UWB nodes and
to enhance wireless communication system performance. the laptop, enabling seamless communication and data
The following features/ predictors were used for training the transfer. The DecaWave DWM1000 UWB radio module,
machine learning models as shown in Table I. renowned for its high precision and reliability in ultra-
wideband communication, ensures accurate measurement
TABLE I: Description of Various Features and recording of signal characteristics across various indoor
environments.
Parameter Description
NLOS 1 if NLOS, 0 if LOS Together, the integrated system of hardware and software
Measured range Time of flight components forms the foundation for the research
FP_IDX Index of detected first path element in endeavour, empowering the exploration and classification of
CIR accumulator LOS and NLOS signal conditions in indoor environments.
FP_AMP1 First path amplitude (part 1)
FP_AMP2 First path amplitude (part 2) STUDY OF DIFFERENT ALGORITHM
FP_AMP3 First path amplitude (part 3)
STDEV_NOISE Standard deviation of noise Before training the classification models, a crucial
CIR_PWR Total channel impulse response power preprocessing step was undertaken to enhance the efficiency
MAX_NOISE Maximum value of noise and effectiveness of the machine learning algorithms.

Authorized licensed use limited to: Chang'an University. Downloaded on March 18,2025 at 13:42:36 UTC from IEEE Xplore. Restrictions apply.
Feature extraction was performed utilizing the correlation The score is calculated as wTX where W is the weight vector
matrix technique to identify and eliminate highly correlated and X is the data point , after calculating the score it is
features, thereby mitigating multicollinearity issues that linked to a logistics function / sigmoid function that brings
could adversely affect model performance. The process not down the numerical values to its probabilistic values i.e.
only optimized computational resources but also ensured the ranges between (0,1).
retention of relevant and independent features crucial for Depending upon the the probabilities of the data points, it
accurate classification. can be either classified to class A or B.
Furthermore, to streamline the training process and alleviate
the curse of dimensionality, dimensional reduction 3) SUPPORT VECTOR MACHINE ALGORITHM
techniques were applied. By reducing the number of
features while preserving the most significant information, The support vector machine algorithm, a supervised learning
these techniques enhance model interpretability and algorithm, aims to find a weight vector that determines a
generalization performance. Principal Component Analysis hyperplane that separated the data points into 2 different
(PCA) technique was employed to transform the feature classes with maximum possible separation also called
space into a lower-dimensional subspace, capturing the gamma margin separation [17].
essential variance in the data while minimizing information Given the dataset {(x1,y1), (x2,y2),….(xn,yn)} where each yi
loss. belongs to either class A (LOS) or class B (NLOS) & the
Subsequently, four distinct classification algorithms were goal is to classify a new test data point xtest in
trained on the preprocessed dataset: Support Vector class A (LOS) or class B(NLOS) . The
Machine (SVM), Random Forest, k-Nearest Neighbors objective function is such that, yi(sgn(wTxi +b))
(KNN), and Logistic Regression. SVM, renowned for its ≥ 1.
ability to handle complex decision boundaries, was
employed to classify the data into distinct LOS and NLOS 4) RANDOM FOREST CLASSIFICATION
categories by maximizing the margin between them. ALGORITHM
Random Forest, a versatile ensemble learning method,
utilized a multitude of decision trees to achieve robust It is a hierarchical node based classification algorithm where
classification performance through ensemble averaging. each node is set to have a threshold for its decision
KNN, a non-parametric algorithm, classified data points boundary and the final classification is dependent upon the
based on the majority class among their nearest neighbors, leaf node of the forest starting from the root node [18].
offering simplicity and effectiveness in handling multi-class Given the dataset {(x1,y1), (x2,y2),….(xn,yn)} where each yi
classification tasks. Logistic Regression, a widely-used belongs to either class A (LOS) or class B (NLOS) & the
linear model, provided probabilistic predictions by modeling goal is to classify a new test data point xtest in class A (LOS)
the probability of a data point belong to a particular class or class B (NLOS) .
[14].
By leveraging these diverse classification techniques, each EXPERIMENTAL RESULT
with its unique strengths and characteristics, the research
aimed to develop robust models capable of accurately To assess the performance of the trained classification
distinguishing between LOS and NLOS signal conditions in models, rigorous testing procedures were employed,
indoor environments, thereby advancing the reliability and leveraging metrics such as the confusion matrix, F-score,
performance of wireless communication systems. precision, recall, and accuracy. The confusion matrix
The above algorithms are discussed in detail below: provided a comprehensive summary of the classification
results, delineating the true positives, true negatives, false
1) KNN Algorithm positives, and false negatives for each class (LOS and
NLOS). By analyzing the confusion matrix, insights into the
The naïve Supervised machine learning algorithm model’s ability to classify instances correctly of both LOS
specifically used for classification and based on the fact that and NLOS scenarios were gained, facilitating a nuanced
point can be classified into a particular class if its nearest understanding of their strengths and limitations.
data points also belong to that class itself [15]. The distance Furthermore, metrics including the F-score, precision, recall,
is calculated as Euclidian distances here. and accuracy were computed to quantify the models'
Given the dataset {(x1,y1), (x2,y2),….(xn,yn)} where each yi performance across various evaluation criteria. The F-score,
belongs to either class A (LOS) or class B (NLOS) & the a harmonic mean of precision and recall, offered a stabilized
goal is to classify a new test data point xtest in class A (LOS) assessment of the models' potency in correctly identifying
or class B (NLOS) . both positive and negative instances. By rigorously
analyzing these metrics, comprehensive insights into the
2) LOGISTICS REGRESSION CLASSIFICATION performance of the classification models in distinguishing
The probabilistic classification model which uses the score between LOS and NLOS signal conditions were obtained,
and linking function to calculate the probability of a thereby informing further refinements and optimizations to
particular class [16]. increase the reliability and efficacy of wireless
Given the dataset {(x1,y1), (x2,y2),….(xn,yn)} where each yi communication systems in real-world applications.
belongs to either class A (LOS) or class B (NLOS) & the
goal is to classify a new test data point xtest in class A (LOS)
or class B (NLOS) .

