Predictive energy detection for inferring radio frequency activity

Next generation cognitive radar/radio systems rely on dynamic spectrum access (DSA) to adaptively and ef-ficiently utilize the radio frequency (RF) spectrum. Such technology must detect, predict, and avoid channels occupied by RF interference. Conventional spectrum sensing methods may fail to determine signal occupancy states during transition periods. Predicting RF activity reduces the probability of interference during such transition periods and improves the overall efficiency of DSA schemes. This work employs a one-step ahead prediction approach to determine future busy or idle states through linear support vector regression (SVR). Supervised learning forecasts future signal energy which then acts as a decision statistic to determine occupancy in a sub-band of interest. The scheme's prediction accuracy is evaluated with respect to input signal-to-noise ratio (SNR) and RF activity as a function of mean busy/idle time. Generalizing RF activity as an alternating renewal process allows exponential random variables to generate simulated data for SVR training and testing. The results show that this approach predicts RF activity with high accuracy over various signal traffic statistics and SNRs. Prediction accuracy is also evaluated with respect to the expected busy/idle transitions given activity statistics.