【作者】 徐毅琼；

【导师】 葛临东；

【作者基本信息】 解放军信息工程大学 ， 军事情报学， 2011， 博士

【摘要】 自动调制识别是软件无线电和非合作接收的关键技术之一。近年来,随着现代通信和信号处理技术的突飞猛进,通信信号体制和调制样式日趋复杂和多样,另外,信号传输的环境也变得越来越恶劣,所有这些变化都使得对通信信号识别的要求越来越高,识别难度越来越大。通信信号自动调制识别,特别是非合作接收中的自动调制识别面临的主要问题是:多径信道中的单载波与多载波信号的盲识别;参数估计精度和适用范围很难同时兼顾,获得较好的参数估计性能依赖于一定的先验信息;基于瞬时参量的调制识别方法容易受到噪声和多径衰落的影响;基带信号的调制识别方法受载波频偏和初始相位的影响大等。论文主要研究数字调制信号自动识别技术,包括信号检测、参数估计、特征提取与选择,以及分类器设计等内容,论文工作是作者所在实验室承担的军队某重点科研项目的一个子任务。论文重点致力于较少依赖先验信息的调制类型识别、与调制类型无关的参数估计、对噪声、多径衰落和频偏不敏感的调制特征提取等关键技术的研究。论文的主要工作和创新性成果是:1、针对基于时域的信号存在性检测算法易受噪声影响大的不足,本文提出了自适应的基于自相关函数的信号存在性检测算法,该算法将信号的自相关函数值,重新排序、求商,确定自适应门限,实现信号检测,该算法实现简单、运算量较小,且与调制类型、载波频率和载波初始相位无关。仿真实验表明该方法在低信噪比的性能优于传统的基于能量的检测算法,利用实际信号的测试表明该方法适应非协作接收信号的多样性。2、已有的单载波与多载波信号(OFDM)的识别方法大都针对基带信号,为解决适合中频信号的单载波与多载波的盲识别问题,本文引入Shintaro_K特征实现了单载波和OFDM信号的识别。本文在推导单载波信号与OFDM信号两次小波变换结果方差不同的基础上,提出了基于二次小波变换VAR (|| WT ||)特征的单载波与OFDM信号识别方法。实验结果证实了上述算法的可行性。在加性高斯白噪声信道下和多径衰弱信道下,上述方法与基于高阶累积量的单载波与OFDM信号识别方法进行比较。实验结果表明基于高阶累积量的识别方法需要对信号进行载波频率估计,低信噪比下识别性能不佳,且运算量大。基于Shintaro_K特征参数的分类方法,对高斯白噪声不敏感,运算复杂度最低,但抗多径性能差。基于VAR (|| WT ||)的分类方法,抗高斯白噪声和多径衰弱的能力强。3、为解决非协作通信中的参数估计问题,本文首先改进了基于子空间分解技术的盲信噪比估计算法,采用MDL准则确定信号子空间维数。仿真结果表明,本文提出的基于子空间的盲信噪比估计算法性能要明显优于传统的采用归一化比值算法的基于子空间的信噪比估计算法,该算法不需要人为设定门限值,无需先验信息,受观测数据长度影响较小。本文为了改善基于最大似然频偏估计算法估计范围小的缺点,利用载波频率粗估计结果对原信号进行移频,提出了扩大频偏估计范围的基于最大似然的频偏估计算法,并利用分段FFT算法,提出了增强精度的频偏估计算法。本文为解决与调制方式无关的符号速率估计问题,针对原基于小波变换的符号估计方法不适合基带成形信号,要求信噪比条件高等不足,提出了一种基于小波变换波形整理的盲符号速率估计算法。该算法采用MAC谱线增强算法,缩小了搜索范围,减少运算量的同时,降低毛刺对MAC谱的影响。仿真实验表明本文所提出的符号速率估计算法适用范围广,估计精度高,运算量适中。4、针对已有文献中对谱线和谱峰特征多以描述性为主,操作性不强的问题,本文给出了Pl ( n)、N l( n)离散谱线特征和N p( n)谱峰数特征的具体定义和提取方法。该方法首先确定谱线检测与谱峰搜索的范围,然后在谱线特征提取中,采用了MAC谱线增强的谱线检测算法,该谱线增强算法削弱了色噪声和频谱起伏对谱线检测的干扰,其次在谱峰搜索中提出了基于二值削波的谱峰搜索算法,该谱峰搜索算法有效地提高了FSK阶数判别的准确性。本文提出的谱线增强算法和波峰搜索算法具有广泛的适用性。本文以上述特征为基础,提出了基于时域、频域、星座图Radon变换特征和决策树(DT)的调制识别方法,该方法采用特征少,判决流程简单。5、针对已有的基于Haar小波变换的调制识别方法存在尺度选择问题,且不适合于基带成形信号的不足。本文提出了基于最优尺度Haar小波变换的调制识别方法。该方法利用调制信号Haar小波变换前后信噪比增益的最大值确定最优缩放尺度。仿真实验结果表明最优尺度下调制信号的Haar小波变换幅度阶梯状和幅度跳变最明显,所提取的Haar小波变换特征识别性能最好。本文在推导FSK、PSK和QAM信号的Morlet小波变换结果的基础上,提出了基于Morlet小波变换的调制识别方法。针对Morlet小波变换运算速度慢的不足,本文采用了快速卷积计算小波变换,提高了Morlet小波变换的运算速度。仿真实验表明与Haar小波相比,Morlet小波变换尺度选择相对简单,抗噪性能较好,且对滚降系数不敏感。6、针对非协作接收信号,无先验知识的问题,本文分析了通信信号独立分量分析(ICA)的可实现性,提出基于ICA的调制识别方法。该方法首先对通信信号进行ICA,利用特征的分类能力γ,选择有效特征子集,采用RBF神经网络分类器,实现调制识别。仿真实验验证了该方法的可行性和特征选择的有效性。为克服ICA算法运算速度慢、难收敛的不足,本文选择将FastICA算法应用到调制特征提取中,将Aitken算法和m步牛顿迭代相结合,提出了加速的M-FastICA算法。该算法在不影响识别性能的前提下,加快了收敛速度。7、针对基于ICA的调制识别方法,要求信噪比条件高的不足,本文提出了在通信信号小波域进行ICA的调制识别方法。在推导通信信号小波域进行ICA可行性的基础上,该方法利用通信信号的小波系数进行ICA,所得到的分离矩阵也是原始通信信号的分离矩阵。仿真实验表明,小波系数与原始信号相比超高斯性更强,基于WTICA的调制特征提取方法具有抗噪声能力强,收敛速度快的优点。更多还原

【Abstract】 Automatic modulation recognition is one of the key technologies, and also a hard nut to crack in the Software Radio and uncooperative receiver field. In recent years, due to rapid development of modern communication and signal processing technologies the communication systems and the corresponding modulation types are becoming more and more complicated and diverse. Besides, the channel environment is also becoming more and more deteriorated. All of above changes make modulation recognition become more and more difficult.The primary difficult problems the automatic modulation recognition must face, especially in the uncooperative environment, are summarized as follows: blind separation of single-carrier and multi-carrier signals in multipath channels; simultaneous high accuracy and wide scope of parameter estimation; dependency of estimation performance on some prior information of the signals that is not easy to obtain; sensitivity of instantaneous parameters to noise and multipath fading; uncertainty of Carrier frequency