A TinyML Soft-Sensor Approach for Low-Cost Detection and Monitoring of Vehicular Emissions

Area of Science:

• Environmental Science
• Automotive Engineering
• Embedded Systems

Background:

• Vehicles are significant urban air polluters, emitting harmful gases like CO2.
• Accurate indirect measurement of vehicle emissions via OBD-II is challenging due to data complexities.
• Soft-sensor approaches offer a viable solution for estimating emissions from engine data.

Purpose of the Study:

• To develop an embedded soft-sensor system for estimating vehicle CO2 emissions.
• To process engine combustion data retrieved through the OBD-II interface.
• To improve the accuracy of CO2 emission estimation using embedded processing.

Main Methods:

• Utilized an embedded system to interface with vehicle OBD-II ports.
• Developed two distinct algorithms based on available vehicle data for CO2 estimation.
• Implemented an unsupervised TinyML approach for outlier detection and data cleaning.
• Tested feasibility on a Freematics ONE+ board with 1Hz acquisition frequency.

Main Results:

• Demonstrated the feasibility of the embedded soft-sensor system.
• Achieved CO2 emission granularity measurement in gCO2/km.
• The TinyML approach enhanced overall soft-sensor accuracy by removing outliers.
• Real-time monitoring and analysis of CO2 emissions were achieved.

Conclusions:

• The proposed embedded soft-sensor system effectively estimates vehicle CO2 emissions.
• The integration of TinyML enhances data accuracy and system reliability.
• This solution provides a feasible method for real-time, on-board air pollution monitoring.