Search

CN-121987144-A - Method for automatically identifying sleep based on one-dimensional convolutional neural network

CN121987144ACN 121987144 ACN121987144 ACN 121987144ACN-121987144-A

Abstract

The application relates to the field of intelligent wearing, and provides a method for automatically identifying sleep based on 2 one-dimensional convolutional neural networks and signal characteristics, which comprises the steps of acquiring a PPG signal and a triaxial acceleration signal, wherein the PPG signal and the triaxial acceleration signal are measured by wearing equipment on a measuring part of a user; and acquiring the operation results of the training 2 one-dimensional Convolutional Neural Networks (CNN) on the characteristics of the measurement signals. According to the scheme, the accuracy and stability of intelligent wearing automatic sleep detection are effectively improved, 2 one-dimensional CNNs are used for guaranteeing the accuracy, and meanwhile, the operation amount of the intelligent wearing automatic sleep detection is remarkably reduced compared with a single multi-dimensional CNN, so that the intelligent wearing automatic sleep detection device is beneficial to being applied to a single chip microcomputer.

Inventors

  • LI JIUCHAO
  • DING HUI

Assignees

  • 深圳市维亿魄科技有限公司

Dates

Publication Date
20260508
Application Date
20241101

Claims (11)

  1. 1. A method for analyzing characteristics of a measurement signal to automatically identify sleep using two one-dimensional convolutional neural networks, the method comprising: 1) Acquiring a PPG signal and a triaxial acceleration signal which are measured by a wearable device on a measurement part of a user; 2) Extracting features of the measurement signal, the features comprising 1 piece of feature data buffered in units of every 5 minutes, the data comprising an average pulse rate per minute, a wear detection result, a 5 minute total step, a 5 minute total motion amount, a 20 second motion amount extracted based on continuous 20 second data, a 20 second signal quality extracted based on continuous 20 second PPG signals, and a pose; 3) Inputting the characteristic data into two one-dimensional convolutional neural networks which are subjected to training, wherein the networks are improved one-dimensional LeNet-5 networks, each network consists of 7 layers, including a C1, a C3 and a C5 convolutional layer, an S2 and an S4 pooling layer, an F6 full-connection layer and a Gaussian connection layer which is classified by using a softmax function, and acquiring an operation result of the networks to identify the sleeping state of a user.
  2. 2. The method according to claim 1, wherein the motion amount is a difference of a sum of successive triaxial acceleration vectors, an amount of change of which absolute value is accumulated, for reflecting an activity level of the user.
  3. 3. Method according to claim 1, wherein the signal quality comprises a peak amplitude mean, a peak amplitude variance, a peak interval mean and a variance of peak points of the pulse wave in the PPG signal for assessing stability and reliability of the PPG signal.
  4. 4. The method of claim 1, wherein the 20 second motion and 20 second signal quality are data sources with continuous 20 second triaxial acceleration and PPG signals, respectively, and the corresponding data extraction features are captured per minute for different time periods (0 to 20 seconds, 10 to 30 seconds, 20 to 40 seconds, 30 to 50 seconds, 40 to 60 seconds), and 25 20 second motion and 25 20 second signal quality are acquired in total within 5 minutes to provide finer time resolution features.
  5. 5. The method according to claim 1, wherein the gestures are classified into 12 gestures according to the size and positive and negative of x, y and z three-axis data based on the features extracted by the 3-axis acceleration, the gestures are counted once per second according to the average value of the three-axis data, and the mode of each second gesture is counted as a gesture of 5 minutes in 5 minutes to reflect the overall activity gesture of the user.
  6. 6. The method of claim 1, wherein the measured PPG signal comprises green light and near infrared light, the near infrared light being used for wear detection, the green light being used for calculating pulse rate and extracted signal quality.
  7. 7. The method of claim 6, wherein the wearing detection is proximity detection, i.e. detecting whether an object is approaching a signal acquisition module of the device by comparing the value of the reflected PPG signal with a preset threshold.
  8. 8. The method according to any one of claims 1 to 7, wherein the training data is in the form of an array of 1 x 32, and the two one-dimensional convolutional neural networks operate on the feature data independently, and the output results thereof are comprehensively analyzed to determine the sleep state of the user.
  9. 9. The method of claim 1, further comprising a training process for the two one-dimensional convolutional neural networks, including data preprocessing, network construction, parameter setting, model training and verification, to ensure accuracy and generalization capability of the network.
  10. 10. The method according to claims 1 to 9, further comprising subsequent processing of the recognition result, such as generating a sleep report from the recognition result, providing the user with knowledge of his sleep quality and sleep habits.
  11. 11. Method according to any of claims 1 to 10, wherein the wearable device may be a smart watch, a smart bracelet, a health monitoring patch or the like, capable of continuously monitoring the PPG signal and the tri-axial acceleration signal of the user.

