Enhancing automotive products with TinyML and MEMS sensors: a preliminary approach

Abstract

The Tiny Machine Learning (TinyML) research field has seen exponential growth in recent years, mostly due to advances in the usage of IoT devices and microcontrollers. Despite major challenges related to hardware constraints, TinyML has been expanding the scope of applications in several domains, with the automotive industry taking special interest in these models for their potential to enhance existing Advanced Driver Assistance Systems (ADAS). ADAS have a vast array of sensors that produce huge quantities of quality data, that can in turn be used to solve different perception tasks. In this work, we present an architecture based on TinyML and Microelectromechanical systems (MEMS) sensors to accurately classify different handling and misuse cases on a test car. After showcasing the experimental setup for data capture, with a description of sensor placing along the car and wheels, we propose a methodology to guarantee accurate labelling of the datasets, as well as our preliminary approach to the classification model. The Random Forests classifier, implemented alongside Principal Component Analysis (PCA) to reduce the dimensionality of the dataset and Synthetic Minority Oversampling Technique (SMOTE) algorithm to address class imbalance, shows promising results with a global accuracy of 0.938 and F1-Score values ranging between 0.65 and 0.99. Having shown the potential of this methodology, future steps to improve performance and computational cost are discussed, as well as new use cases to further enhance ADAS capabilities.