Emulating Touch Signals from Multivariate Sensor Data Using Gated RNNs
Seung-Chan Kim, Byung-Kil Han · 2019
This paper proposes a sensor substitution system that generates time-series sensor data via recurrent neural networks (RNN)-based sequence analysis to regress a virtual sensor sequence. More specifically, the proposed system estimates capacitive touch sensor signals by exploiting tiny motion and audio signals generated by touch. The proposed system was validated in a supervised learning task in which multiple sensors-specifically, an omnidirectional microphone, an accelerometer, a gyroscope, and a capacitive touch sensor-were employed. The multivariate temporal information of the input sequence was modelled using a gated recurrent unit (GRU). The experimental results obtained verified the feasibility of the proposed system and indicated that, compared to inertial signals (e.g., acceleration and angular velocities), audio signals are better for estimating the corresponding touch sensor sequences.