Application of neural ordinary differential equation in online learning of dynamic responses

To study the application of neural ordinary differential equation on smart sensors, numerical simulations of a second order system sensor under online learning of dynamic responses condition are performed. The time series prediction abilities under different environment noises of neural ordinary differential equation and fully connected network are tested. It is found that neural ordinary differential equation has higher training efficiency and no extra memory cost for larger prediction length. These features can improve the existing smart sensors and enhance the ability of edge computing.