Learning Pressure Sensor Drifts from Zero Deployability Budget

This work addresses the challenging problem of performing on-device online learning for a neural network to accomplish a regression task, under extreme memory and processing constraints, which a sensor typically presents. Specifically, we tackle the issue within the context of compensating the drifting behavior of the state of art LPS22DF absolute pressure sensor, operating under sporadic reference availability. The proposed solution is based on Gaussian Radial Basis Function networks and features a method for the allocation and removal of hidden neurons that dynamically updates the topology over the time. Additionally, we present an innovative adaptive distance threshold mechanism designed to ensure robust adaptivity of the model to sudden changes in input pattern distribution. The experimental assessment demonstrated significant error reductions ranging from 47.3% up to 93.4%, depending on reference availability frequency, when applied to sensors subject to different thermal stresses. The maximum memory footprint of 524 bytes (26 neurons) in all the performed experiments, proved the feasibility of performing the learning process directly within the sensor.