offset and initial phase which are critical in baseband recognition.This paper focuses on the key technologies of automatic digital modulation recognition including signal detection, parameter estimation, feature extraction and selection and classifier designing while laying emphasis on the following problems: recognition methods less depending on the prior information; parameter estimation independent of the modulation type; feature extraction insensitive to noise, multipath fading and frequency offset. The work finished in this paper is a part of a key engineering project of the army undertaken by the laboratory the author works with. The main work and innovative achievements obtained in this paper are summarized as follows.(1)Targeting at the problem that the time-domain-based signal detection algorithm is sensitive to noise, an adaptive signal detection algorithm based on the autocorrelation function is proposed. The algorithm set the adaptive threshold by reordering the autocorrelation function values and determining the quotient, thus is of less computational complexity and independent of modulation type, carrier offset and initial phase. Simulation results have proved the improved performances over the conventional energy-based detection algorithms under low SNR. Test on practical signals shows that the method can meet the needs caused by the diversity of the signals in uncooperative environment.(2)Targeting at the problem that the most currently available separation methods of single-carrier and multi-carrier signals (OFDM) can only be used for baseband signals, a method based on the Shintaro_K feature is proposed in this paper for IF single-carrier and OFDM signals. Besides, another method based on VAR (|| WT ||)for the same purpose is proposed by the use of the differences between the twice wavelet transform results of single-carrier and OFDM signals. Experiment results have verified the feasibility and show that the method, compared with the conventional methods in additive white Gaussian noise channel and multipath fading channel, does not need carrier offset estimation any longer and is of less computational complexity, insensitive to Gaussian white noise, however poorer anti-multipath performance.(3)With regard to parameter estimation, an improved blind SNR estimation algorithm based on subspace decomposition is proposed by the use of MDL criterion for determination of the subspace dimension. Simulation results show that the performance of the above algorithm is significantly better than the conventional ones, does not need threshold setting and without the need of prior information. To widen the estimation scope of the conventional maximum likelihood estimation algorithm, an improved algorithm based on the maximum likelihood estimation is proposed for expanding the estimation scope by roughly estimating the carrier frequency of the original signal in advance. A frequency offset estimation algorithm with enhanced accuracy is proposed based on segmented FFT. Targeting at the problem that the conventional wavelet-based symbol rate estimation methods are not suitable to the pulse shaping baseband signals and need too high SRN, a new blind symbol rate estimation algorithm based on sorting the waveforms of the wavelet transform is proposed. Because the algorithm uses the MAC spectrum line enhancement algorithm now only in a narrower search range, not only the computational complexity is remarkably reduced but also the influence of the noise pulse on the MAC spectrum is also reduced. Simulation results show that the new symbol rate estimation algorithm is of wide estimation range, high precision and moderate computational complexity.