ERI: An Adaptive Incremental Deep Learning Architecture for Real-Time Inference of RF Signals in Dynamic Spectrum Sharing Environments

5G及更高的下一代无线通信技术将提供更高的容量、更快的速度和全球连接，改变我们的生活、工作、学习和娱乐方式。仅5G的发展就会影响我们的经济和劳动力，在未来十年为美国国内生产总值贡献1.4至1.7万亿美元，并在2034年前创造460万个与5G相关的就业机会。无线设备的戏剧性增长和不断增长的功能和性能已经挤占了电磁频谱。5G及以后的动态频谱共享可以满足频谱稀缺、流量空前、服务质量更好的需求。例如，未经许可和获得许可的频段，如2.4-5 GHz工业、科学和医疗频段、6 GHz无线电频段、60 GHz毫米波频段，正在共享用于商业和科学用途。然而，动态共享频谱给在民用、政府和国防之间自主、可靠和安全地共享频谱带来了额外的挑战。因此，了解周围的无线信号对于支持共享频谱上的无线用户共存至关重要。本项目致力于开发一种自适应增量式深度学习结构，以在动态频谱共享环境中实时推断射频信号。这项拟议的研究将向电气和电子工程师协会(IEEE)标准P1900.8提供关于评估机器学习频谱感知模型性能的标准的建议。研究和教育活动与外展活动的整合将扩大科学、技术、工程和数学领域中未被充分代表的群体的参与，并通过基于服务学习的课程开发有利于当地社区的合作。该项目的目标是开发一个增量深度学习架构，以适应和高效地检测、分类和解调动态频谱共享环境中的无线频率信号，并具有在线学习能力。所提出的增量式深度学习结构将(1)自适应地学习广泛的无线通信场景，从具有有限已知信号或场景的小数据集开始，然后增量地学习新的信号或场景，并以不间断的在线方式更新深度学习网络，以重新训练整个网络和中间人来标记信号；(2)推进频谱自学习，包括频谱感知、信号分类和射频参数实时表征；(3)用自适应增量学习改进基于深度学习的信号解调，它可以取代传统的基于块的解调过程，并以高度灵活性保持相同的性能。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

5G and beyond, the next generation of wireless communication technology, will provide higher capacity, faster speeds, and world-wide connectivity, transforming the way we live, work, learn and entertain. The development of 5G alone impacts our economy and workforce by contributing $1.4 to $1.7 trillion to US gross domestic product over the next decade and create 4.6 million 5G-related jobs through 2034. The dramatic growth and ever-increasing functionality and performance of wireless devices has crowed the electromagnetic spectrum. Dynamic spectrum sharing in 5G and beyond can meet the demands of scarce spectrum, unprecedented traffic, and better quality of service. For example, unlicensed and license-aided bands, such as 2.4 - 5 GHz industrial, scientific and medical bands, 6 GHz radio frequency bands, 60 GHz millimeter wave bands, are being shared for commercial and scientific use. However, dynamically sharing spectrum poses additional challenges to share autonomously, reliably, and securely among civilian, government, and defense. Therefore, learning surrounding wireless signals is essential to support wireless user coexistence over shared spectrum. This project focuses on developing an adaptive incremental deep learning architecture to infer radio frequency signals in dynamic spectrum sharing environments in real time. The proposed research will provide recommendations to Institute of Electrical and Electronics Engineers (IEEE) Standard P1900.8 on criteria for evaluating the performance of machine learned spectrum awareness models. The integration of research and education activities and outreach activities will broaden participation of underrepresented groups in science, technology, engineering, and mathematics fields and benefit local community collaborations through service learning-based curriculum development.The goal of this project is to develop an incremental deep learning architecture for adaptively and efficiently detecting, classifying, and demodulating radio frequency signals in dynamic spectrum sharing environments in real time with online learning capabilities. The proposed incremental deep-learning architecture will (1) adaptively learn a wide range of wireless communication scenarios, starting from a small dataset with limited known signals or scenarios and then incrementally learning new signals or scenarios as well as updating the deep learning network in an online manner without interruption to re-train the whole network and a man-in-the-middle to label the signal; (2) advance self-learning the spectrum including spectrum sensing, signal classification and radio frequency parameter characterization in real-time; and (3) improve deep learning-based signal demodulation with the adaptive incremental learning, which can replace conventional block-based demodulation processes and preserve the same performance with high flexibility.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.