Forecasting Task Optimization for A New Architecture of MEMS Reservoir Computing Using Stiffness Modulation

This paper reports a novel reservoir computing (RC) architecture based on MEMS resonant accelerometer. Our method makes use of stiffness modulation that the input is reflected in stiffness disturbance of nonlinear resonator. We eliminated the discretization of data before injected into the reservoir which is often necessary in conventional delay-based RC, so that the sensor can process natural signal directly which reduces system complexity and power consumption, providing a novel sensing paradigm. The new system integrates sensing and computing in a single device, and we specifically enhanced its forecasting performance because the original non-delay-based MEMS RC we proposed earlier is only suitable for recognition tasks. We carried out nonlinearity tuning for physical reservoir and optimization of the post-processing algorithm. Forecasting task related to chaotic system was successfully done by the measured response used in simulation. The proposed architecture shows a concept of intelligent sensor which can handle a variety of tasks and scenes.