A TinyML Soft-Sensor for the Internet of Intelligent Vehicles

The increased number of sensors in modern cars offers the opportunity to develop algorithms that can monitor and diagnose vehicle performance more efficiently. We present the results of applying and deploying a TinyML model into a typical OBD-II automotive scanner to serve as a soft-sensor and estimate carbon dioxide emissions. A TinyML workflow based on TensorFlow, TensorFlow Lite, and Micro was designed to a 32-bit microcontroller target (Machhina A0
^TM
) and considering different quantization methods and Multi-layer Perceptron Regressors (MLP). Train, test, and validation were conducted using real-world data fetched from several kinds of vehicles through an emission measurement system. The results suggest that the soft-sensor can estimate Carbon Dioxide emissions with a Mean Absolute Percentage Error (MAPE) of approximately 27% and processing time averages around 37 to 173 microseconds (depending on activation functions adopted) in the target hardware and using intake manifold absolute pressure, intake air temperature, and vehicle speed as independent variables. The results of this study also demonstrated quantization has a major impact on memory usage. On average, 10 to 17 times less memory is required to achieve the same result on MAPE.