Gaussian Process Regression for Improving the Performance of Self-powered Time-of-Occurrence Sensors

In our previous work, we had demonstrated a CMOS timer-injector integrated circuit for self-powered sensing of time-of-occurrence of mechanical events. While the sensor could achieve an improved time-stamping accuracy by averaging the output across over multiple channels, the mismatch between the channels made the calibration process cumbersome and time-consuming. In this paper, we propose the use of non-parametric machine learning techniques to achieve more robust and accurate event reconstruction. This is demonstrated using training and testing data that were obtained from fabricated prototypes on a $0.5-\mu \mathrm {m}$ CMOS process; the model trained using Gaussian process regression can achieve an average recovery accuracy of 3.3% on testing data, which is comparable to the performance of using an averaging technique on calibrated injection results. The experimental results also validate that scalable performance can be achieved by employing more injection channels.