Domain-Robust Pre-Training Method for the Sensor-Based Human Activity Recognition

Abstract

Transfer learning improves problem-solving efficiency by transferring the learned knowledge from the source domain to the target domain. In transfer learning, using a large amount of data for pre-training is beneficial to improve the robustness of the model. Data differ significantly when the domain changes in Sensor-Based human activity recognition (HAR). Currently, in HAR, data usage is relatively independent, lacking source domains with massive data and rich labels. This paper proposes a new pre-training method using multiple domain datasets to construct a domain-robust pre-training model. We divide the pre-training dataset into basic and complex activities scenarios by considering the difference in activity classification. We evaluate the classification scenarios that are most beneficial for sensor-based HAR based on the constituted dataset and using deep convolutional networks. We show that our method verified the influence of the source domain on transfer learning in sensor-based HAR. By constructing a sizeable correlated source domain, our method can enhance the generalization ability of the network model. This paper also demonstrated that large-scale and basic activity classification datasets can be better used as pre-training models to participate in HAR classification tasks.