Learning Online MEMS Calibration with Time-Varying and Memory-Efficient Gaussian Neural Topologies

• Published in: Sensors (Basel, Switzerland) Vol. 25; no. 12; p. 3679
• Main Authors: Pau, Danilo Pietro, Tognocchi, Simone, Marcon, Marco
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 12.06.2025; MDPI
• Subjects: Accelerometers; Accuracy; Algorithms; Analysis; Artificial intelligence; Calibration; Datasets; Deep learning; Digital signal processors; Global positioning systems; GPS; in-sensor tiny machine learning; inertial sensor calibration; intelligent sensor processing unit; Machine learning; Microelectromechanical systems; Navigation systems; Neural networks; on-device learning; radial basis functions; Real time; Sensors; in-sensor tiny machine learning; intelligent sensor processing unit; radial basis functions; inertial sensor calibration; on-device learning
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s25123679

This work devised an on-device learning approach to self-calibrate Micro-Electro-Mechanical Systems-based Inertial Measurement Units (MEMS-IMUs), integrating a digital signal processor (DSP), an accelerometer, and a gyroscope in the same package. The accelerometer and gyroscope stream their data in real time to the DSP, which runs artificial intelligence (AI) workloads. The real-time sensor data are subject to errors, such as time-varying bias and thermal stress. To compensate for these drifts, the traditional calibration method based on a linear model is applicable, and unfortunately, it does not work with nonlinear errors. The algorithm devised by this study to reduce such errors adopts Radial Basis Function Neural Networks (RBF-NNs). This method does not rely on the classical adoption of the backpropagation algorithm. Due to its low complexity, it is deployable using kibyte memory and in software runs on the DSP, thus performing interleaved in-sensor learning and inference by itself. This avoids using any off-package computing processor. The learning process is performed periodically to achieve consistent sensor recalibration over time. The devised solution was implemented in both 32-bit floating-point data representation and 16-bit quantized integer version. Both of these were deployed into the Intelligent Sensor Processing Unit (ISPU), integrated into the LSM6DSO16IS Inertial Measurement Unit (IMU), which is a programmable 5–10 MHz DSP on which the programmer can compile and execute AI models. It integrates 32 KiB of program RAM and 8 KiB of data RAM. No permanent memory is integrated into the package. The two (fp32 and int16) RBF-NN models occupied less than 21 KiB out of the 40 available, working in real-time and independently in the sensor package. The models, respectively, compensated between 46% and 95% of the accelerometer measurement error and between 32% and 88% of the gyroscope measurement error. Finally, it has also been used for attitude estimation of a micro aerial vehicle (MAV), achieving an error of only 2.84°.