基于特征学习的ECG身份识别

1、 基于特征学习的ECG身份识别 ECG biometric authentication based on feature learning 摘 要基于特征学习的ECG身份识别随着社会信息数据的快速发展，人类的信息安全面临着巨大的安全隐患。生物特征识别技术作为具有高度安全性与唯一性的新型识别技术，逐渐进入大众视野。心电信号因其处理简单，易采集，难伪造等特点，使得ECG身份识别逐渐成为生物身份识别领域的一个研究热点。这一技术不但推动身份识别领域的快速发展，并且有效补充了现有的生物特征识别技术。目前为止，尽管在该方面取得了许多技术突破，但仍存在识别精度不高，时效性不好的问题。针对这一问题本文从心拍特征提取，和特征学习两方面进行研究。为了更贴近实际应用场景，采用数据来源中每个个体的心率、身体健康状况与情绪状态不受限制。在特征提取的过程中，根据信号的采样频率、心电信号与干扰噪声的频率特点，采用九层小波去噪，得到较纯净的信号。然后采用二阶差分阈值法进行心拍检测，最后分别提取了信号的形态学特征与小波特征。为了获取最优的分类心拍特征，通过在不同分类器下的实验对比可知，相对于采用单一形态学特征（维度4

2、25维，心拍分类准确率为74%，身份识别准确率90%）或小波特征（维度172维，心拍分类准确率为72%，身份识别准确率93%），采用形态学与小波的组合特征（维度624维，心拍分类准确率为76%，身份识别准确率93%）分类准确率更好。然而采用组合特征作为系统的输入特征，虽然提高了身份识别的准确率，但同时也造成特征维度急剧增加从而引入了过多特征冗余，导致身份识别模型的计算复杂度高、存储空间消耗大，识别效率低下。针对此问题的解决，本文采用核主成分分析法(KPCA)，弥补了线性变换PCA无法深层表示非线性信号内在联系的不足。通过实验可知KPCA算法（维度500维，心拍分类准确率为76%，身份识别准确率94%）能够降低特征维度，使得在不影响分类准确率的同时提高系统的时效性。但是KPCA算法并不适用于现实ECG身份识别的应用场景，为解决此问题采用特征学习网络来进一步提高系统的时效性。采用稀疏自编码网络来设定特征学习网络的初值，利用全局参数微调来提高此网络的识别性能，最后采用L-BFGS算法对网络参数寻优，从而降低ECG特征学习算法的时间复杂度与空间复杂度。最后通过实验对比，特征学习网络(维度50维

3、，心拍分类准确率为87%，身份识别准确率96%)与KPCA算法相比较，不仅能够有效地对特征降维，并且提高身份识别的分类准确率，从而保证识别模型的身份识别准确率，时效率与鲁棒性。关键词：身份识别，特征组合，层次型SVM，KPCA，稀疏自编码，特征学习AbstractECG biometric authentication based on feature learningWith the rapid development of social information data, information security of human beings are facing the huge security risk. As the new identification technology with high security and uniqueness, biometric identification technology is gradually entering the public. As a new biometric identification technology,

4、 ECG signal has simple preprocessing, easy collection and difficult falsification characteristics and gradually become a research hotspot in the field of biometric authentication. The technology not only promotes the rapid development of the field of biometric authentication, but also effectively complements the existing biometric identification technology. Although many technologies have made breakthroughs in the respect so far, there are still some problems of low identification precision and

5、bad efficiency. To solve these problems, the paper researches the feature extraction of heart beats and feature learning.In order to be closer to the actual application, the sources of data are not be restricted that include heart rate, physical condition and emotional state of every individual. In process of the feature extraction, according to the signal sampling frequency, frequency characteristics of ECG signal and noise, the paper adopts wavelet denoising of nine layer to obtain the pure si

6、gnal. Then we use two-order difference threshold method to detect heart beats and extract the morphological features of signal and wavelet feature. In order to obtain the optimal heart beats features for classification, the experimental contrast for different classifier has been made. Compared with the single morphological features (dimension is 425, heart beat classification accuracy is 74%, identification accuracy is 90%) and wavelet features (dimension is 172, heart beat classification accura

7、cy is 72%, identification accuracy is 93%), the compound feature (dimension is 624, heart beat classification accuracy is 76%, identification accuracy is 93%) could achieve higher classification accuracy.While the compound feature improves identification accuracy as input feature for system, the sharp increasing of feature dimension leading too much feature redundancy which causes high complexity and low efficiency of identification system. To solve this problem, the paper uses kernel principal

8、component analysis (KPCA) to make up the deficiency of linear transform PCA which couldnt express the intrinsic connection among nonlinear signal. We realize that KPCA algorithm (dimension is 500, heart beat classification accuracy is 76%, identification accuracy is 94%) could reduce feature dimension and improve system efficiency without affecting the classification accuracy. But KPCA algorithm is not suitable for the practical application of ECG identification, the paper adopts feature learnin

9、g network to further improve system efficiency. Firstly the paper uses sparse autoencoder to set initial of feature learning network and utilizes global parameter tuning to improve the recognition performance of the network. At last, we adopt L-BFGS algorithm to optimize network parameters and reduce time complexity and space complexity of ECG feature learning algorithm. Finally, compared with KPCA algorithm, the feature learning network (dimension is 50, heart beat classification accuracy is 87%, identification accuracy is 96%) not only can effectively reduce feature dimension and improve identification accuracy through experiments. So it ensures the accuracy, efficiency and robustness of authentication system.Keywords:Identity recognition; Compound feature; Hier

《基于特征学习的ECG身份识别》由会员wo7****35分享，可在线阅读，更多相关《基于特征学习的ECG身份识别》请在金锄头文库上搜索。