Authorized licensed use limited to: Chang'an University. Downloaded on March 18,2025 at 13:42:36 UTC from IEEE Xplore. Restrictions apply.
 Logistic 86% 86% 86% 86%
Regression

KNN 80% 82% 80% 80%

CONCLUSION

The research paper has explored the classification of LOS

and NLOS signal conditions in indoor environments using
machine learning techniques. Through rigorous
experimentation and analysis, paper demonstrated the
effectiveness of Support Vector Machine, Logistic
Regression, KNN, and Random Forest classifiers in
distinguishing between these signal propagation scenarios
[15]. Furthermore, paper identified areas for improvement,
including the exploration of deep learning techniques and
modern classification algorithms, to enhance the accuracy
and reliability of classification tasks in future endeavors. By
continuing to innovate and refine our methodologies, it
strive to contribute to the advancement of wireless
communication systems, ensuring robust performance
across a wide range of applications.
Figure 1: Comparison Metrics of different algorithms. Looking ahead, leveraging deep learning techniques and
modern classification algorithms presents a promising
Based on the evaluation metrics provided, the classifiers can avenue for further improving the accuracy and efficacy of
be ranked in increasing order of performance as follows: signal classification in distinguishing between LOS and
KNN, Logistic Regression, Support Vector Machine NLOS conditions. Deep learning models, such as
(SVM), and Random Forest. convolutional neural networks (CNNs) and recurrent neural
KNN achieved the lowest accuracy among the classifiers, networks (RNNs), offer the capability to automatically learn
but it showed relatively balanced precision, recall, and F1 intricate patterns and representations from raw data,
score. Despite its simplicity, KNN performed adequately, potentially capturing nuanced features that may be
especially considering its ease of implementation and overlooked by traditional machine learning approaches.
interpretability. Transfer learning, a technique where pre-trained models are
Logistic Regression demonstrated slightly improved adapted to new tasks, could expedite the development
performance compared to KNN across all metrics, process by taking the advantage of the knowledge gained
showcasing higher accuracy, precision, recall, and F1 score. from large-scale datasets in related domains. Additionally,
Support Vector Machine (SVM) exhibited further advancements in ensemble learning methods, such as
enhancement in performance compared to Logistic stacking and gradient boosting, could further enhance
Regression, with all metrics showing consistent classification performance by combining the strengths of
improvement. multiple base models. Furthermore, attention mechanisms
Finally, Random Forest emerged as the top-performing and other innovations in neural network architectures can
classifier, showcasing the highest accuracy, precision, recall, help focus on relevant information and improve model
and F1 score among the evaluated models. interpretability. By embracing these cutting-edge techniques
In summary, while all classifiers demonstrated competence and continuously refining the classification pipeline, we aim
in distinguishing between LOS and NLOS signal conditions, to achieve even greater accuracy and reliability in
Random Forest emerged as the most effective model, identifying LOS and NLOS signal conditions, thereby
followed by Support Vector Machine, Logistic Regression, advancing the capabilities of wireless communication
and KNN, in increasing order of performance. These systems in diverse real-world applications.
insights can lead the selection of the best suitable classifier
based on the specific needs and constraints of the REFERENCES
application at hand. [1] C. Huang et al., "Machine Learning-Enabled LOS/NLOS Identification
for MIMO Systems in Dynamic Environments," in IEEE Transactions on
TABLE II: Different Algorithms and their percentage Wireless Communications, vol. 19, no. 6, pp. 3643-3657, June 2020, doi:
10.1109/TWC.2020.2967726

Classifier Accuracy Precision Recall F1 Score [2] S. Mahato, M. Goswami, S. Kundu and A. Bose, "GPS-GLONASS
Random 89% 89% 89% 89% combined position performance accuracy for Indian Defence
Forest Application," 2021 2nd International Conference on Range Technology
(ICORT), Chandipur, Balasore, India, 2021, pp. 1-4, doi:
Support 87% 87% 87% 87% 10.1109/ICORT52730.2021.9581541.
Vector
Machine

Authorized licensed use limited to: Chang'an University. Downloaded on March 18,2025 at 13:42:36 UTC from IEEE Xplore. Restrictions apply.
[3] H. Liu, H. Darabi, P. Banerjee and J. Liu, "Survey of Wireless Indoor
Positioning Techniques and Systems," in IEEE Transactions on Systems,
Man, and Cybernetics, Part C (Applications and Reviews), vol. 37, no. 6,
pp. 1067-1080, Nov. 2007, doi: 10.1109/TSMCC.2007.905750.].

