摘要
随着无线设备及其应用呈现指数级增长,以及各种无线技术的出现,电磁频谱环境呈现环境多域、态势多维、应用多样、行为多变、信号密集的复杂特性.通过对海量数据进行即时有效的分析和处理,实现有效的频谱资源管控,对于频谱高效利用和电磁致胜显得尤为重要.以机器学习为代表的智能技术方法为电磁频谱数据分析提供了新的发展机遇,目前其已经应用于无线网络、频谱管理、资源管理等场景.然而,将机器学习应用于电磁频谱空间存在许多挑战.因为电磁频谱空间环境任务动态变化、高质量标记样本稀缺、频谱决策对高实时性的需求、知识利用和迁移困难等难点,现有机器学习算法难以很好地适用于电磁频谱空间的智能认知和决策.以干扰识别的研究为例,传统的技术包括信号特征提取和模式识别两个阶段.随着计算机算力的进步,深度学习技术凭借强大的特征表征能力逐渐占据主导地位.然而电磁频谱空间环境和数据往往存在未知、动态变化等不确定性因素,深度学习难以快速学习和适应未观测到的任务,其泛化性较差,且依赖大量高质量标记样本和人工调整超参数进行离线训练.因此,尽管深度学习在大多数研究场景中展现出了传统机器学习难以望其项背的结果,传统的机器学习图像识别模型仍然在某些场景中具备优势.在干扰识别中,若根据环境和任务特征动态地调度传统机器学习算法和深度学习算法及其超参数,则可以避免单一算法的缺陷,使总体性能得到改善.由此可见,提出一个能够自适应动态环境与任务、不断积累新知识、对错误样本数据具有鲁棒性、部分可解释的学习范式,对于从感知智能向认知智能跨越、迈向通用人工智能具有重要的研究意义和实际应用价值.本文从频谱特性出发,提出了电磁频谱空间机器学习新范式:认知学习,试图在“认知”的层面使用机器学习技术.认知学习范式包括离线训练、在线执行、在线结果反馈、离线自学习过程,对应大脑认知机制的执行控制和内省过程.离线训练探索算法结构和超参数与环境和任务的匹配关系,在线时能够有效地根据实际环境执行快速决策,选择合适的算法和超参数并输出学习结果,同时利用学习结果的反馈传递进行离线自学习的自我优化,形成离线-在线-离线的结构.具体来说,认知特征提取模块建立环境和任务特征与算法类型和超参数之间的匹配关系,从而得到合适的算法类型和超参数;记忆模块调用具体算法和超参数值,进行算法的重构;学习网络模块根据数据执行所选择的算法,输出在线学习结果;认知评估模块对在线学习结果进行内省,调整记忆中的频谱数据库与认知案例库,更新最合适的算法和超参数.在离线阶段,根据当前的学习结果判断是否将其作为新的认知案例存储至记忆空间;认知控制模块重新训练算法和超参数的选择,从而不断优化算法和超参数选择性能.此外,本文揭示了认知学习中的动态匹配定理和最优匹配定理,并根据没有免费午餐定理的思路完成了定理证明.动态匹配定理论证了最优算法和超参数的存在性与动态性,即对于任意环境和任务,总存在一组性能最优的算法和超参数.不同的环境与任务对应不同的最佳算法和超参数,并且本文提出的框架可以为动态环境和任务选择出这组算法和超参数.本文还揭示了最优匹配定理,即当算法和超参数选择的知识增加时,候选的算法类型和超参数搜索空间减小,能够选择到最合适的算法类型和超参数的概率提升.本文设计了一种面向干扰识别的认知学习方案.通过认知特征提取,基于差异性的任务需求和信号模态,对不同问题采取不同算法和超参数产生测试结果,生成认知案例库.将大量认知案例作为训练样本训练算法和超参数选择的神经网络.对于神经网络结构,采用了径向基函数神经网络,其优势是隐藏单元的数量不需要手动设置.当各种问题特征的训练样本数量增加时,隐藏层的大小可以根据近似线性相关准则自动扩展,缓解了神经网络中灾难性遗忘的问题.本文进行了不同数据集和不同识别要求情况下的干扰识别仿真.仿真结果表明,所提框架能适应动态变化的环境和任务、能通过自学习提升性能、能缓解错误标记样本的影响,对电磁频谱空间机器学习的发展具有启发式意义.认知学习的发展仍有进步空间,未来需要从多模态计算、知识的可解释性、大规模任务、学习架构的可扩展性等方向寻求突破.
