随着无线通信的发展和电磁环境的日益复杂，对智能通信系统的需求越发显著。根据已有的研究成果，实际无线环境中存在的非高斯噪声导致通信环境更加复杂，基于高斯噪声建模的无线通信算法在非高斯噪声下性能下降。研究存在非高斯噪声情况下的智能通信系统具有重要的意义。针对上述问题本论文选择α-稳定分布模型，对复杂通信环境中能够高效、可靠、自适应传输的智能通信系统进行了研究。首先验证非高斯噪声对基于高斯建模的无线通信性能带来的影响，从而引出无线通信系统在非高斯噪声背景下需要解决的几个关键技术，主要包括非高斯噪声背景下的频谱检测、非高斯噪声背景下的信道估计以及智能通信系统的核心—智能决策引擎。然后结合复杂电磁环境下的通信需求和已有的研究成果，利用深度强化学习算法构建具备学习和决策能力的“新通信系统”—智能通信系统，并着重介绍了将深度Q网络(Deep Q-Net，DQN)应用于智能决策引擎建模的方法。仿真结果验证了智能决策引擎的可行性，α-稳定分布噪声下与成熟的自适应调制编码(Adaptive Modulation and Coding，AMC)算法对比，获得了更好的决策效果，为智能通信在物理层的研究提供了理论支撑和新的思路。 首先，针对α-稳定分布噪声下的频谱检测算法自适应能力差的问题提出了一种基于动态阈值的分数低阶矩(Fractional Low-order Moment，FLOM)检测算法。在分数低阶矩检测算法的基础上，引入了一个平衡参数来动态地调整检测门限以适应噪声背景的变化。该方法降低了检测算法的复杂度，并且依据实际检测序列的统计量获得的检测门限更加精确。仿真结果表明，本文提出的频谱感知算法相对于已有的分数低阶矩检测算法检测概率得到了有效提升，且依据实际序列计算的实时检测门限使信算法具有良好的自适应能力。 其次，为降低α-稳定分布噪声下的信道估计算法对先验知识的严格依赖性，改进稀疏信道估计算法中损失函数f(〓)的分段门限计算方式，并保持对α-稳定分布噪声中的大幅值脉冲的抑制能力。同时提出了一种基于畸变约束的群稀疏信道估计新算法，利用二次复杂性框架在递归最小二乘(Recursive Least Squares，RLS)群稀疏信道估计算法中引入畸变约束函数来抑制非高斯噪声带来过大的估计误差，结合混合范数提高了对非高斯噪声下信道群稀疏特性的发掘能力。仿真结果表明，改进的稀疏信道估计算法降低复杂度的同时维持了估计性能，基于畸变约束的群稀疏信道估计算法在α-稳定分布噪声下可以获得小于-18dB最小均方偏差(Mean Square Deviation，MSD)的精确结果，满足智能通信系统的需求。 最后，针对复杂电磁环境对智能通信系统的需求，本文结合当前通信系统的优势和深度强化学习的特点设计了一种新的智能决策引擎，在此基础上给出了一种具备自我学习和决策能力的智能通信系统结构。智能决策引擎的核心是DQN，本文创新性地提出了基于离线知识库和在线更新样本的交叉样本策略的学习方案，长短期记忆(Long Short-Term Memory，LSTM)神经网络依据自身的记忆能力保证了更好地学习和决策。智能决策引擎基于频谱感知结果选择通信频段，并根据通信的反馈信息进行训练和学习以逐渐适应当前通信环境，给出合理的通信方案，获得可靠稳定的传输性能。在相同的仿真条件下通过仿真对比了智能决策引擎和当前的AMC技术的决策方案获得的吞吐率性能，本文提出的智能决策引擎在α-稳定分布噪声下给出了更好的决策方案，验证了提出的智能决策引擎的可行性。 关键词：决策引擎；DQN；α-稳定分布噪声；频谱感知；信道估计

 The rapid development of wireless communication and increasingly complicated electromagnetic environment pose unprecedented demands on intelligent communication systems. Non-Gaussian noise in practical wireless environment leads to more complicated electromagnetic environment as described in the existing research results. And the performance of the communication algorithms based on Gaussian noise decrease under the non-Gaussian noise. It is meaningful to study the intelligent communication system under non-Gaussian noise. In this paper, the intelligent communication system is studied to obtain more reliable and stable communication performance and α-stable distribution model is selected to describe the complex non-Gaussian noise. Firstly, the key technologies to be solved are given based on the influence of α-stable distribution noise on the performance of communication algorithms, and spectrum sensing, channel estimation and the design of intelligent decision engine are all considered. Then, a new communication system with abilities of learning and decision is constructed based on Deep Reinforcement Learning (DRL), named intelligent communication system. And DQN (Deep Q-Net) is applied to the modeling of intelligent decision engine. The simulations verified the feasibility of the intelligent decision engine. Throughput rate performance of different decision results given by mature Adaptive Modulation and Coding (AMC) algorithm and intelligent decision engine are compared. Obviously, intelligent decision engine made a better decision, which provides theoretical support and new ideas for the research of intelligent communication in the physical layer. The key technologies of intelligent communication systems under α-stable distributed noise can be summarized as follows: Firstly, a fractional low-order moment (FLOM) detection algorithm with dynamic threshold is proposed for spectrum sensing under α-stable distributed