Research On ECG Signal Classification Algorithms Based On Deep Learning

Electrocardiogram(ECG)is a non-invasive detection tool,which is an important basis for detecting arrhythmia symptoms,reflects the physiological state of various parts of the heart,and is an important auxiliary means for detecting cardiovascular diseases.Due to the lack of ECG signal sample data of arrhythmia,the number of signals of various types is seriously unbalanced.In addition,the features of ECG signals are complex,which makes the detection of ECG signals of arrhythmia by current classification algorithms prone to misdiagnosis and missed diagnosis.Based on the rapid development of deep learning technology in recent years,this thesis studies the classification of ECG signals from the perspectives of data enhancement algorithm and classification algorithm.The main research contents are as follows:1.An ECG signal classification method based on the combination of CNN and Bi LSTM is presented.Aiming at the problems that the current feature extraction methods lack the ability to capture details and the morphological features of ECG signals extracted at a single scale are not comprehensive enough,a point-to-point cross layer fusion of features is proposed to enhance the residual network to increase the amount of information contained in a single feature,and then a parallel convolutional neural networks(CNN)structure is constructed by combining the scale decomposition strategy to expand the dimension space,enhance the ability of multi-scale analysis of complex features;In view of the lack of relevance of the multi-level features extracted by CNN,the bi-directional long short-term memory network(Bi LSTM)is introduced to improve the correlation decision-making ability of the model between the multi-level ECG signal features extracted by CNN by increasing the reverse information association ability of the long short-term memory network and returning the hidden state of different time steps.2.A Bi LSTM-DCGAN ECG signal generation and classification method based on dynamic time warping algorithm is presented.When Deep Convolution Generative Adversarial Networks(DCGAN)has multiple categories in its own training data,it cannot directly classify ECG signals while generating them,so each training can usually only target one ECG signal data,or need to train a new classification algorithm for classification.In view of the above problems,the class labels of different classes of ECG signals are introduced as the conditional discriminant information generated by the generator network data,and a dual-discriminator structure is proposed by combining the Dynamic Time Warping(DTW)algorithm with DCGAN.The newly added similarity discriminator is used to guide the training process of the model,and the classification of the generated ECG signals is realized by calculating the DTW scores between the signals generated during the training process and the real signals of each category;For the problem of feature extraction from local sequences,Bi LSTM is combined with CNN to improve the sequence learning ability of the generator network.3.A data enhancement method for generative adversarial networks based on threshold transfer is presented.Aiming at the problem that it is difficult for generative adversarial networks to generate valid data through training models when the amount of data in the training set is too small,a threshold-based deep generative transfer learning method is proposed by introducing the DTW score and adding a threshold to guide the parameter transfer process of the generative adversarial network.Aiming at the problem that the generated signal is single and cannot perform effective data enhancement,mini-batch discrimination is introduced by calculating the distance between each sample and other samples during training.And working together with the DTW similarity discriminator to improve the diversity of generated signals.Finally,by enhancing the S-category signal,the classification performance of the classification algorithm for N,S,V,and F ECG signals is improved.