Time-series data optimized AR/ARMA model for frugal spectrum estimation in Cognitive Radio

Abstract

Abstract Wideband and agile Spectrum Estimation (SE) is a fundamental component of the Cognitive Radio (CR) system. However, CR systems generally utilize the classical sensing techniques for SE due to heteroscedasticity of the available spectrum. Unfortunately, analysis of the Time-series data for SE using a testbed is rare to find out. A novel Goodness-of-Fit (GoF) based accurate SE technique for CR system has been proposed in this work involving Time-series data samples generated from Field Programmable Gate Array (FPGA) based Wireless open Access Radio Protocol (WARP) testbed having a sampling frequency of 40 MHz. Anderson-Darling (AD) rejection based Null-Hypothesis testing has been employed to implement the CR system within a frequency range of 9 kHz to 10 MHz. 1 MHz sinusoidal signal has been generated by the testbed for digital transmission/reception through Radio Board 1 and 3. Statistical parameters like Mean Square Error (MSE), Final Prediction Error (FPE), Loss Function and Fit(%) of the received samples adjudicate the Convex optimization of the data length. Akaike Information Criteria (AIC) and Bayesian Information Criteria (BIC) are responsible for the selection of the Auto Regressive Moving Average (ARMA) (3,2) model for optimal signal processing. Finally, the Power Spectral Density (PSD) confirms the superiority of the proposed work with the most optimized data length and lag order in real-time. Computation of complexities of the proposed algorithms also indicates a parsimonious choice of the model.