AbstractsEngineering

Adaptive Machine Learning for Modeling and Control of Non-Stationary, Near Chaotic Combustion in Real-Time.

by Adam Vaughan




Institution: University of Michigan
Department: Mechanical Engineering
Degree: PhD
Year: 2015
Keywords: Real-time Adaptive Neural Network; Weighted Ring - Extreme Learning Machine; Engine Combustion Cyclic Variability; Nonlinear Model Predictive Control; Dynamical Systems; Chaos Theory; Mechanical Engineering; Engineering
Record ID: 2057953
Full text PDF: http://hdl.handle.net/2027.42/111333


Abstract

Fuel efficient Homogeneous Charge Compression Ignition (HCCI) engine combustion phasing predictions must contend with non-linear chemistry, non-linear physics, near chaotic period doubling bifurcation(s), turbulent mixing, model parameters that can drift day-to-day, and air-fuel mixture state information that cannot typically be resolved on a cycle-to-cycle basis, especially during transients. Unlike many contemporary modeling approaches, this work does not attempt to solve for the myriad of combustion processes that are in practice unobservable in a metal engine. Instead, this work treads closely to physically measurable quantities within the framework of an abstract discrete dynamical system that is explicitly designed to capture many known combustion relationships, without ever explicitly solving for them. This abstract dynamical system is realized with an Extreme Learning Machine (ELM) that is extended to adapt to the combustion process from cycle-to-cycle with a new Weighted Ring-ELM algorithm. Combined, the above techniques are shown to provide unprecedented cycle-to-cycle predictive capability during transients, near chaotic combustion, and at steady-state, right up to complete misfire. These predictions only require adding an in-cylinder pressure sensor to production engines, which could cost as little as $13 per cylinder. By design, the framework is computationally efficient, and the approach is shown to predict combustion in sub-millisecond real-time using only an iPhone generation 1 processor (the $35 Raspberry Pi). This is in stark contrast to supercomputer approaches that model down to the minutiae of individual reactions but have yet to demonstrate such fidelity against cycle-to-cycle experiments. Finally, the feasibility of cycle-to-cycle model predictive control with this real-time framework is demonstrated.