With the exponential growth in the number of wireless devices and wireless applications,and the emer⁃gence of various wireless technologies,the electromagnetic spectrum environment has presented complex characteristics such as multi-domain environments and situations,diverse applications,changeable behaviors,and dense signals.Spectrum resource control through immediate and effective analysis and processing of massive data is vital for efficient spectrum utili⁃zation and electromagnetic spectrum dominance.Intelligent technology represented by machine learning methods provides new opportunities to analyze electromagnetic spectrum data.Machine learning methods have been applied to wireless net⁃works,spectrum management,resource management,and other scenarios.There are many challenges when applying machine learning to the electromagnetic spectrum environment.Because of the dynamically changing environment and tasks,the scar⁃city of high-quality labeled samples,real-time spectrum decision-making,and the difficulty of knowledge utilization and trans⁃fer in the electromagnetic spectrum environment,existing machine learning algorithms are not ideal for intelligent cognitive decision-making.Taking the research on interference recognition as an example,the traditional methods include two stages:signal feature extraction and pattern recognition.With the growth in computer computing power,deep learning techniques gradually replace traditional methods because of their powerful feature characterization capability.However,the electromag⁃netic spectrum environment and data often have uncertainties caused by unknown,dynamic changes.Adapting deep learning to unobserved tasks quickly is challenging,indicating poor generalization capability.Furthermore,it is limited by its reliance on a large number of high-quality labeled samples and manual adjustment of hyperparameters for offline training.Thus,al⁃though deep learning has far outperformed traditional methods in most research scenarios,traditional machine learning-based image recognition models still have advantages in some scenarios.In interference recognition,if traditional machine learning algorithms and deep learning algorithms and their hyperparameters can be dynamically scheduled according to the environ⁃ment and task characteristics,the shortcomings of using a single algorithm can be avoided,resulting in improved overall per⁃formance.It is necessary to propose a learning paradigm that continuously adapts to dynamic environments and tasks,accu⁃mulates new knowledge,is robust to erroneous sample data,and is partially interpretable.The leap from perceptual intelli⁃gence to cognitive intelligence toward general artificial intelligence has research significance and practical value.Given the characteristics of the electromagnetic spectrum,this paper proposes a new paradigm of machine learning meth⁃od in the electromagnetic spectrum environment—cognitive learning,attempting to use machine learning techniques at a“cog⁃nitive”level.It includes offline training,online execution,online feedback and offline self-learning processes,corresponding to the executive control and introspective processes of the brain's cognitive mechanisms.Offline training explores the match⁃ing relationship between the algorithm type and hyperparameters with the environment and the task.During the online pro⁃cess,it can perform decision-making based on the actual environment,select appropriate algorithms and hyperparameters,and output learning results.The feedback of learning results can be used for self-optimization in offline self-learning,form⁃ing an offline-online-offline structure.The cognitive feature extraction module establishes matching relationships between features of environments and tasks and algorithm types and hyperparameters to obtain the appropriate algorithm types and hy⁃perparameters.The memory module calls specific algorithms and hyperparameter values for algorithm reconstruction.The learning network module executes