Charles Bokor
The primary aim of this investigation is the development of a machine learning based methodology that extracts unknown dynamics and physical laws from systems that exhibit complex/nonlinear and high dimensional traits. This will then be implemented in multi-objective experimental designs and control structures where observability (sensors) and controllability (actuators) are also limiting factors.
Current challenges in this area associated to a lack of interpretability, limited feature extraction, and the number of required data points are known to restrict the application of data driven machine learning to real engineering control and experimental design situations. Optimization techniques encompass a large portion of this problem, with system dynamic identification described above also being highly interlinked to methods for model order reduction, feature extraction, pattern recognition, defining control rules etc. Relevant cases within the experimental/research stages of the engineering process, such as design of experiment, test rig design and calibration, will be focused on.
Consideration will then be given to its intended implementation within a finished control system (controller rules, actuator/sensor impact/fidelity, dynamic Vs pre-recorded data, etc). This will manifest initially with a focus on system dynamics identification using deep learning and ensemble frameworks with simplified test cases. Ultimately a modern direct injection gasoline engine will be utilized as a validation case study for challenging dynamic behaviour. This will allow the exploration of limitations, simplifications and model-reductions that may be required to achieve implementable/realistic performance improvements and/or calibrations given practical factors such as analogue-digital transfer, reduced data points, system complexity, etc.