Wireless Spectrum Analysis Based On Cloud Computing

With the rapid development of mobile communication technology and the continuous improvement of spectrum monitoring network,radio monitoring stations have accumulated a large amount of historical monitoring data.The rapid increase of data volume not only brings storage pressure to computer hardware,but also challenges to data management and data analysis.This dissertation collects 9KHz~3GHz spectrum monitoring data through long-term spectrum observation in a certain region of sichuan province,and cloud computing technology is combined with the relevant theories of data mining to predict and analyze the frequency band occupancy data.It has important theoretical value and practical significance for the allocation of radio spectrum resources and the management of monitoring stations.Therefore,the dissertation focuses on the problem of how to dig out the frequency rule of spectrum resources from the massive historical monitoring data.My main research is:1.Aiming at the characteristics that the frequency band occupancy data is time sequence.Autoregressive Integrated Moving Average(ARIMA)model was used to conduct fitting analysis of multiple frequency band occupancy data,and the fitting results were evaluated from multiple perspectives.Finally,the paper analyzes the limitations of ARIMA model.2.Aiming at the limitations of ARIMA model,ARIMA model is introduced into Kalman filtering algorithm to construct ARIMA-Kalman hybrid model.This method can not only overcome the limitations of ARIMA model,but also well fit the frequency occupation data,and has a good ability to predict the unknown time.3.Aiming at the problem that both ARIMA model and ARIMA-Kalman hybrid model are linear models,which cannot capture the nonlinear relationship in data.To solve this problem,Long Short-Term Memory(LSTM)neural network is used to extract the nonlinear relation in the data,and two kinds of nonlinear information extraction schemes are designed.In the first scheme,LSTM network is used to extract the linear and nonlinear relations in the data directly.In the second scheme,ARIMA-Kalman mixed model was first used to extract the linear relationship existing in the data,and then LSTM network was used to extract the residual residuals of the mixed model in a non-linear way.The feasibility of the two schemes is verified by analyzing multiple data sets.4.Spark platform was used for parallel training of ARIMA-Kalman hybrid model,and data and model were designed respectively from two aspects of data parallelization and model parallelization.The results show that Spark platform can significantly improve the training speed of the model.The effectiveness of the methods are validated by simulations.The results show that compared with the other three methods,the ARIMA-Kalman hybrid model performs well in both algorithm complexity and prediction of unknown time,and is worthy of recommendation in actual practice.