Wireless Link Quality Prediction Based on Temporal Convolutional Networks and Self-Attention Fusion
Y.H. Wang, Linlan Liu · 2024
Most current deep learning-based link quality prediction methods rely on statistically derived link quality parameters over sampling periods, which makes short-term correlations in link quality difficult to capture, and the prediction task often requires multiple consecutive probing cycles, increasing energy consumption. To this end, this paper proposes a method for link quality prediction using sequences of physical layer parameters over probing cycles, including a Temporal Convolutional Network based on improved self-attention (TCNS), where a self-attention mechanism (SAM) is used to capture the time series in the global dependencies in the time series, thus improving the prediction accuracy. Experiments on laboratory-collected and public datasets show that this method outperforms other link quality prediction models. In this paper, we also investigate the effect of probe cycle length on the link quality prediction. In general, the accuracy of a link quality prediction model improves as the length of the probing cycle is lengthened. Still, a certain threshold value exists that makes the prediction performance optimal.