Machine Learning Approach for Constructing Wet Clutch Torque Transfer Function 2021-01-0712

A wet clutch is an established component in a conventional powertrain. It also finds a new role in electrified systems. For example, a wet clutch is utilized to couple or decouple an internal combustion engine from an electrically-driven drivetrain on demand in hybrid electric vehicles. In some electrical vehicle designs, it provides a means for motor speed reduction. Wet clutch control for those new applications may differ significantly from conventional strategy. For example, actuator pressure may be heavily modulated, causing the clutch to exhibit pronounced hysteresis. The clutch may be required to operate at a very high slip speed for unforeseen behaviors. A linear transfer function is commonly utilized for clutch control in automating shifting applications, assuming that clutch torque is proportional to actuator pressure. However, the linear model becomes inadequate for enabling robust control when the clutch behavior becomes highly nonlinear with hysteresis. The use of linear transfer function also leads to errors in powertrain simulation. This paper presents a machine learning approach to construct non-linear clutch torque transfer functions for robust powertrain control and simulation. Several regression methods are evaluated for accuracy and computational efficiency as compared to the conventional linear fitting approach. It is found that Gaussian Process Regression outperforms other approaches for accurately capturing non-linear clutch behavior during training. It also provides a confidence interval for predicted clutch torque. A simulation study is conducted for a hybrid powertrain system to illustrate the impact of the non-linear clutch transfer function on engine restart behavior, as compared to the conventional linear clutch model. The use of machine learning based regression method enables accurate representation of wet clutch behavior for improved simulation and robust control development.