Department of Computer Science, College of Humanities, Arts, and Sciences, University of Northern Iowa, Cedar Falls, IA 50613, USA.
Department of Computer Science and Engineering, BML Munjal University, Gurugram 122413, India.
Sensors (Basel). 2023 Jul 24;23(14):6631. doi: 10.3390/s23146631.
A good night's sleep is of the utmost importance for the seamless execution of our cognitive capabilities. Unfortunately, the research shows that one-third of the US adult population is severely sleep deprived. With college students as our focused group, we devised a contactless, unobtrusive mechanism to detect sleep patterns, which, contrary to existing sensor-based solutions, does not require the subject to put on any sensors on the body or buy expensive sleep sensing equipment. We named this mechanism Packets-to-Predictions(P2P) because we leverage the WiFi MAC layer traffic collected in the home and university environments to predict "sleep" and "awake" periods. We first manually established that extracting such patterns is feasible, and then, we trained various machine learning models to identify these patterns automatically. We trained six machine learning models-K nearest neighbors, logistic regression, random forest classifier, support vector classifier, gradient boosting classifier, and multilayer perceptron. K nearest neighbors gave the best performance with 87% train accuracy and 83% test accuracy.
良好的睡眠对于我们认知能力的无缝执行至关重要。不幸的是，研究表明，美国成年人口中有三分之一严重睡眠不足。以大学生为重点群体，我们设计了一种非接触式、不引人注目的机制来检测睡眠模式，与现有的基于传感器的解决方案不同，这种机制不需要受试者在身体上佩戴任何传感器或购买昂贵的睡眠感应设备。我们将这种机制命名为“数据包到预测（P2P）”，因为我们利用家庭和大学环境中收集的 WiFi MAC 层流量来预测“睡眠”和“清醒”期。我们首先手动确定提取这种模式是可行的，然后我们训练各种机器学习模型来自动识别这些模式。我们训练了六个机器学习模型——K 最近邻、逻辑回归、随机森林分类器、支持向量分类器、梯度提升分类器和多层感知器。K 最近邻的表现最好，训练准确率为 87%，测试准确率为 83%。