[4] Zeynep Turgut, Gulsum Zeynep Gurkas Aydin, Ahmet Sertbas,”Indoor

Localization Techniques for Smart Building Environment”,Procedia
Computer Science,Volume 83,2016,Pages 1176-
1181,https://doi.org/10.1016/j.procs.2016.04.242.

[5] D. Liu et al., "WLAN new technologies in IEEE 802.11," 2014 XXXIth
URSI General Assembly and Scientific Symposium (URSI GASS),
Beijing, China, 2014, pp. 1-4, doi: 10.1109/URSIGASS.2014.6929347.

[6] R. Joseph and S. B. Sasi, "Indoor Positioning Using WiFi

Fingerprint," 2018 International Conference on Circuits and Systems in
Digital Enterprise Technology (ICCSDET), Kottayam, India, 2018, pp. 1-3,
doi: 10.1109/ICCSDET.2018.8821184.

[7] Chih-Ning Huang, Chia-Tai Chan, “ZigBee-based indoor location

system by k-nearest neighbor algorithm with weighted RSSI”,Procedia
Computer Science,Volume 5,2011,Pages 58-
65,https://doi.org/10.1016/j.procs.2011.07.010.

[8] E. Essa, B. A. Abdullah and A. Wahba, "Improve Performance of

Indoor Positioning System using BLE," 2019 14th International Conference
on Computer Engineering and Systems (ICCES), Cairo, Egypt, 2019, pp.
234-237, doi: 10.1109/ICCES48960.2019.9068142.

[9] M. A. Al-Ammar et al., "Comparative Survey of Indoor Positioning

Technologies, Techniques, and Algorithms," 2014 International Conference
on Cyberworlds, Santander, Spain, 2014, pp. 245-252, doi:
10.1109/CW.2014.41.

[10] W. -L. Chin and M. -J. Lu, "Signal Detection for Mobile
Devices," 2019 7th International Conference on Information and
Communication Technology (ICoICT), Kuala Lumpur, Malaysia, 2019, pp.
1-4, doi: 10.1109/ICoICT.2019.8835236.

[11] Alarifi A., Al-Salman A., Alsaleh M., Alnafessah A., Al-Hadhrami S.,
Al-Ammar M.A., Al-Khalifa H.S.,”Ultra wideband indoor positioning
technologies: Analysis and recent advances”2016, Sensors Switzerland, 16
(5), art. no. 707,doi: 10.3390/s16050707.

[12] Y. Dong, T. Arslan and Y. Yang, "Real-Time NLOS/LOS

Identification for Smartphone-Based Indoor Positioning Systems Using
WiFi RTT and RSS," in IEEE Sensors Journal, vol. 22, no. 6, pp. 5199-
5209, 15 March15, 2022, doi: 10.1109/JSEN.2021.3119234.

[13] Bregar, Klemen & Hrovat, Andrej & Mohorcic, Mihael. (2016). NLOS
Channel Detection with Multilayer Perceptron in Low-Rate Personal Area
Networks for Indoor Localization Accuracy Improvement.

[14] Caruana, Rich, and Alexandru Niculescu-Mizil. "An empirical

comparison of supervised learning algorithms." In Proceedings of the 23rd
international conference on Machine learning, pp. 161-168. 2006.

[15] N. U. Sadad, A. Afrin and M. N. I. Mondal, "Binary Classification

using K-Nearest Neighbor Algorithm on FPGA," 2021 International
Conference on Computer, Communication, Chemical, Materials and
Electronic Engineering (IC4ME2), Rajshahi, Bangladesh, 2021, pp. 1-4,
doi: 10.1109/IC4ME253898.2021.9768439.

[16] X. Zou, Y. Hu, Z. Tian and K. Shen, "Logistic Regression Model

Optimization and Case Analysis," 2019 IEEE 7th International Conference
on Computer Science and Network Technology (ICCSNT), Dalian, China,
2019, pp. 135-139, doi: 10.1109/ICCSNT47585.2019.8962457.

[17] Q. Wang, "Support Vector Machine Algorithm in Machine

Learning," 2022 IEEE International Conference on Artificial Intelligence
and Computer Applications (ICAICA), Dalian, China, 2022, pp. 750-756,
doi: 10.1109/ICAICA54878.2022.9844516.

[18] Schonlau, M., & Zou, R. Y. (2020). The random forest algorithm for
statistical learning. The Stata Journal, 20(1), 3-
29. https://doi.org/10.1177/1536867X20909688.

Authorized licensed use limited to: Chang'an University. Downloaded on March 18,2025 at 13:42:36 UTC from IEEE Xplore. Restrictions apply.