A Spectrum Prediction Network Based on Time-Frequency Correlation Fusion Mechanism

With the development of modern communication technology, the electromagnetic spectrum environment is becoming increasingly complex. Frequency Spectrum Prediction is an important technology in Cognitive radio (CR). It plays a significant role in spectrum user authorization management, spectrum resource scheduling and spectrum interference avoidance. Traditional spectrum prediction methods are susceptible to noise and its performance is limited when the amount of spectral data is large. Therefore, we designed a Time-Frequency Correlation Capture module as a network branch for correlation feature extraction. We also designed the multi-scale Decoder module to build a spectrum prediction model based on Sequence-to-Sequence architecture deep network. The dataset is generated by using software radio equipment to collect electromagnetic environment around the campus. As verified by the measured spectrum data, our method improves significantly in spectrum multi-channel prediction compared with the traditional method and existing networks.