A New Multi-Dimensional Hybrid Deep Neural Network Based Spectrum Inference for Cognitive Radio Network

In wireless communications, cognitive radio (CR) technology has significantly enhanced radio spectrum efficiency. Spectrum sensing is a key process in CR along with other major functions namely spectrum decision, sharing and mobility. Minimizing the processing delays, energy consumption of these functions and enhancing spectrum utilization is a major challenge. Spectrum inference has been proposed as a viable solution to overcome these problems. Many machine learning-based spectrum inference techniques using artificial neural networks (ANNs) and deep neural networks have been proposed in literature. In this paper we aimed to determine whether hybrid deep neural network based spectrum inference model outperform single model in time and frequency domains for spectrum occupancy dataset. Radial basis function (RBF) neural network tend to excel in extracting spatial features of spectrum data whereas bidirectional long short-term memory (BiLSTM) work very well for temporal dependencies of this data. Spectrum dataset exhibit both short-and-long term temporal/spectral dependencies. In this paper we have proposed spectrum inference based on a hybrid deep neural network RBF and BiLSTM. The proposed algorithm has been simulated using real time spectrum measurement data with time dimension ranging from (1 to 7 days), spectral range (0.7 GHz to 2.7 GHz) across three geographically varying locations Pune, Solapur and Kalaburagi in India. Hybrid deep neural network integration of RBF and BiLSTM is built, tested and compared with single models LSTM, BiLSTM for accuracy and speed. The hybrid method has outperformed single models to achieve Precision, Recall, F1 scores of 0.9959, 0.9575, 0.9763 respectively and training time improvement of 57.60% for GSM and whole band in frequency and time dimensions.