基于成分分解的自适应滤波降噪方法研究

发布时间：2018-05-25 23:22

本文选题：信号降噪 + 经验小波变换EWT　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：降噪问题是信号处理领域中的一个经典问题,信号与噪声类型多种多样,而目前的大多降噪方法都只针对特定的信号与噪声,对适用性更广的降噪方法的研究一直是人们努力的方向。自适应滤波降噪尽管需要额外的参考噪声作为输入,但对各种噪声的适应性较强,因此十分具有研究价值。自适应成分分解方法能够根据信号自身性质将其分解成多个成分,根据噪声在不同成分中分布的不同,通过设定阈值或某种方法对各成分进行选择性的保留与重构,也被广泛用于降噪研究。为充分利用成分分解方法的自适应性,并克服LMS型自适应滤波算法某些时候收敛速度过慢且对于非平稳信号去噪效果不理想的问题,本文主要研究将自适应成分分解方法与自适应滤波器相结合的降噪方法。针对包含多种频率成分或相关度高的噪声LMS型算法收敛速度很慢降噪效果差的问题,本文引入了一种成分分解方法——经验小波变换EWT。EWT方法是一种基于对频谱划分的自适应频带分解方法,通常的频谱划分方法是根据极大极小值点寻找分割边界,为了应对频谱更加复杂的信号,本文研究了一种基于尺度空间的无参数频谱分割方法。在此基础上本文提出了基于EWT的自适应滤波降噪方法,该方法先通过EWT将噪声信号分解成若干子带,保留过程中得到的经验小波函数并用其分解混合信号,对每个噪声信号与混合信号对应的子带独立进行自适应滤波降噪,最后对每个子带的滤波结果进行累加得到最终降噪后的信号。通过实验研究了EWT中参数对分割的影响,从重构误差上看对于周期性确定性信号,其归一化误差在10-4数量级,进一步降低重构误差将能够提高降噪效果;分解后子带信号在频谱动态范围的改善也通过实验得以验证。最后进行了信号降噪实验,仿真实验可以看出该方法可以取得比直接自适应滤波更好的收敛性能,对于在粉红噪声背景下的语音信号降噪效果也更好。针对对于非平稳信号/噪声LMS型算法降噪效果有限的问题,引入另一种成分分解方法——经验模态分解EMD方法。EMD方法脱离了傅立叶分析框架,通过对信号极值包络不断取平均,迭代提取出单分量信号,能更好地反映信号局部特征,适合对非平稳信号进行分析。本文首先对两种其他研究者提出的EMD与自适应滤波结合的降噪方法进行研究,分析其中存在的一些问题,在此基础上提出采用多元经验模态分解MEMD对混合信号与参考噪声信号同步分解,分解后各IMF的平稳性得到了加强,更适合使用自适应滤波进行处理,通过降噪实验验证了方法的有效性。本文的研究初步提出了将自适应成分分解与自适应滤波结合这一新的降噪思路,从两个角度出发分别引入两种不同的成分分解方法,降噪实验的结果验证了这一思路的可行性。
[Abstract]:Noise reduction is a classical problem in the field of signal processing. There are a variety of signal and noise types, but most of the current noise reduction methods only focus on specific signals and noises. The research of more widely used noise reduction methods has been the direction of people's efforts. Adaptive filtering noise reduction needs additional reference noise as input, but it has strong adaptability to all kinds of noise, so it is of great research value. The adaptive component decomposition method can decompose the signal into several components according to its own properties. According to the different distribution of noise in different components, the adaptive component decomposition method can selectively retain and reconstruct each component by setting a threshold or a certain method. It is also widely used in noise reduction research. In order to make full use of the self-adaptability of the component decomposition method, and to overcome the problem that the LMS adaptive filtering algorithm converges too slowly at some times and the denoising effect for non-stationary signals is not satisfactory. This paper mainly studies the noise reduction method which combines the adaptive component decomposition method and the adaptive filter. In order to solve the problem that the convergence speed of noise LMS type algorithm with multiple frequency components or high correlation is very slow, the effect of noise reduction is poor. In this paper, a component decomposition method, empirical wavelet transform (EWT.EWT), is introduced, which is an adaptive frequency band decomposition method based on spectrum partitioning. In order to deal with the more complex spectrum of signals, this paper studies a non-parametric spectrum segmentation method based on scale space. On this basis, an adaptive filtering noise reduction method based on EWT is proposed. Firstly, the noise signal is decomposed into several sub-bands by EWT, and the empirical wavelet function is used to decompose the mixed signal. The sub-bands corresponding to each noise signal and the mixed signal are independently filtered and de-noised. Finally, the filtering results of each sub-band are accumulated to obtain the final de-noised signal. The effect of parameters on segmentation in EWT is studied experimentally. For periodic deterministic signals, the normalized error is in the order of 10-4 from the reconstruction error point of view, and further reducing the reconstruction error will improve the noise reduction effect. The improvement of the subband signal in the spectrum dynamic range after decomposition is also verified by experiments. Finally, the signal denoising experiment is carried out, and the simulation results show that the proposed method can achieve better convergence performance than direct adaptive filtering, and it is also better for speech signal denoising under the background of pink noise. In order to solve the problem that the noise reduction effect of the non-stationary signal / noise LMS algorithm is limited, another component decomposition method, empirical mode decomposition (EMD) method, is introduced, which breaks away from the Fourier analysis framework, and the envelope of the signal extremum is continuously averaged. Iterative extraction of single component signals can better reflect the local characteristics of the signals and is suitable for the analysis of non-stationary signals. In this paper, two kinds of noise reduction methods proposed by other researchers, which are combined with EMD and adaptive filtering, are studied, and some problems are analyzed. On this basis, the multivariate empirical mode decomposition (MEMD) is proposed to synchronously decompose the mixed signal and the reference noise signal. After the decomposition, the stability of each IMF is enhanced, which is more suitable for processing with adaptive filtering. The effectiveness of the method is verified by noise reduction experiments. In this paper, a new method of noise reduction is proposed, which combines adaptive component decomposition with adaptive filtering. Two different methods of component decomposition are introduced from two angles. The results of noise reduction experiments verify the feasibility of this idea.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN911.7

【参考文献】