Machine learning-based motion type classification from 5G data

Abstract. To improve the quality of their services and products, nowadays every industry is using artificial intelligence and machine learning. Machine learning is a powerful tool that can be applied in many applications including wireless communications. One way to improve the reliability of wireless connections is to classify motion type of the user and hook it with beamforming and beam steering. With the user equipment’s motion type classification ability, the base station can allocate proper beamforming to the given class of users. With this motivation, the studies of ML algorithms for motion classification is conducted in this thesis. In this work, the supervised learning technique is used to predict and classify motion types using the 5G data. In this work, we used the 5G data collected in 4 different scenarios or classes which are (i) Walking (ii) Standing (ii) Driving and (iv) Drone. The data is then operated on for cleaning and feature engineering and then is fed into different classification algorithms including Logistic Regression Cross Validation (LRCv), Support Vector Classifier (SVC), k-nearest neighbors (KNN), Linear Discriminant Analysis (LDA), AdaBoost, and Extra Tree Classifier. Upon analyzing the evaluation metrics for these algorithms, we found that with the accuracy of ~99% and log-loss of 0.044, Extra Tree Classifier performed better than others. With such promising results, the output of classification process can be used in another pipeline for resource optimization or hooked with hardware for beamforming and beam steering. It can also be used as an input to a digital twin of radio to change its variables dynamically which will be reflected in the physical copy of that radio