Description

Method for automatically identifying sleep based on one-dimensional convolutional neural network Technical Field The application relates to the technical field of intelligent wearing, in particular to a method for realizing automatic sleep recognition based on a one-dimensional convolutional neural network and signal characteristics. According to the method, the PPG signal and the triaxial acceleration signal are comprehensively analyzed, the characteristics are extracted and input into two one-dimensional convolutional neural networks for operation, and finally, the accurate judgment of the sleep state is realized. Background The intelligent wearable devices on the market at present mainly adopt an acceleration sensor and a PPG sensor to monitor the sleep state of a user. However, due to the large difference of individual sleep characteristics, the traditional modeling method has limited effect of detecting sleep, and has high probability of missed detection and false detection. Therefore, there is a need for a more accurate and reliable sleep detection method to meet the needs of users for sleep monitoring. Disclosure of Invention The application provides a method for automatically identifying sleep based on two one-dimensional convolutional neural networks and signal characteristics. According to the method, the characteristics of the PPG signal and the triaxial acceleration signal are comprehensively analyzed, and the two one-dimensional convolutional neural networks are utilized for deep learning and recognition, so that the accurate judgment of the sleep state is realized. Compared with the traditional modeling method, the method has higher accuracy and stability, and is suitable for sleeping characteristic differences of different individuals. Drawings Fig. 1 is a flow chart of calculation of sleep results. S1, obtaining original data, S2, processing the original data, extracting signal characteristics, S3, taking the extracted characteristics as the input parameters of the neural network, and obtaining the operation result of the neural network. Fig. 2 is a calculation flow of the neural network. Each network consists of 7 layers, including a C1, C3, C5 convolution layer, an S2, S4 pooling layer, an F6 full connection layer, and a Gaussian connection layer classified by using a softmax function, and the operation result of the network is obtained to identify the sleeping state of the user. Detailed Description The application is described in detail below with reference to the drawings and the specific embodiments, but does not limit the scope of the application. 1. And (3) data acquisition: o Acquiring physiological signals of a user, including photoplethysmography (PPG), using a wearable device Signals and triaxial acceleration signals. o The PPG signal is collected by green LED lights and photodetectors on the device for monitoring heart rate and heart rate variability. o The triaxial acceleration signals are collected through an accelerometer arranged in the device and are used for monitoring the movement and posture change of a user. 2. Data preprocessing: o And filtering the collected PPG signal and the triaxial acceleration signal to remove high-frequency noise and direct-current components, wherein PPG data keep heart rate related signals of 0.5Hz to 5Hz, and acceleration data keep motion related signals of 0.1Hz to 3 Hz. o The continuous signal is segmented according to a fixed time window for subsequent processing. o And checking the missing value and the abnormal value of each section of signal, and filling by using an interpolation method or an average value method to ensure the data quality. 3. Feature extraction: o The average pulse rate per minute is calculated according to the PPG signal to reflect heart rate variation of a user, and the method specifically comprises the steps of detecting a heart peak value in the filtered PPG signal, determining an effective heart peak value by searching a local maximum value of the signal, setting a threshold value, calculating an average time interval of the effective peak value, namely, dividing 60 by the average time interval to obtain an instantaneous heart rate, and counting the average value of the instantaneous heart rate within 1 minute, namely, the average pulse rate per minute. o Wearing detection result, using near infrared light to make proximity detection and judging that said equipment is correctly worn Worn at the measurement site of the user. This can ensure validity and accuracy of the data. The specific operation is as follows: the average value of the infrared light signals of each second is taken, the threshold value is set to be larger than the average value, the wearing is failed when the average value is smaller than the threshold value, the wearing is failed when the average value is larger than the threshold value, the infrared light signals are continuously worn for 5 times within n minutes, and the wearing is failed when