noise due to the poor adaptive ability of existing algorithms. Balance parameter is introduced to the FLOM detection algorithm to adjust the detection threshold dynamically to adapt to the complicated communication environment. It can reduce the repeated calculation for detection threshold every time. And the detection threshold is more accurate as it is obtained based on the statistics of the actual detection sequence, which reduces the complexity of the detection algorithm. It can be concluded that the spectrum sensing algorithm proposed in this paper has improved the detection probability according to the simulation results, and the detection threshold based on the actual sequence makes the algorithm have good adaptability. Secondly, the channel estimation algorithms under α-stable distributed noise are studied. Calculation method for segmentation threshold of the loss function is improved in the sparse channel estimation algorithm, which reduces the strict dependence on prior knowledge and maintains the ability to suppress large-value pulses in α-stable distribution noise. Besides, a new algorithm for group sparse channel estimation based on distortion constraint is proposed. The quadratic complexity framework applied the distortion constraint function to the Recursive Least Squares (RLS) group sparse channel estimation algorithm, which aimed at suppressing the excessive estimation error caused by non-Gaussian noise and improving the ability to explore the group sparse characteristics of channel under non-Gaussian noise. According to the simulation results, it can be concluded that the improved sparse channel estimation algorithm reduces the complexity while maintaining the estimation performance. The group-sparse channel estimation algorithm based on distortion constraint can obtain smaller minimum mean square deviation (MSD) under α-stable distribution noise, which meet the demand of intelligent communication systems Finally, a new intelligent decision engine is designed by combining the advantages of current communication systems with the characteristics of deep enforcement learning (DRL). And an intelligent communication system with capabilities of self-learning and decision-making is proposed, which addresses the urgently demand for intelligent communication systems in complex electromagnetic environments. Deep Q-Net (DQN) is the core of the intelligent decision engine, and we have innovatively proposed a learning scheme based on the cross-sample strategy of offline knowledge and online update sample. Long Short-Term Memory (LSTM) neural network guarantees better learning and decision results based on its own memory ability. The intelligent decision engine selects the communication frequency based on the spectrum sensing results, and trains and learns according to the feedback information of the communication to adapt to the current communication environment gradually. And a reasonable communication scheme will be given to obtain reliable and stable transmission performance. Throughput performance obtained by the intelligent decision engine and AMC decision-making scheme are compared under the same simulation conditions. A better decision scheme given by intelligent decision engine verifies the effectiveness of the proposed intelligent decision engine in the α-stability distribution noise. Keywords: decision engine, DQN, α-stable distribution noise, spectrum sensing, channel estimation