Towards Enhancing Spectrum Sensing: Signal Classification Using Autoencoders
Towards Enhancing Spectrum Sensing: Signal Classification Using Autoencoders
复制标题
DOI：
10.1109/access.2021.3087113
复制
发表时间：
2021-01-01
期刊：
影响因子：
3.9
通讯作者：
Gifford, Kevin
中科院分区：
文献类型：
--
作者：
Subray, Siddhartha;Tschimben, Stefan;Gifford, Kevin
对依赖于无线电频谱的技术（例如移动的通信和物联网）的需求一直呈指数级增长。因此，提供对无线电频谱的接入变得越来越重要。不断增长的无线业务和可用频谱的日益稀缺保证了无线电频谱的有效管理。与此同时，机器学习（ML）正变得无处不在，并因其识别模式和辅助决策过程的能力而在许多领域得到应用。最近，机器学习算法已被用于解决无线通信领域中的挑战，例如无线电频谱感测，并且已经显示出比传统感测方法（例如能量检测）更好的性能。频谱感测是一种用于检测和识别在无线电频谱的相同频带中传输的不同无线信号的方法，对于改善动态频谱共享至关重要，其具有增强相同频带中不同无线技术的共享和共存并最终提高频谱效率的潜力。为此，本研究评估了不同类型的自动编码器，如深度、变分和长短期记忆（LSTM）自动编码器，以识别和区分LTE和Wi-Fi传输。我们的目标是研究不同类型的自动编码器在由LTE和Wi-Fi信号（IEEE 802.11ax和IEEE 802.11ac）组合组成的I/Q数据集上的性能，以确定最佳算法的复杂度，精度和召回率。对于此分类任务，我们的模型实现了高达99.9%的准确率和88.1%的召回率。此外，这些模型的最短训练时间约为47秒，适合在动态RF环境中进行在线学习和部署。
The demand for technologies relying on the radio spectrum, such as mobile communications and IoT, has been growing exponentially. As a consequence, providing access to the radio spectrum is becoming increasingly more important. The ever-growing wireless traffic and the increasing scarcity of available spectrum warrants efficient management of the radio spectrum. At the same time, machine learning (ML) is becoming ubiquitous and has found applications in many fields for its ability to identify patterns and assist with decision-making processes. Recently, machine learning algorithms have been used to address challenges in the wireless communications domain, such as radio spectrum sensing, and have shown better performance than traditional sensing methods, such as energy detection. Spectrum sensing, a method for detecting and identifying different wireless signals being transmitted in the same band of the radio spectrum, is crucial for improving dynamic spectrum sharing, which has the potential to enhance sharing and coexistence of different wireless technologies in the same frequency band and ultimately improve spectrum efficiency. To this end, this research evaluates different types of autoencoders, such as deep, variational and Long Short-Term Memory (LSTM) autoencoders, to identify and differentiate between LTE and Wi-Fi transmissions. The goal is to investigate the performance of the different types of autoencoders on an I/Q dataset consisting of LTE and a combination of Wi-Fi signals (IEEE 802.11ax and IEEE 802.11ac) for the classification task in terms of complexity, precision, and recall to identify the best algorithm. Our models have achieved up to 99.9% precision and 88.1% recall for this classification task. Additionally, with a shortest training time of approximately 47 seconds, the models are suitable for online learning and deployment in a dynamic RF environment.