【作者】 王迪；

【导师】 司玉娟；

【作者基本信息】 吉林大学 ， 通信与信息系统， 2020， 博士

【摘要】 近年来,随着信息技术的快速发展和网络的普及,个人身份识别已经在金融安全、访问控制、安防监控和资料保密等领域得到了广泛应用,可靠、稳定、简单和低成本的个体身份识别技术在信息安全领域具有巨大的应用前景。然而传统身份识别工具,由于存在易丢失、易遗忘与易被盗用的缺点,已经无法满足人们的需求,因此生物特征识别技术应运而生。心电信号作为一种医学生理信号,不仅能够反映心脏的健康状态,同时蕴涵了个体丰富的身份信息,满足作为生物特征用于身份识别的必要特性。不仅如此,与人脸、指纹等体外密钥相比,心电信号还有其自身独特的优势:(1)心电信号属于体内密钥,难以被窃取;(2)心电信号自带活体检测属性,可以有效避免指模、照片等人造工艺带来的仿制问题。因此对心电身份识别技术进行研究,能够极大提升身份识别的安全性,具有广阔的应用前景。目前国内外心电身份识别研究广泛采用模式识别方法,通过计算机自动分析探索心电信号与个体身份之间的联系。在过去的数十年里,心电身份识别取得了长足的进步,大量研究成果相继涌现。然而,受心电信号自身复杂性与噪声的干扰,在应用中实现鲁棒、准确的心电信号身份识别仍然是一个很大的挑战。针对心电信号身份识别存在的问题,本文结合心电信号自身特点,分别从抗噪、抗心率变化与抗基准点检测误差三个方面展开鲁棒性心电身份识别算法的研究。研究工作归纳如下:1.心电信号在采集过程中会不可避免地混入噪声,这些噪声会扭曲信号形态,导致身份识别准确率下降。在传统心电身份识别过程中,去噪预处理虽然能有效抑制噪声,但是其同时存在算法普适性差、容易滤除身份识别有用信息的问题。为了消除传统去噪预处理给身份识别带来的不利影响,本文结合离散小波变换技术与稀疏自编码网络,通过将去噪思想融入特征提取,提出了一种基于时-频域自编码的抗噪声心电身份识别算法。算法首先利用离散小波变换将心拍分解为多个时-频成分,然后结合心电信号各波段与主要噪声的时-频分布特点进行特征选择,通过只保留关键波段对应的时-频成分滤除噪声影响,最后使用稀疏自编码网络通过特征学习在抑制残余噪声的同时,实现个体鉴别性特征提取并完成分类。在不考虑心率变化影响的条件下,通过在ECG-ID心率稳定数据集与MIT-BIH-AHA数据库上的对比实验,验证算法能够在无需去噪预处理的条件下,直接从原始含噪信号中提取有效特征,取得良好的身份识别精度。2.当人体发生情绪或运动状态改变时,心电信号因心率变化而产生T波偏移,这种偏移会使心电信号在形态上发生改变,导致识别准确率下降。为了满足实际场景应用,提出了一种面向心率变化的无去噪心电身份识别方法。算法首先采用本文提出的T-QRS-T重采样方法对心拍进行T波校正。与传统T波校正方法相比,T-QRS-T重采样方法的优势在于不依赖Q、S等微弱波幅点的检测,因此更能满足复杂实际场景的应用需求。在特征提取环节,针对传统神经网络,如自编码网络参数繁多且寻优过程复杂的问题,采用仅含少量超参数且具有抗噪性能的主成分分析网络直接从校正后的含噪心拍中挖掘个体间差异化信息,最后使用线性支持向量机完成分类。为了验证算法的有效性,在ECG-ID心率变化数据集与ECG-ID全数据集上开展多组对比实验,结果证明所提出的方法能够有效解决T波偏移问题,实现在心率变化下的鲁棒性心电身份识别。3.针对现有基准点检测算法定位精度不足给心电身份识别带来的不利影响,提出了一种基于窗口组合特征的抗基准点检测误差心电身份识别方法。该方法首先采用固定时长滑动窗将原始心电信号划分为多个窗口信号段,以这些窗口信号段代替心拍作为基本识别单元。然后,考虑到不同窗口信号段间的形态、信息含量等差异,对窗口信号段分别从统计域与时域角度进行全局与关键局部信息表示,其中统计域表示由自相关/离散余弦变换获取,时域表示由本文提出的窗口信号段MAX特征提取方法得到;第三,使用PCANet分别对统计域表示与时域表示进行特征提取,并组合两种特征实现二者信息互补,丰富身份识别判据;最后使用线性支持向量机完成分类。通过在ECG-ID心率稳定数据集、ECG-ID心率变化数据集、ECG-ID全数据集与MIT-BIH-AHA数据库上的多组实验对比,证明提出的算法能够在无基准点检测的条件下,有效改善心电信号身份识别准确率。更多还原

