Modeling, control and design of automotive transmission components based on artificial neural network (ANN)

This thesis studies first-principle enhanced hybrid neural network and its application to modeling, control and design of automotive transmission components, including friction components and hydraulic valve. The research objective is to design accurate yet numerically efficient hybrid neural network models that outperform the pure first principle and black-box like neural network models.; An improved hybrid neural network approach is proposed to demonstrate the feasibility of applying explicit first principles to enhance the neural network model performance. To make the resultant model applicable for dynamic simulation with variable sampling time steps, an advanced hybrid neural network (AHNN) architecture is developed. With improvement over the previous hybrid model, the time pattern information is added to the inputs and a simpler architecture is developed through more explicit utilization of the physical laws. It is shown that the AHNN model can compensate for time step variations, and therefore can be easily implemented for dynamic simulations.; To further advance the state of the art of the hybrid approach, new neural network architecture with parallel modules (namely the PMNN model) reflecting the first principle structures, is proposed in this investigation. The PMNN-based friction component model isolates the contribution of engagement pressure on engagement torque while identifying the nonlinear characteristics of the pressure-torque correlation. Furthermore, it provides a simple torque formula that is scalable with respect to engagement pressure. These features make the PMNN model a promising tool for powertrain system simulation and controller design. A pressure-profiling scheme through a quadratic polynomial pressure-torque relationship from the PMNN model is successfully implemented for the transmission shifting optimization.; Non-dimensional artificial neural network ( NDANN) based automotive hydraulic valve fluid field model is developed. A growth and trimming procedure is proposed to identify trivial non-dimensional network inputs and optimize the neural network architecture. A hydraulic valve testing bench is designed and built to provide data for neural network model development. NDANN-based fluid force and flow rate estimator are established based on the experimental data. The NDANN model outperforms the classical equations in that it provides much more accurate prediction under most operating conditions, and captures features missing from the classical equations.