Nonlinear dynamic feature fusion in Wearable Sensor-Based Human Activity Inference

The human movement process is a nonlinear and complex system with many definite instabilities, and the activity behaviors originate from the mutual non-linear interactions of various motor subsystems of the human body. Sensor-based Human Activity Recognition (HAR) in mobile application scenarios often face changes in sensor patterns, the use of shallow features in traditional machine learning leads to poor performance in incremental or unsupervised learning, and machine learning is limited to the constraints of linear relationships in the data, in this study of HAR, it is proposed to differentiate from a single traditional machine learning approach by using Lyapunov Exponent, Poincare plot, Recurrence Plot and other nonlinear dynamics features to achieve accuracy of 85.90%,95.91%,93.71% on mHealth dataset, proves that Nonlinear dynamics features are available in human activity recognition, and at the same time, fusing multiple nonlinear dynamics features to obtain the different deep multi-dimensional representations of the same action during human activity, better describing the process of motor subsystems during human activity, and solving the complex nonlinear system of human activity recognition in a deeper level, with an accuracy of 97.88% in the mHealth dataset, which indicates that the model designed in this work for the fusion of multiple nonlinear dynamics features is effective in human activity recognition.