Link failure prediction in LoRa networks

Low-Power Wide-Area Networks allow long-range communication with Internet-of-Things (IoT) devices. The LoRa (Long Range) is one of the most widely used LP WAN technologies thanks to unlicensed bands, extensive availability of low-cost transceivers and the standardized LoRaWAN networking stack. The messages in LP WANs are rarely extended (e.g. few times a day). The radio link characteristics may change significantly between two consecutive data transmissions due to changes in interference level, collisions with other transmissions or the presence of obstacles on the radio wave transmission path. It makes it very hard to estimate the optimal transmission parameters and delays the detection of any problems in the transmission. We investigate the problem of prediction of the link state, namely the prediction if a packet transmitted by a LoRa arrives at the gateway. We analyze the time sequence of packets delivered to LoRa gateways from a city-wide telemetry network, consisting of 15 gateways and more than 1 mln data points. We evaluate two prediction methods, based on moving averages of received signal strength (RSSI) measurements and using Support Vector Machines (SVMs). Using a time series of 20 days, it is possible to achieve the specificity and sensitivity of the prediction above 0.6 with the SVM method. The mean accuracy of the SVM method is approximately 0.6. Using the comparison of RSSI moving averages, we could predict the link failure with a lower sensitivity, but very high specificity - e.g. prediction of 10% of link breaks with 91% specificity.