A Seq2Seq Learning Approach for Link Quality Estimation Based on System Metrics in WSNs

Link quality estimation (LQE) is a fundamental problem in Wireless Sensor Networks (WSNs). LQE is not only a prerequisite for efficient routing but also significantly impacts the energy consumption of sensor nodes. Despite its importance, LQE remains an open problem due to the time-varying nature of WSNs. Existing approaches mainly rely on physical layer measurements to estimate link quality. However, considering the hardware and environment variations, modeling the correlation between physical layer measurements and link quality is a nontrivial task, rendering it difficult to obtain accurate link quality estimations. For example, our study reveals that various packet delivery rates may correspond to the same RSSI value. In this paper, we propose a novel method SeqLQE to predict link quality using system metrics (e.g., radio-on time, number of packets received) rather than physical layer measurements. We systematically design and collect runtime metrics during network operation. We then adopt a Seq2Seq learning-based model to capture the structure of correlation between link quality and system metrics. Through extensive experiments, we show that SeqLQE achieves an MSE error of 0.0226, which is 6 times better than widely used linear models.