基于BP算法和遗传算法的自适应噪声抵消器/

2019-04-13 21:22:40

遗传算法 算法 GA BP noise



一、引言
自适应噪声抵消技术是一种能够很好的消除背景噪声影响的信号处理技术,应用自适应噪声抵消技术,可在未知外界干扰源特征,传递途径不断变化,背景噪声和被测对象声波相似的情况下,能够有效地消除外界声源的干扰获得高信噪比的对象信号。
从理论上讲,自适应干扰抵消器是基于自适应滤波原理的一种扩展,简单的说,把自适应滤波器的期望信号输入端 改为信号加噪声干扰的原始输入端,而它的输入端改为噪声干扰端,由横向滤波器的参数调节输出以将原始输入中的噪声干扰抵消掉,这时误差输出就是有用信号了。
在数字信号采集、处理中,线性滤波是最常用的消除噪声的方法。线性滤波容易分析,使用均方差最小准则的线性滤波器能找到闭合解,若噪声干扰类型为高斯噪声时,可达到最佳的线性滤波效果。在实际的数字信号采集中,叠加于信号的噪声干扰往往不是单一的高斯噪声,而线性滤波器所要求的中等程度噪声偏移,使线性滤波器对非高斯噪声的滤波性能下降,为克服线性滤波器的缺点,往往采用非线性滤波器,所以本文采用神经网络对信号进行滤波处理。
二、基于BP算法和遗传算法相结合的自适应噪声抵消器
在本文中,作者主要基于自适应噪声对消的原理对自适应算法进行研究,提出了一种新的算法,即BP算法和遗传算法相结合的自适应算法。作者对BP网络的结构及算法作了一个系统的综述,分析了BP算法存在的主要缺陷及其产生的原因。传统的BP网络既然是一个非线性优化问题,这就不可避免地存在局部极小问题,网络的极值通过沿局部改善的方向一小步进行修正,力图达到使误差函数最小化的全局解,但实际上常得到的使局部最优点。学习过程中,下降慢,学习速度缓,易出现一个长时间的误差平坦区,即出现平台。
通过对遗传算法文献的分析、概括和总结,发现遗传算法与其它的搜索方法相比,遗传算法(GA)的优点在于:不需要目标函数的微分值;并行搜索,搜索效率高;搜索遍及整个搜索空间,容易得到全局最优解。所以用GA优化BP神经网络,可使神经网络具有进化、自适应的能力。
BP-GA混合算法的方法出发点为:
(1)利用BP神经网络映射设计变量和目标函数、约束之间的关系;
(2)用遗传算法作实现优化搜索;
(3)遗传算法中适应度的计算采用神经网络计算来实现。
BP-GA混合算法的设计步骤如下:
(1)分析问题,提出目标函数、设计变量和约束条件;
(2)设定适当的训练样本集,计算训练样本集;
(3)训练神经网络;
(4)采用遗传算法进行结构寻优;
(5)利用训练好的神经网络检验遗传算法优化结果。若满足要求,计算结束;若误差不满足要求,将检验解加入到训练样本集中,重复执行3~5步直到满足要求。
通过用短时傅立叶信号和余弦信号进行噪声对消性能测试,在单一的BP算法中,网络的训练次数、学习速度、网络层数以及每层神经元的节点数都是影响BP网络的重要参数,通过仿真实验可以发现,适当的训练次数可以使误差达到极小值,但是训练次数过多,训练时间太长,甚至容易陷入死循环,或者学习精度不高。学习速度不能选择的太大,否则会出现算法不收敛,也不能选择太小,会使训练过程时间太长,一般选择为0.01~0.1之间的值,再根据训练过程中梯度变化和均方误差变化值确定。
基于梯度下降原理的BP算法,在解空间仅进行单点搜索,极易收敛于局部极小,而GA的众多个体同时搜索解空间的许多点,因而可以有效的防止搜索过程收敛于局部极小,只有算法的参数及遗传算子的操作选择得当,算法具有极大的把握收敛于全局最优解。使用遗传算法需要决定的运行参数中种群大小表示种群中所含个体的数量,种群较小时,可提高遗传算法的运算速度,但却降低了群体的多样性,可能找不出最优解;种群较大时,又会增加计算量,使遗传算法的运行效率降低。一般取种群数目为20~100;交叉率控制着交叉操作的频率,由于交叉操作是遗传算法中产生新个体的主要方法,所以交叉率通常应取较大值,但若过大的话,又可能破坏群体的优良模式,一般取0.4~0.99;变异率也是影响新个体产生的一个因素,变异率小,产生个体少,变异率太大,又会使遗传算法变成随机搜索,一般取变异率为0.0001~0.1。由仿真结果得知,GA与BP算法的混合算法不论是在运行速度还是在运算精度上都较单纯的BP算法有提高,去噪效果更加明显,在信噪比的改善程度上,混合算法的信噪比针对傅立叶信号提高了16db左右,针对余弦信号提高了23db左右。
三、结论
用短时傅立叶信号和余弦信号进行噪声对消性能测试,通过分别使用单一的BP算法和混合算法作比较发现遗传算法具有很强的处理能力和优化能力,用它优化BP神经网络的权值,与原有单一的BP算法相比,可以节省大量的学习和计算时间,而且提高了信噪比。