【Abstract】 In recent years,with the rapid development of information technology and widespread of the internet,personal identification has been widely used in applications such as financial security,access control,security monitoring and material confidentiality,et al.Nowadays,reliable,stable,simple and low-cost personal identification technologies have broad prospects in the field of information security.Since the traditional identification tools are prone to be lost,forgotten and stolen,they can no longer meet people’s need for personalized privacy protection.To solve the above problem,biometric identification is proposed.As a physiological signal,the electrocardiogram(ECG)can not only reflect the health state of heart,but also contains rich personal identity information.ECG satisfies the necessary requirements for being a biometric,and thus can be used for biometric identification.Moreover,compared with such external biometrics as face and fingerprint,ECG also has its own unique advantages:(1)As an electrical signal inside the body,ECG is much more difficult to be stolen;(2)ECG comes with the attribute of liveness detection,which makes it able to effectively avoid problems caused by synthetic samples,e.g.,photos or gummy fingers that have fingerprint impressions.Due to the above advantages,ECG biometric identification can greatly improve the security of identification,and thus has broad application prospects in the future.Currently,pattern recognition methods are widely adopted by researchers to explore the relationship between ECG signals and personal identity through automatic computer analysis.In the past decades,great progress has been made in ECG biometric identification,and a lot of researches have been implemented.However,due to the complexity of ECG signal and the interference of noise,it is still a big challenge to realize robust and accurate ECG biometric identification in practical application.In order to solve the existing problems in ECG biometric identification,this thesis makes robust ECG biometric identification researches from three respects,namely,robustness to noise,robustness to heart rate change and robustness to fiducial-point detection error,based on the characteristics of ECG signal.The main contents and contributions of this study are presented as follows:1.In the data acquisition process,ECG signal will inevitably be mixed with a variety of noise.The noise will distort the ECG waveform and lead to the decline of identification accuracy.Although denoising process can effectively remove the noise,it also exists the problems of poor universality and easy to filter out identity-related information.To eliminate the adverse effects of denoising process on ECG identification,this thesis combines Discrete Wavelet Transformation and sparse autoencoder to propose a novel temporal-frequency autoencoding based method for noise robust ECG identification,whose main idea is to incorporate denoising process into feature extraction step.In the method,heartbeats are firstly decomposed into multiple time-frequency components by Discrete Wavelet Transformation.Then,feature selection is performed on the obtained components according to the time-frequency distribution of ECG waveforms and noise.By only preserving time-frequency components corresponding to the key waveforms,noise can be efficiently removed.At last,the stacked sparse autoencoder is introduced to implement discriminative feature extraction and classification while depressing the residual noise in a feature learning way.Without considering the impact of heart rate change,the effectiveness of the method is verified by performing comparison experiments on the ECG-ID heart rate stabilization dataset and the MIT-BIH-AHA database.Experimental results show that the proposed method can directly extract effective features from the original noisy signals and achieve high identification accuracy without requiring denoising process.2.When the state of motion or emotion changes,T wave shift will occur on ECG signals due to the heart rate change.The T wave shift can change the morphology of ECG signals,leading to the decline of identification accuracy.To solve the above problem and satisfy the real scenario application,a novel ECG identification method,which does not require denoising process and is robust to heart rate change,is proposed.In the method,firstly,all the heartbeats are normalized by a novel T-QRS-T resampling strategy to correct T wave shift.Compared with traditional T wave shift correction methods,the significant advantage of the proposed T-QRS-T resampling strategy is that it does not rely on the detection of fiducial points with weak amplitude value,e.g.,Q and S.Thus it can better satisfy the requirement of complex real scenario application.Then in the feature extraction process,to solve theproblem that traditional neural networks,e.g.,autoencoder,have numerous parameters and complex optimization process,the principal component analysis network(PCANet)with much less parameters and good noise robustness is adopted to learn discriminative features from normalized noisy heartbeats.Finally,linear support vector machine is adopted to classify the extracted features.To verify the effectiveness of the proposed method,comparison experiments are performed on both ECG-ID heart rate change dataset and ECG-ID dataset with all data.Experimental results show that the proposed method can effectively solve the problem of T wave shift,and realize noise robust ECG biometric identification under heart rate change condition.3.To eliminate the adverse effects of insufficient accuracy of existing fiducial-point detection method on ECG biometric identification,a novel ECG identification method based on combined segment features,which is robust to fiducial-point detection error,is proposed.In the method,firstly,a fixed-length sliding window is used to partition the original ECG signal into multiple short segments,which will replace heartbeat as basic identification unit in the following steps.Secondly,considering the difference of morphology and content among the obtained segments,global and key local representation of the segments are respectively obtained from the perspective of statistical domain and time domain.Here,the statistical domain representation is obtained by Autocorrelation/Discrete Cosine Transformation,and the time domain representation is obtained by the proposed MAX feature extraction method.Thirdly,PCANet is adopted to respectively extract features from the statistical domain representation and the time domain representation,and the extracted features are combined together to realize information complementation.Finally,linear support vector machine is adopted to classify the combined features.To verify the effectiveness of the proposed method,comparison experiments are conducted on ECG-ID heart rate change dataset,ECG-ID heart rate stabilization dataset,the MIT-BIH-AHA database and ECG-ID dataset with all data.Experimental results demonstrate that the proposed method can effectively improve the identification accuracy without requiring fiducial point detection in the preprocessing step.更多还原