摘要
通过分析非线性函数各输入参数对函数值的影响,发现当输入参数间无相互作用时,表征其灵敏度的傅立叶振幅主要集中在基频上.基于该发现,提出一种基于扩展傅立叶振幅灵敏度方法的神经网络隐含层神经元快速修剪算法.其实质是通过计算神经网络隐含层各神经元输出的指定基频上的傅立叶振幅,求取隐含层神经元对神经网络输出的灵敏度.根据各神经元的灵敏度,按照一定的准则削减冗余神经元,获得紧凑的神经网络结构.将提出的神经网络结构修剪算法用于污水水质参数化学需氧量(COD)的软测量过程中,实验结果与扩展傅里叶振幅灵敏度算法相比,在修剪效果相同的情况下,其运行时间得到明显减小.
By analyzing the impact of the input factors on the output value in nonlinear function,it is suggested that when the input factors are independent,the Fourier Amplitudes which showing sensitivity values are relied mainly on the fundamental frequency.As a result,a fast pruning algorithm for the hidden neurons in the neural network is proposed based on the Fourier amplitude sensitivity test method.In essence,the Fourier amplitudes on the assigned frequencies of the hidden layer outputs are computed.Then the sensitivity of each hidden neuron to the neural network output is obtained.Finally,the redundant hidden neurons are pruned according to their sensitivity values to obtain a network with compact structure.The propose method is used in the soft measurement for Chemical Oxygen Demand（COD）,which is a quality parameter of waste water.The experimental result shows that our proposed method is much faster than the Fourier amplitude sensitivity test method.The remaining neurons are the same after pruning for the two methods.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2010年第4期830-834,共5页
Acta Electronica Sinica
基金
国家863高技术研究发展计划(No.2009AA04Z155
No.2007AA04Z160)
国家自然科学基金(No.60873034
No.60674066)
教育部博士点基金(No.200800050004)
北京市自然科学基金(No.4092010)
