A time-frequency feature prediction network for time-varying radio frequency interference

The time-varying radio frequency interference has strong nonlinear dynamic characteristics, which is difficult to be predicted by linear method effectively, making the anti-interference decision without sufficient information support. To solve this problem, a recurrent neural network for spectrum prediction based on time-frequency correlation features is proposed. A sliding window is used to characterize the two-dimensional correlation of time-frequency series, and the spectrum prediction problem is transformed into a problem similar to spatiotemporal sequence prediction. A gradient bridge structure across time frames is added to reduce the attenuation of the gradient in the long time and multi-level network propagation. The training efficiency and network performance are improved by the loss function with better matching. Simulation and experimental results verify the validity of the network prediction results.