Abstract
In this paper, we explore the spectrum inference to achieve the spectrum occupancy in advance through analyzing the historical spectrum. We have conceived an offline-online cooperative framework. Specifically, the hyperparameters can be achieved on an offline way, which will be used for online prediction. Moreover, based on the accuracy of online spectrum inference, the hyperparameters can be further optimized relying on specifically designed grid search and K-fold cross-validation combined method in an iterative manner. We present a long short-term memory (LSTM) aided spectrum occupancy prediction method, relying on adaptive threshold quantization aided data preprocessing (ATQ-DP). To be specific, first, the captured spectrum data may be quantized by the adaptive thresholds in order to lesson the influence of noise imposed on them, where the thresholds are obtained by kernel density estimation (KDE) method. Then, LSTM will be activated to perform spectrum prediction based on the quantized data, thus, future spectrum occupancy can be inferred in advance. Additionally, performance evaluations show that the accuracy of spectrum inference is always better than that of the LSTM aided spectrum inference relying on the traditional fixed threshold quantization aided data preprocessing (FTQ-DP), where the FTQ-DP is used for comparison purposes.
Highlights
Nowadays, many emerging technological services need to be provided through wireless channels, leading to increasing demand for spectrum resources [1]
We propose an ATQ-DP scheme based on kernel density estimation (KDE) [14] to make the quantization results of power spectral density (PSD) values closer to the true spectrum occupancy, which will make the prediction results more reliable and will be of benefit to prediction accuracy
We evaluate fixed threshold quantization aided data preprocessing (FTQ-DP) based long short-term memory (LSTM) models as a baseline, the evaluation results show that the prediction accuracy of the proposed ATQ-DP based LSTM models always outperform the baseline FTQ-DP based LSTM models
Summary
Many emerging technological services need to be provided through wireless channels, leading to increasing demand for spectrum resources [1]. In [12], an LSTM network was adopted to predict real-world spectrum occupancy with fixed threshold quantization in time-frequency domain. In [13], an LSTM network was applied to predict real-world power spectral density (PSD) values in time domain with uniform quantization. Differing from [11]–[13], we conceive a joint offline-online spectrum prediction framework, consisting of multiple adaptive threshold quantization aided data preprocessing (ATQ-DP) based LSTM models. We propose an ATQ-DP scheme based on kernel density estimation (KDE) [14] to make the quantization results of PSD values closer to the true spectrum occupancy, which will make the prediction results more reliable and will be of benefit to prediction accuracy.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
