Channel Quality Prediction for TSCH Blacklisting in Highly Dynamic Networks: A Self-Supervised Deep Learning Approach

Abstract

Cross-Technology Interference (CTI) is a severe issue for wireless networks operating in the license-free ISM bands. In particular, CTI can significantly affect the performance of low-power wireless links used in the Internet-of-Things (IoT). The IEEE 802.15.4 standard adopts a channel hopping scheme in its Time-Slotted Channel Hopping (TSCH) mode as a means to mitigating the adverse effect of CTI. However, the indiscriminate procedure by which TSCH nodes hop over different channels can suffer from severe interference in specific channels. Adaptive channel blacklisting is a technique to alleviate this issue by leaving out low-quality channels from the hopping list. To enable an effective blacklisting, especially in highly varying networks, an accurate real-time prediction of the quality of all available channels is of paramount importance. Previous studies rely on the past records of the channels as an indication of their quality in near future. Evidently, such approaches cannot extend to highly dynamic environments. This paper present a self-supervised approach for training deep neural networks capable of predicting the future behaviour of the frequency channels. The trained models can then substitute the quality assessment approaches in blacklisting schemas. Considering in-vehicle wireless networks as a target application, we evaluate this idea using a real-world experimental dataset, consisting of three measurement scenarios inside a moving vehicle. The experimental results show that using the proposed technique for TSCH blacklisting significantly improves the reliability of networks experiencing such highly dynamic interference and performs at least as good as the existing channel assessment methods in low-interference conditions.