Development of an Artificial Calibrator for Shift Program Optimization to Enhance Fuel Efficiency

The calibration of automotive electronic control units is a critical and resource-intensive task in modern powertrain development. Optimizing parameters such as transmission shift schedules for minimum fuel consumption traditionally requires extensive prototype testing by expert calibrators. This process is costly, time-consuming, and subject to variability in environmental conditions and human judgment. In this paper, we introduce an artificial calibrator – a software agent that autonomously tunes transmission shift maps using reinforcement learning (RL) in a Software-in-the-Loop (SiL) simulation environment. The RL-based calibrator explores shift schedule parameters and learns from fuel consumption feedback, thereby achieving objective and reproducible optimizations within the controlled SiL environment. Applied to a 7-speed dual-clutch transmission (DCT) model of a 1.5 L Mild Hybrid Electric Vehicle (MHEV), the approach yielded significant fuel efficiency improvements. In a case study on a 4.7 km Worldwide harmonized Light-Duty vehicles Test Cycle (WLTC) driving segment, the RL-optimized shift strategy reduced fuel consumption from a baseline of 0.46 L to 0.37 L. Furthermore, when starting from an already optimized shift map representative of a series production vehicle’s calibration, the artificial calibrator further enhanced fuel efficiency, achieving approximately a 0.6 % reduction in fuel consumption for the 4.7 km segment and nearly a 5 % reduction for the full WLTC. The artificial calibrator thus demonstrates a promising methodology to frontload calibration tasks in simulation, thereby offering the potential to reduce reliance on resource-intensive physical testing and to significantly accelerate the development of fuel-efficient powertrain control software.. The direct compatibility of parameter files with real vehicle Electronic Control Unit (ECUs) and the validated SiL behavior suggest high transferability of learned strategies, offering the potential for minimal fine-tuning on physical vehicles post-simulation.