Abstract
Recently, various studies have been conducted on the quality of sleep in medical and health care fields. Sleep analysis in these areas is typically performed through polysomnography. However, since polysomnography involves attaching sensor devices to the body, accurate sleep measurements may be difficult due to the inconvenience and sensitivity of physical contact. In recent years, research has been focused on using sensors such as Ultra-wideband Radar, which can acquire bio-signals even in a non-contact environment, to solve these problems. In this paper, we have acquired respiratory signal data using Ultra-wideband Radar and proposed 1D CNN (1-Dimension Convolutional Neural Network) model that can classify and recognize five respiration patterns (Eupnea, Bradypnea, Tachypnea, Apnea, and Motion) from the signal data. Also, in the proposed model, we find the optimum parameter range through the recognition rate experiment on the combination of parameters (layer depth, size of kernel, and number of kernels). The average recognition rate of five breathing patterns experimented by applying the proposed method was 93.9%, which is about 3%~13% higher than that of conventional methods (LDA, SVM, and MLP).