Research on Prediction Model of Physical Activity Energy Expenditure with Wearable Sensors

Abstract

To solve the contradiction between multiple wearable sensor features and the limited computing power and storage capacity of embedded devices, feature engineering is used to select the best features for predicting physical activity energy expenditure (PAEE) on the basis of data fusion of multiple sensors (accelerometer and gyroscope sensors). In the data preprocessing stage, time-domain and frequency-domain features of the sensor are extracted by using sliding window technology, and sinusoidal curve fitting is used for dataset at three velocity levels, finally hypothesis testing is carried out to check data outliers. A WEKA experimental platform is constructed based on filtering, warpper and embedded feature selection algorithms and machine learning prediction models such as multiple linear regression, regression tree, support vector machine and neural network. Finally, the optimal model is selected by evaluating the correlation coefficient and mean absolute error of each model during the decision level fusion. The dataset with jitter is used as the test data, which shows that feature selection can mitigate model overfitting and improve the model’s generalization ability and robustness. Embedded feature selection adopts classical elastic network algorithm. Experimental results show that the features extracted from accelerometer sensors play a more decisive role than those from gyroscope sensors in PAEE and the neural network model of multi-sensor feature fusion based on correlation coefficient method is the optimal model.

Keywords

• |physical activity energy expenditure|feature selection|neural network