Calibration of Haptic Sensors using Transfer Learning

Abstract

This paper introduces a novel sensor calibration scheme, the Hybrid Analytical and Network (HAN) model, which is especially suitable for regression based on neural networks with small training datasets. Through the HAN model, qualitative physical knowledge can be used to facilitate network training so that the demand on training data can be significantly reduced. Furthermore, training results can be transferred to the calibration of other types of sensors. The HAN model has been used to calibrate two haptic sensors, namely, the balance sensor and the skin sensor, which both operate on the principle of Frustrated Total Internal Reflection (FTIR), but for two different applications. The balance sensor is designed for human balance ability assessment, while the skin sensor is designed for evaluation of human skin properties. The calibration results using the HAN model are far more accurate and faster than other neural network training methods, especially for a small experiment dataset. The strong transfer ability of the HAN model is also demonstrated by the transfer attempt from balance sensor calibration to that of the skin sensor. The HAN model provides a new method to integrate model-based and data-driven sensor calibration methods, and shows greater efficacy. Its application is not restricted to sensor calibration but has broader potential value in all neural network regression tasks.