1. Introduction
The technology of the adaptive noise canceller is on signal processing ,which can clear off background noise effect better. When the disturbance source from environment is not available ,its way to transmit is a constant change ,background noise and tested sound wave are similar .Using the technology of the adaptive noise canceller ,we can clear off disturbance from environment effectively better and acquire high SNR of tested signal .
Theoretically, the adaptive noise canceller is a kind of expansion based on the adaptive filter. In brief, we change the expected signal input of the adaptive filter into the primitive input end that a signal adds the noise and interferes. And its input end changes into an interferes noise end, we may offset primitive input noise of interfere by adjusting parameter of transversal filter, at this moment the error outputted is a useful signal.
In the collection and processing of digital signal, the linear filter is a method of the most frequently used elimination noise. It is easy to analyse linear filter, because using minimum criterion of variance it can find close solving. If the type of noise is one gauss of noises, it can reach best filtering result. While gathering in the real digital signal, the noise disturbance superposing signal is not single gauss noise. But the linear filter required medium noise of wave filter is skew, it makes linear filter to non-gausses noise filtering performance drop. In order to overcome the shortcoming of linear filter, we often use non-linear filter. So in this paper the method of filter processing to signal is by neural network.
2. The adaptive noise canceling based on the back propagation algorithm and genetic algorithm joined
In this paper, the author studied the adaptive algorithm on the basis of the adaptive noise canceller principle mainly. The paper proposed a kind of new algorithm that a designed method to the adaptive noise canceller based on back propagation network optimization with genetic algorithms. The author made a summary to the BP network structure and algorithm, had analysed its main defect and reason which it produced. Since traditional BP network is a non-linear optimization question, the local minimum problem will exist unavoidably in this. The extreme value of the network , through revising along an improved a little step of direction, try hard to reach the global solutions of making the error function minimize ,but in fact we obtain the local optimal solution. In the learning process, it is slow to drop, it is slow to study the speed, it is easy to appear in a long-time smooth district of error, namely appear in the platform.
Through an analysis of GA document, generalization and summary, we find that the GA compares with other search methods, The advantage of the genetic algorithm (GA) lies in: do not need differential value of goal function; run side by side and search for, it is with high efficiency to search for; it search for space entirely and spread all over, it is easy to reach the global optimal solutions. So we optimize BP neural network with GA, can make the neural network have evolving, adaptive ability .
The inspiration of the structure optimization method of BP-GA is:
(1) Utilize BP neural network mapping the relation of design variable and goal function, restriction;
(2) Realize optimizing and searching for with the GA;
(3) The calculation of fitness function in the GA adopts the neural network to realize.
The design step of BP-GA mixed algorithms is:
(1) Analyse the question, put forward goal function, design the variable and restraining the condition;
(2) Establish the proper training sample, calculate and train the sample;
(3) Train the neural network;
(4) Adopt the GA to seek the optimal structure;
(5) Use the trained neural network to check up the GA results, if they meet the demands, it is over to calculate; if the error does not meet the demands, will make the examine results add to the trained samples, then carry out 3-5 steps repeatedly until meeting the demands;
Through using short-term Fourier signal and cosine signal to offset testing, in single BP, the training number of times, the learning speed , the network layers and determination the number of each layer, which are important parameter influenced BP network. Through simulation experiment we can find proper training times may make the error minimum, but if training times are too much, training time will be too long, so it is easy to fall into bad circulation, or learning precision is decrease. Learning speed can’t choose too heavy, otherwise the algorithm appear non-convergence, can’t choose too light either, the training time will be too long. So as generally we choose the value of 0.01~0.1, then we confirm the value according the gradient change and the square error change in the course of training.
BP based on gradient principle, searches for only alone in solution space, and it is extremely easy to convergence to the local minimum, but numerous individual of GA search for a lot of solution in space at the same time, therefore GA can prevent effectively to convergence to the local minimum during the searching course. Nothing but the parameter and the GA operation of operator of algorithm are chose properly, the algorithm has great assurance to convergence to the global solution. Using genetic algorithm need operation parameter that determine population size show including individual quantity in the population, when the populations are small, it can raise the operation speed of GA , but it has reduced the variety of the population, can't find out the optimum solutions; When the populations are bigger , it can increase calculations , and make GA operational efficiency decrease . so we choose the value of 20-100 as the populations generally; The probability crossover is controlling the frequency of operated alternately, because it produces the new individual's main method in the GA to operate alternately, so the probability crossover should usually choose greater value , but if too big , it may destroy the optimum mode of the population, generally we choose the value of 0.4-0.99; The probability mutation(pm) is also one factor that it influence the new individual producing, if pm is little, produce individual little, if pm is too heavy, it can make GA turn into at random searching for, so we choose the value of 0.0001-0.1 generally. Learned by the simulation result, the mix algorithm of
BP and GA algorithms raise no matter at operation speed or operation precision compared with single BP, and it reach obvious result at improvement degree of SNR, the SNR of the mix algorithms raise about 16db to the Fourier signal, and improve about 23db to the cosine signal
3. Conclusion
The author made the noise offset testing with the Fourier signal and the cosine signal, through using single BP algorithm and the mixed algorithms relatively we find the GA algorithm has strong processing and optimize ability separately, optimizing the weight of BP with it ,compared with original single BP algorithms , the mixed algorithms can save a large amount of learning and calculation time, and improve SNR.