ECG Signals Classification Method for Wireless Body Area Network Based on Quantized Residual Network

ECG Signals Classification Method for Wireless Body Area Network Based on Quantized Residual Network

연구 분야: Infrastructure

학회: ICBDT '22: Proceedings of the 5th International Conference on Big Data Technologies

In recent years, rapidly developing deep neural network methods have been widely used for ECG signals classification. However, deep neural networks generally have problems such as too many network redundant parameters, high computational complexity, and large model storage, which are not suitable for application in resource-constrained wireless body area network (WBAN). To address the above problems, this paper proposes a quantized residual network-based ECG signals classification method for WBAN. The method firstly proposes a full-precision residual network, which is activated by adding the input and output of the residual unit through the form of a hopping layer connection and using a one-dimensional convolutional layer as the residual information to enhance the characteristics of the residual block. The test of the MIT-BIH arrhythmia database shows that the accuracy of ECG classification is 99.18%. Secondly, the parameters of the above full-precision residual network are fixed-point binary quantized, and the same database test shows that the final model memory occupation is compressed by 32 times, and the classification accuracy is 99.07%. The proposed method effectively solves the problems of many redundant parameters and large model storage in deep neural networks while ensuring high classification accuracy, and provides some reference for the research of low power consumption in WBAN.