Applying deep neural networks for duty cycle estimation
Al-Tahmeesschi, Ahmed; Umebayashi, Kenta; Iwata, Hiroki; López-Benítez, Miguel; Lehtomäki, Janne (2020-06-19)
A. Al-Tahmeesschi, K. Umebayashi, H. Iwata, M. López-Benítez and J. Lehtomäki, "Applying Deep Neural Networks for Duty Cycle Estimation," 2020 IEEE Wireless Communications and Networking Conference (WCNC), Seoul, Korea (South), 2020, pp. 1-7, doi: 10.1109/WCNC45663.2020.9120720
© 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
https://rightsstatements.org/vocab/InC/1.0/
https://urn.fi/URN:NBN:fi-fe2020111290061
Tiivistelmä
Abstract
A pro-active spectrum usage prediction is a key technique in decision making and spectrum selection for dynamic spectrum access systems. This work focuses on the estimation of the duty cycle (DC) metric to reflect spectrum usage. The prediction is formulated as a time-series regression problem. Deep neural networks (DNNs) is selected to obtain accurate predictions of channel usage. Namely, Multilayer perceptron (MLP), Long short term memory (LSTM) and a hybrid model based on convolutional neural network followed by an LSTM (CNN-LSTM) layer are selected. The hyper-parameters selection has been optimised utilising both grid search and multi-stage grid search. Moreover, in many cases, the spectrum usage is measured on a smaller time scale from the actual required one. Hence, down-sampling and averaging is required. Averaging operation results in flattening the data and losing essential features to assist DNN to predict the channel usage. We show what is the minimum required time resolution to have a pro-active prediction system. Then, we propose utilising feature engineering to improve prediction accuracy. All the proposed DNNs approaches are trained on real-life measurements. The experimental evaluation demonstrated a high potential of DNNs to learn from previous spectrum usage and accurately predict the spectrum usage. Moreover, adding input features significantly assists the system to achieve accurate predictions in a pro-active manner.
Kokoelmat
- Avoin saatavuus [31657]