Estimating Light-Duty Vehicle Gaseous Emissions Using a Data-Driven Approach in Off-Cycle Measurements 2024-01-2705
As global regulations on automotive tailpipe emissions become increasingly stringent, developing precise tailpipe emissions models has garnered significant attention to fulfill onboard monitoring requirements without some drawbacks associated with traditional sensor-based systems. Within the European Union, there is consideration of mandating real-time measurement of emission constituents to enable driver warnings in cases where constituent standards are exceeded. Presently, available technology renders this approach cost-prohibitive and technologically challenging, with most sensor suppliers either unable to meet the demand or unwilling to justify the development costs associated with sensor commercialization. Efforts to circumvent the sensor-based approach through first principle models, incorporating thermokinetics, have proven to be both computationally expensive and lacking in accuracy during transient operations.
We propose a data-driven solution based on DL (deep learning) to create virtual sensors capable of accurately estimating instantaneous emissions comparable to fast gas analyzers as an alternative to these approaches. To construct such DL models, a highly accurate dataset is essential for training, validation, and testing. This level of precision was achieved by utilizing a PEMS (portable emissions measurement system) to analyze real-world exhaust stream constituents, complemented by the logging of critical powertrain variables. The data recorded by the PEMS comprises a comprehensive inventory of THC (total hydrocarbons), CO (carbon monoxide), and NO (nitrogen oxide) concentrations in the tailpipe, correlated with engine speed, air intake charge, ignition timing, catalyst temperatures, and other key powertrain signals.
During the collection of emissions and powertrain characteristics, test vehicles were driven over diverse city and highway routes, encompassing various ambient conditions, to create an extensive dataset conducive to training. The generated datasets exclude cold-start events, which are subject to rigorous scrutiny in vehicle certification efforts. Furthermore, the model relies on the closed-loop operation of the fuel control system, which is often not the case during cold start conditions. The trained networks exhibit good accuracy, as R2 and error metrics demonstrate. The resulting data-driven model can be integrated into production vehicles as an independent virtual measuring module or with OBM (onboard monitoring).
