Decisão de espectro em redes de sensores sem fio empregando aprendizado de máquina

Abstract

Wireless Sensor Networks (WSNs) use ISM (Industrial, Scientific and Medical) spectrum bands, which are currently overloaded due to various technologies such as WLANs, WSNs and Bluetooth. Because of the overhead, WSNs suffer with great loss of information, causing high waste of energy. Therefore, in order to reduce such losses, WSNs must employ clever methods, such as Cognitive Radio (CR) to select the best channel, coexisting with other networks that use the ISM band. One of the CRs implementation paradigms, which has been highly adopted by the literature, employs Machine Learning (ML), which creates a method to choose the best channel, process which is also known in the CR area as Spectrum Decision. Despite the good performance of the existing proposals, most of them have high computational costs, which prevent their use on platforms with limited resources. Another limitation is the fact that most of the works consider a small set of ML algorithms. Moreover, the majority of studies are validated only through simulations and empirical data, which reduces the reliability on the effectiveness in real scenarios. This study evaluates an extensive set of ML algorithms for channel selection in low-resource WSN platforms. The best algorithms are implemented on real sensor nodes, experimentally demonstrating their effectiveness. Finally, we show a comparison of the proposed algorithms with fixed channel and energy-based Spectrum Decision methods. The obtained results show that ML-based RC methods enhance the overall communication performance, when compared with the implemented baselines, reducing the number of retransmissions, and thus reducing the delay.