the selected algorithm based on the input data and outputs online learning results.The cognitive evaluation module introspects the online learning results and adjusts the data in the memory database and cognitive case base,and updates the most appropriate algorithm and hyperparameters.In the offline stage,the current learning results are judged whether to be stored as a new cognitive case in the memory.The cognitive control module re⁃trains the selection of algorithms and hyperparameters to continuously optimize the performance of selecting algorithms and hyperparameters.Moreover,we reveal and confirm the dynamic matching theorem and the optimal matching theorem in cognitive learn⁃ing according to the No Free Lunch Theorem.The dynamic matching theorem states the existence and dynamics of optimal algorithms and hyperparameters.For any environment or task,there is always a set of algorithms and hyperparameters with optimal performance,which differs for different environments and tasks.The proposed cognitive learning paradigm can se⁃lect appropriate algorithms and hyperparameters.The optimal matching theorem states that as the knowledge of algorithm and hyperparameter selection increases,the search space of candidate algorithm types and hyperparameters decreases,and the probability of selecting the most suitable algorithm type and hyperparameters increases.A cognitive learning scheme for radio frequency interference identification is designed.Different algorithms and hy⁃perparameters are adopted for different problems to generate test results and a cognitive case base based on the differential task requirements and signal modalities acquired by cognitive feature extraction.A large number of cognitive cases are used as training samples to train the neural network for algorithm and hyperparameter selection.A radial basis function neu⁃ral network is used for the neural network structure,which has the advantage that the number of hidden units does not need to be set manually.The size of hidden layers can be automatically expanded according to the approximate linear correlation criterion when the number of training samples for various problem features increases,alleviating the problem of catastroph⁃ic forgetting in neural networks.We performed interference recognition simulations with different datasets and different recognition requirements.The simulation results demonstrate that the proposed paradigm can adapt to dynamically changing environments and tasks,im⁃prove performance by self-learning,and alleviate the impact of mislabeled samples.The cognitive learning paradigm has a pioneering significance for developing machine learning methods applied to the electromagnetic spectrum environment.There remains room for improvement in cognitive learning.We must seek future research breakthroughs in multimodal computing,knowledge interpretability,large-scale tasks,and learning architecture scalability.
作者
阮天宸
吴启晖
赵世瑾
周福辉
黄洋
RUAN Tian-chen;WU Qi-hui;ZHAO Shi-jin;ZHOU Fu-hui;HUANG Yang(Key Laboratory of Dynamic Cognitive System of Electromagnetic Spectrum Space,Ministry of Industry and Information Technology,Nanjing University of Aeronautics and Astronautics,Nanjing,Jiangsu 211106,China)
出处
《电子学报》
EI
CAS
CSCD
北大核心
2023年第6期1430-1442,共13页
Acta Electronica Sinica
基金
国家自然科学基金重点项目(No.62231015)。
关键词
电磁频谱空间
机器学习
认知学习
算法选择
干扰识别
electromagnetic spectrum environment
machine learning
cognitive learning
algorithm selection
inter⁃ference identification
作者简介
阮天宸,男,南京航空航天大学电子信息工程学院/集成电路学院在读博士生.主要研究方向为认知学习、信号源识别与搜索.E-mail:ruantianchen@nuaa.edu.cn;通讯作者:吴启晖,男,南京航空航天大学特聘教授、副校长,教育部“长江学者”特聘教授,南京航空航天大学电子信息工程学院/集成电路学院教授、博导.主要研究方向为认知科学与应用、电磁空间频谱认知智能管控、无人机认知集群.E-mail:wuqihui@nuaa.edu.cn;赵世瑾,女,南京航空航天大学电子信息工程学院/集成电路学院在读博士生.主要研究方向为认知强化学习、视觉导航.E-mail:shijin_zhao@nuaa.edu.cn;周福辉,男,国家优秀青年基金获得者,南京航空航天大学电子信息工程学院/集成电路学院教授、博导.主要研究方向为频谱智能管控和资源鲁棒优化、认知智能与知识图谱.E-mail:zhoufuhui@nuaa.edu.cn;黄洋,男,南京航空航天大学电子信息工程学院/集成电路学院副教授.主要研究方向为强化学习、深度学习、最优化理论、协同智能、频谱管控.E-mail:yang.huang.ceie@nuaa.edu.cn。
