Under-Sampled UWB NLOS/LOS Channel Classification using Machine Learning

This paper investigates the ability of different machine learning (ML) algorithms to classify ultra-wideband channels into line-of-sight and non-line-of-sight channels. The examined algorithms include convolutional neural network, K-nearest neighbors, logistic regression, long-short term memory, stochastic gradient descent, support vector machine, and ensemble ML. For consistency and generality, multiple experimental and simulated datasets are used. We examine the classification performance with the raw data of the channel impulse response (CIR) or some extracted features. The promising features are energy, peak to lead delay, kurtosis, mean excess delay, RMS delay spread, and skewness, among others. Due to the ultrawide bandwidth used, the associated sampling rate is very high, and the required processing is costly. This work demonstrates that we can work with down-sampled data without deteriorating the feature extraction or the classification performance. Under-sampling the experimental data by a factor of 10 still guarantees high classification accuracy. This also reduces the complexity and accelerates the classification process. Ensemble ML algorithms are recommended because they provide the largest accuracy for most of the considered datasets. They achieve ~ 90% classification accuracy for dataset-C and IEEE802.15.4a) and ~ 80% accuracy for dataset-B when the CIR is downsampled by a factor of 20.