Research On Physical Activity Energy Expenditure Prediction Model Based On Wearable Devices

Physical activity energy expenditure(PAEE)monitoring is one of the hottest issues in the field of kinematics.Moderate exercises can effectively lower the risk of obesity,cardiovascular disease,osteoporosis,and among others.At present,the standard testing equipment for PAEE is not only cumbersome to wear,but also expensive for daily uses.Using wearable inertial measurement units to estimate PAEE has both theoretical and practical impacts.Aiming at the problems in feature engineering,we mainly study how to improve the effectiveness of deep learning methods in calculating PAEE.Given that simple signal fusion in processing time series data may incur loss of effective features,we design a multi-head signal fusion network using convolutional neural network(CNN).There are few public data sets available for PAEE estimation,the data sizes and the type of activities are generally small for machine learning tasks.Since training deep learning models requires a lot of training data to effectively learn parameters,we propose a transfer learning method to first train a human activity recognition classifier,and then apply it on PAEE estimation.Our main work includes:1.We first study how to combine accelerometer and gyroscope signals in the estimation of PAEE.We compare both machine learning methods and deep learning ones on their generalization ability for various speeds of walking and running.The results show that machine learning methods outperform deep learning methods on estimating energy consumption,and but deep learning methods are superior on their generalization abilities while dealing with different distributions of data.2.In contrast with traditional 1D convolutional neural networks in processing time series signals,we implement a multi-head convolutional neural network to from different channels extract hierarchical features from different sensors.The Comparative experiments show that our method can better fuse multiple categories of signals and mine the motion characteristics from different sensors.It can overcome the common issue of feature loss while using the traditional convolutional neural networks.In addition,the multi-head convolutional neural network shows improved generalization ability in PAEE application on activities with different types and intensities.3.In the Transformer-based PAEE model,we also design a transfer learning method to mitigate the shortage of training data.The experimental results show that transfer learning can greatly improve our Transformer model’s generalization ability even under the condition of using different sensors,wearing positions,and physical activities.