In-Sensor Learning for Pressure Self-Calibration

An interesting research challenge for the TinyML community concerns the capability to achieve accurate online incremental learning for a regression task. This without relying on computational and memory demanding algorithms, such as back-propagation, while ensuring deployability on very assets-constrained devices. Pressure sensors are at the top of this challenge, because of the limited access to the ground truth reference, once deployed on tiny devices. This paper presents a viable, tiny solution to learn at any time how to provide compensations to the errors generated during the sensor lifetime, with minimal complexity while achieving satisfactory accuracy. The proposed solution is based on a Radial Basis Function network that is dynamically updated during online learning thanks to the sporadic availability of the reference values. The achieved results demonstrate robust performance, with an error reduction of up to 96% compared to the initial precision of the sensors and a remarkable accuracy even in scenarios with low availability over the time of the reference values. At maximum, only 6 nodes were required by the network corresponding to a memory footprint of 106 bytes.