Page 120 - Read Online

Page 10 of 13 Zhang et al. Soft Sci. 2025, 5, 17 https://dx.doi.org/10.20517/ss.2024.68

pillow with a high pressure, and thus has a high height (~1.0 relative current change). Its length depends on
when the person sits up. The waveform of snoring comes from the inhale action with a small pressure, and
thus has a small height (~0.04-0.07 relative current change). Its length depends on the snoring time, which
can range from 2 to 8 s. Therefore, abnormal snoring activity can be timely detected by the intelligent
pillow, and a warning could be made to wake the sleeper for safety.

Application of sensor array on airbag pillow for head posture recognition
Monitoring the sleeping posture changes is of great significance for evaluating sleep quality, and the head
posture is one of the important indicators to judge the health status during sleep. Although there are many
sleep monitoring devices on the market to collect body movement information, most of them are wearable
electronics that need to be tightly worn on the body and thus inevitably hinder the quiet sleeping
experience. To overcome this practical limitation, herein we develop an intelligent pillow equipped with the
developed five flexible and breathable sensors, which can detect and recognize the head posture with the
help of the one-dimensional (1D) convolutional neural networks (CNNs) algorithm prediction model
[Figure 5A]. The five sensors are aligned along the length direction on the surface of the pillow and labeled
as A to E to distinguish different signal sources. Three head positions are set, including lying on the back,
lying on the right side, and lying on the left side of the body. When the head comes into contact with the
sensors at different positions during its flipping process, each sensor will individually output a
corresponding real-time response signal upon its contacting state with the head. A multi-channel signal
acquisition system is explored to simultaneously acquire the five sensors’ output signals in a real-time mode,
and the collected five-channel data information is fed into the built-in machine learning model for training
(see detailed information on architecture, hyperparameters and sample sets in Supporting Information).
Through one convolutional layer to extract the local features of the waveforms and one pooling layer to
reduce the dimensionality of the data, a CNN prediction model is built after training and calibration. Then
future testing data can be input into the trained model to obtain the prediction results.

In practice, volunteers are asked to lie down with heads on the pillow and perform the three different head
positions so that characteristic waveforms obtained by the five sensors are collected [Figure 5B]. Totally 350
sets of data are obtained and randomly classified into three groups according to a ratio of 5:1:1. Among
them, 250 sample data are used for the training set, 50 sample data are used in the testing set, and 50 sample
data are used in the validation set. The gesture data sets show obvious differences after training [Figure 5C].
The 300-sample data including the training and test sets are input into the CNN framework for model
training, yielding a high recognition rate for the three different head positions (Figure 5D; the confusion
matrix shows the training results with an average accuracy of up to 98%).

CONCLUSIONS
In summary, a flexible and breathable pressure sensor is developed on an all-paper platform. The sensor
adopts a planar device structure with one MXene/dust-free paper as the piezoresistive layer on top of one
dust-free paper with printed interdigital silver electrodes. SEM characterization confirms the achievement of
the MXene-coating fabric network and the multilayered device structure. Measurements show that the
sensor not only possesses good flexibility, air/moisture permeability, and wearing biocompatibility but also
-1
exhibits excellent sensing performance with a high sensitivity of 16.7 kPa , a wide detection range of
100 kPa, fast response time of 50 ms and good stability over 10,000 cycles. Towards practical sensing
applications for sleep monitoring, the single sensor unit is integrated into the airbag pillow for detecting
snoring with different modes, and the five-sensor array is aligned for recognizing sleeping head posture with
machine learning. The proposed paper-based sensor, along with the intelligent airbag pillow, provides a
promising approach for sleep monitoring in a comfortable unrestricted way.