(4)As regards spectrum peak detection, an approach to define the characteristics of discrete spectral lines Pl ( n), N l( n) and peak number N p( n) along with a corresponding extraction approach are provided. Firstly, the method determines the range of spectral detection and spectral peak search, and then an enhanced spectral line detection algorithm based on MAC algorithm is employed in line feature extraction, remarkably reducing the interference of color noise and spectral fluctuation on the spectral line detection. Secondly, a spectral peak search algorithm based on binary clipping is proposed which can improve the accuracy of FSK order decision. The proposed spectral line enhancement algorithm and spectral peak search algorithm are of wider applicability and based on above algorithms a comprehensive modulation recognition scheme based on the features extracted from time domain, frequency domain and constellation Radon transform are provided. This method in fact employs less features and the decision process is also simplified.(5)Targeting at the problem of uncertainty in scale selection of the conventional recognition algorithms based on Haar wavelet transform which are not suitable to pulse shaping baseband signals a modulation recognition method based on optimal scale Haar wavelet transform is proposed. The method determines the optimal zoom level by using the maximum SNR gain before and after the transform. Simulation results show that the amplitude jumping and the ladder shapes become clearer if the optimal scale is used, thus the performances of the extracted features are remarkably improved. A modulation recognition method based on the Morlet wavelet transform is also proposed based on the derivation of Morlet wavelet transform of FSK, PSK and QAM signals. Fast convolution calculation of wavelet transform is used to speed up the Morlet wavelet transform. Simulation results show that, compared with the Haar wavelet, Morlet wavelet transform is of simplicity of scale selection, better noise immunity and less sensitivity on roll-off factor.(6)Targeting at the problem of the lack of prior knowledge in non-cooperation receiving, the feasibility of independent component analysis (ICA) towards communication signals is analyzed first, and then a modulation recognition method based on ICA is proposed. The method does ICA to signals first, then selects effective feature subset, and finally uses RBF neural network for modulation recognition. Simulation results show that the method is feasible and feature selection is effective. To speed up the algorithm and to guarantee the convergence of ICA, the FastICA algorithm, some of its key iterative steps are optimized in this paper, is applied to the modulation feature extraction and. Besides, an accelerated M-FastICA algorithm to speed up the convergence without affecting the recognition performance is also provided.(7)Targeting at the problem that the ICA-based modulation recognition method requires higher SNR, a method based on ICA in the wavelet domain is proposed. The method does ICA to the wavelet coefficients of communication signals, and the resulted separation matrix is also the separation matrix of the original communication signal. Simulation results show that signal wavelet coefficients are of more super-Gaussian compared with the original signal, and this modulation feature extraction method based on WTICA has the advantages of enhanced anti-noise performance and fast convergence.更多还原