心律失常是一类重要的心血管疾病，心电图信号的形态不确定性和多义性给心电信号自动化识别和心律失常分类带来巨大挑战。机器学习对于实现心电信号自动化识别，提高心律失常诊断正确率有重要意义。目前基于QRS波、忽略P波的诊断方法无法真正识别心律，对心律失常仅能粗粒度分类；卷积神经网络容易丢失更高层次空间信息，无法处理心电信号的多义性表达。针对这些问题，本课题提出了基于P-QRS模板提取和演化胶囊网络的特征提取和心电信号分类方法。首先借助小波变换和主成分析方法，实现QRS波和P波检测，完成P-QRS模板提取；建立胶囊网络，获取心律失常信号存在概率和姿态信息，使用一致性动态协议路由处理拥挤场景下的心电信号多义性表达。通过心电图数据的测试，验证P-QRS模板提取对分类的有效性和胶囊网络多特征提取的优势。

Arrhythmia is an important cardiovascular disease. The uncertainty and non-stationary of ECG signals bring great challenges to automatic recognition of ECG and arrhythmia classification. Machine learning is important to realize automatic recognition of ECG signals and improve the accuracy of arrhythmia diagnosis. At present, the diagnosis method based on QRS wave group and neglecting P wave cannot recognize rhythm. It can only distinguish small number of arrhythmia. In feature extraction and classification, the convolution neural network can extract the local morphological features of the ECG signal, but it is easy to lose the higher level space information. It is poor in dealing with Polysemous expression of electrocardiogram signals. This project proposes a method of features extraction and classification of ECG signals based on P-QRS template extraction and capsule network. First, we abstract signals of normal rhythm and arrhythmia. Wavelet transform, Hilbert transform and principal component analysis are used to extract features of QRS wave and P wave. Neural network and support vector machine are used to realize ECG QRS wave and P wave detection and classification problems; then, for normal ECG and arrhythmia signals extraction, we employ a multi-layer capsule network with a group of neurons (capsule structure). The capsules are used to get the objective signal presence probability and attitude information (features) of the normal rhythm and arrhythmia. Meanwhile, an agreement dynamic protocol routing between inter capsules for distinguishing an object in congestion scenario is proposed; finally, we use a heuristic optimization algorithm of hyper parameters of capsule network. This method combines genetic algorithm/evolutionary strategy and gradient descent to improve the global convergence of capsule network. It makes the capsule network jumping out of the local optimal trap. The super parameters of capsule network will be obtained. The capsule network model is trained on clinical electrocardiogram data, and the generalization performance is tested on the test set, which verifies the superiority of the P-QRS template extraction and capsule network in ECG signal feature extraction and arrhythmia classification.