Towards new statistical modelling techniques combining expert knowledge and experimental data for propulsion systems

Modern automotive powertrain labs create large amounts of data. The data include various key performance metrics, crank angle resolution cycle events and high frequency recordings of all channels in time traces. Historically, the experimental results in the form of lookup tables and scatter plots have not fully exploited the potential of the data and engineers are increasingly focusing on creating statistical models using the available dataset. High quality statistical models can replace some experimental work as the digital twin of physical systems for predictive analysis and can be embedded directly into automotive controllers for model-based control.

With the wide range of modelling tools available, automotive engineers would benefit from a framework of statistical modelling for specific powertrain systems in the form of an automated tool. This PhD will seek to create such a tool with a help of a large commercial database of experimental data. Familiarity with the physical models for individual components (batteries, motors, engines, fuel cells…) should be the starting point of the study. Open source machine learning libraries, such Keras, will then be used to explore the available dataset to investigate the predictive performance of statistical models, such as Neural Networks, compared to the physical models.

The technical know-how generated in this study is expected to provide the tool users with specific guidance such as:

- whether important inputs are missing for specific technologies; 
- how to reduce the number of input variables of the problem for faster model training.
- how to run iterations of training to eliminate irrelevant areas of the problem space and instead focus on areas of special interest.
- which statistical models are most suitable for the specific engineering problem.
- whether physically inspired rules should be included in the ML structure to improve the model performance.

A likely deficiency of this approach in highly non-linear systems is that the density for experimental data needed to allow the training of a Machine Learning structure would be impractical. If this proves to be the case, an alternative approach should be considered that seeks to embody the engineer’s understanding of the physical causality that underlies the unit under test. This can be represented in the form of physically inspired ‘rules’ or ‘toy models’ that can then be calibrated to represent the unit using an iterative training approach. Such a model could allow a more sparse dataset to be used without sacrificing predictive power.

The outputs from this PhD would be expected to be integrated into a model factory engineering software tool that supports engineers in the creation of mathematical models.

This project is offered within the Centre for Doctoral Training in Advanced Automotive Propulsion Systems (AAPS CDT). The centre aims to create a diverse and stimulating environment where you can deepen your knowledge in your discipline through your PhD whilst giving breath to your skills through collaborations.

As a AAPS CDT student sponsored by AVL, you will benefit from the peer support and professional development offered by AVL’s Systems Engineering Lab. In 2014 AVL’s “SE-Lab” was founded as an interdisciplinary communication & collaboration platform for systems engineering. It comprises around 60 students from various studies. A special developed student trainee program provides sustainable basic and advanced trainings to improve systems engineering competencies and prepare young talents for upcoming challenges in a connected world.

Prospective students will be applying for the AAPS CDT integrated PhD programme, which includes a one-year MRes (full time) followed by a PhD programme. The MRes course will be conducted as a cohort with a focus on technology, team-working and research skills. On successful completion of the MRes, you will progress to a PhD programme which can be conducted on a full-time or part-time basis.

AAPS-CDT is determined to create a welcoming and inclusive environment for all members. The whole CDT community will come together at specific events during the calendar year, most notably the induction events, workshops and guest lectures. All new students joining the CDT will be assigned a student mentor. Each student will be assigned a minimum of 2 academic supervisors at the point of starting their PhD.

Funding is available for four-years (full time equivalent) for Home students.

See our website to apply or for more details (go.bath.ac.uk/aaps-cdt).