Semi-feature Sharing Deep Ensemble Model based on Sensor Data

Abstract

Human activity recognition(HAR) analyzes human activity information through sensors in wearable devices. Research on HAR based on deep learning has been widely used. The deep ensemble model enhances the generalization performance of the model based on the deep model. However,the deep ensemble model must train multiple models simultaneously, which requires a large amount of computing resources and time. In this paper, semi-feature shared blocks are used to construct correlated redundant features on the feature map so that the model training of each path in the deep ensemble model can generate input diversity. At the same time, we perform feature fusion on the base classifiers of the model so that the deep ensemble model can learn deeper data features while adding base classifiers. We use depthwise convolution instead of traditional convolution to reduce the model's computational complexity without changing the baseline model's structure. Compared with conventional ensemble learning methods, our proposed model has better results in terms of parameters and computation.