Deploying Self Learning of Radial Basis Functions Tiny Neural Networks for In Sensor Calibration

This paper devises the deployment of an innovative real-time self-calibration algorithm within sensor-integrated computing resources, designed to compensate for time-varying errors due to thermal and mechanical stresses affecting MEMS accelerometers. The proposed approach employs a Radial Basis Function Neural Network (RBF-NN), which was implemented, trained, and tested within the Intelligent Sensor Processing Unit (ISPU) integrated into the LSM6DSO16IS Inertial Measurement Unit (IMU). This IMU also embeds a 3-axis digital gyroscope and a 3-axis digital accelerometer. The RBF-NN does not use the back-propagation algorithm and does not feature a fixed topology and parameter set. The ISPU integrates a low-power instruction set processor with 32 KiB of program RAM and 8 KiB of data RAM, operating at frequencies between 5 and 10 MHz. Both the 32-bit floating-point precision RBF-NN and its 16-bit integer precision counterpart required less than 21 KiB of the 40 KiB of memory available within the ISPU. The floating-point model, on real-time data acquired by the IMU, achieves an average error reduction of between 90.85% and 99.71%, with, on average, a learning time of 10.18 ms and an inference time of 8.82 ms. The 16-bit model achieves an average error reduction of between 92.90% and 98.92%, with, on average, a learning time of 7.64 ms and an inference time of 4.40 ms.