DSIL: An Effective Spectrum Prediction Framework Against Spectrum Concept Drift

Abstract

Predicting spectrum plays an importance role in cognitive networks, which is the key to address the issue of spectrum scarcity. Deep learning methods for spectrum prediction have attracted significant interests because of the exceptional accuracy. However, when dealing with radio frequency (RF) measurements from real data traffic, the precise distribution of the measurements is often unknown, making model mismatch an inevitable occurrence. This is known as spectrum concept drift, which presents a formidable obstacle for traditional deep learning adapt to the dynamic spectrum environment. Considering spectrum concept drift, we proposed Deep Spectrum Incremental Learning (DSIL) method, a two stage framework including Concept Drift Detection module and Adaptive Spectrum Prediction module. In the first stage, we analysis concept drift detector mechanism and propose an effective spectrum concept drift method by leveraging Hoeffding drift detection method with averaging (HDDM-A). In the second stage, we propose Spectrum Incremental Learning Triple Net (SILTN) for spectrum incremental learning. SILTN, consisted of Multilayer Perceptron (MLP), ConvGRU and ConvLSTM, can effectively extract spectrum spatial and temporal features, and thus, improve spectrum prediction performance. Lastly, we introduce an Adaptive Spectrum Prediction Training (ASPT) method, designed to help SILTN achieve a better balance between past spectrum prediction tasks and incoming spectrum prediction tasks after finetune. The experimental results demonstrate that the DSIL framework can effectively address the issue of concept drift in common deep learning models for spectrum prediction. To the best of our knowledge, this is the first work considering spectrum concept drift detection and corresponding solution.