| In recent years,heart disease has seriously threatened human health.As the most commonly used heart detection method,electrocardiogram(ECG)is very important for the detection of heart disease.The signal recorded by ECG is a non-linear time series data,which shows voltage changes during cardiac pacing.It is not only weak and susceptible to noise interference,but also has serious data imbalance and low linear separability between different categories.These problems make it difficult to obtain effective model coefficients,which seriously restricts the further improvement of heart rhythm recognition.The thesis introduces two methods to solve the problems of severe imbalance in ECG data and difficult to extract discriminative characteristic information.One is the method of heart rhythm recognition based on generative adversarial network and convolutional network,the other is the method of integrated learning based on differential information.(1)The method of heart rhythm recognition based on generative adversarial network and convolutional network.Studies have shown that severe imbalances in ECG data categories have led to recognition models biased to a larger number of categories.The models are often in a state where some categories are overfitted and other categories are underfitted,making effective extraction of features extremely difficult.This method uses the CNN model in collaboration with the DCGAN model to reduce the impact of uneven categories.First of all,train the CNN model under the category imbalance and use this model to filter out the underrepresented ECG data;Then input the under-represented data into the DCGAN model for iterative training,use the generator in the DCGAN model to generate imbalanced and error-prone data,and use the discriminator in the DCGAN model to filter the generated data.The above data expansion method can reduce the impact of data imbalance.Finally,the experimental results show thatthe model can effectively improve the accuracy of heart rhythm recognition.(2)The method of integrated learning based on differential information.Studies have shown that the common information of ECG signals is higher among different types,but the characteristic information of ECG signals is less among each types.These characteristic information is exactly the key of heart rhythm recognition.Because the ECG signal is weak and susceptible to noise,it is difficult to mine the deep feature information of the ECG signal under limited data conditions.This paper introduces ECG differential information to add higher-order information of ECG signals,which can enhance the model’s ability to characterize ECG signals.First of all,use the original ECG data to generate first-order and second-order difference data;Then,the difference information and the original ECG are separately extracted and mapped to different high-dimensional spaces to characterize the ECG from multiple angles;Finally,based on the idea of ensemble learning,the feature mapping of different spaces is integrated to generate deep features and improve the recognition performance of the rhythm recognition method.Experimental results show that using differential information to mine deep features can effectively improve the accuracy of recognition. |
