Deep learning algorithm has become a main method in the image and voice recognition field because of its excellent characteristics under the process of big data in a natural environment.For the purpose of solving the problems that it is difficult to select an appropriate structure of the deep learning network.In this paper,we get an in-depth explore at the structure characteristics of deep learning network and propose a method the preferred method of process to help us to select the structure of the deep learning network.This method uses the variation of the deep reconstruction error of single layer in the deep learning network as a criterion.By introducing the pare to principle in economics into our network,the threshold of the deep reconstruction error of single layer can be determined conveniently and quickly,and deep learning network optimization is given.The preferred method of process can effectively solve the uncertainty of the layers' number and the nodes' number per layer when set up the network.It is validated by experiment that this method has large amount of good effectiveness under a variety of heart sound databases and this method has a certain universal applicability to be used in other studies.As a kind of valuable biological signals,heart sounds can reflect the beating of the heart and are closely related to the health of the heart or the human's body.Heart sounds has been widely used in health monitoring and identity recognition and it has attracted the attention of researchers both domestic and abroad.In this paper,we use the preferred method of process to optimize and build a heart sound deep learning network firstly.Then we design a heart sounds deep belief network by combining the classifier BP neural network with the heart sound deep learning network,the core of it.Compared to other belief network in the same category,this heart sounds deep belief network has a lower error recognition rate and the average error recognition rate is only about 10％.The deep belief network has little limit for input data.So referencing to the previous method of heart sound recognition,an effective method to greatly improve the recognition rate is to extract the feature vector of input data,and then used it as input of the deep belief network.In this paper the average error recognition rate can be reduced to 3 ％ when the network is optimized for a new deep belief network.The new deep belief network is based on the original system data which optimized by the orthogonal wavelet transform and then extract heart sound energy characteristics.The research in this paper has a positive significance to improve the ability to process data of heart sound recognition algorithm under normal natural circumstances.Building better depth of heart sound recognition system,is the next step research work.%深度学习算法因其在自然环境下对大数据处理的优良特性已成为图像、语音识别方面的主流算法.为解决深度学习网络结构选择困难的问题,文中深入探究了深度学习网络的结构特性,提出了一种进程择优法来帮助深度学习网络结构的选择,可方便、快速地给出深度学习网络的优选范围.经实验验证,此方法在多种数据库下都有良好效果,方法具有一定的普适性.而心音作为一种生理信号,反映了人体心脏的跳动情况,与人体心脏的健康息息相关,在心音分类识别、健康鉴定中得到广泛的应用.文中首先使用进程择优法来优选、构建出一种心音深度学习网络,再以心音深度学习网络为核心,加入BP神经网络作为分类器,设计出了一种心音深度信任网络.该网络相比同类其它层次结构的深度信任网络拥有更低的误识别率,平均误识别率在10％左右.特别是将原系统优化为融合心音能量特征输入的心音深度信任网,其平均误识别率可下降到3％.文中的研究对于提高心音识别算法在自然环境下处理数据的能力具有积极的意义.
展开▼
