Machine Learning-Aided Classification Of LoS/NLoS Radio Links In Industrial IoT

Wireless sensors and actuators networks are an essential element to realize industrial IoT(IIoT) systems, yet their diffusion is hampered by the complexity of ensuring reliable communication in industrial environments. A significant problem with that respect is the unpredictable fluctuation of a radio-link between the line-of-sight (LoS) and the non-line-of-sight (NLoS) states due to time-varying environments. The impact of linkstate on reception performance, suggests that link-state variations should be monitored at run-time, enabling dynamic adaptation of the transmission scheme on a link-basis to safeguard QoS. Starting from the assumption that accurate channel-sounding is unsuitable for low-complexity IIoT devices, we investigate the feasibility of channel-state identification for platforms with limited sensing capabilities. In this context, we evaluate the performance of different supervised-learning algorithms with variable complexity for the inference of the radio-link state. Our approach provides fast link-diagnostics by performing online classification based on the analysis of the envelope-distribution of a single received packet. Furthermore, the method takes into account the effects of the limited sampling frequency, bit-depth, and moving average filtering, which are typical to hardware-constrained platforms. The results of an experimental campaign in both industrial and office environments show promising classification accuracy of LoS/NLoS radio links. Additional tests indicate that the proposed method retains good performance even with low-resolution RSSI-samples available in low-cost WSN nodes, which facilitates its adoption in real IIoT networks.