YASA (Yokeless and Segmented Armature) Ltd. is a British manufacturer of electric motors and motor controllers that are used in automotive and industrial applications. Founded in 2009, it supplies world-leading Original Equipment Manufacturers (OEMs) in the automotive industry with innovative powertrain solutions that are based on their next-generation axial-flux technology and controllers.
YASA believe realising the full potential of vehicular electrification requires new, more efficient powertrain solutions than radial electric motors that are based on 50+-year-old legacy technology. Their pursuit of this goal forms the basis of the collaborative research project with ASTUTE 2020+ that explored potential improvements to their overall powertrain system through Computational Engineering Modelling techniques.
The project team created simulations of the electromagnetic, structural, thermal, cooling system, power electronic control, and overall powertrain optimisation with the aim of maximising system performance for electric vehicle (EV) applications.
Challenges – Electric Vehicle Powertrain Optimisation
The main engineering challenge faced by the collaborative team was identifying the optimal product technology and topology specification, including its key components and parameters.
Optimising the EV powertrain involved evaluating the performance, targets, and constraints of the studied vehicle model using more than twenty-five independent variables – a process that can have up to 1x1025 solutions and demands a significant amount of computation time and energy. The objective was to find the most suitable parameters for the vehicle motor, inverter, gearbox and powertrain configuration to minimise costs whilst adhering to key constraints such as battery range, acceleration time, and top speed.
This poses a very complex and computationally expensive challenge that requires advanced techniques, methodologies, and expertise in data science and computational modelling to find a viable solution.
The solution for the electric vehicle powertrain optimisation involved a two-step cascade optimisation approach that was based on a data-driven machine learning vehicle model. The model estimated the behaviour of vehicle performance and was able to run optimisation orders of magnitude faster than YASA’s previous optimisation strategy. The cascade optimiser identified the best set of input parameters from genetic algorithms that used the ML-based vehicle model and the proprietary YASA vehicle model.
The team also investigated and demonstrated the performance of a new open-source software, Pyleecan, that was used to design a radial motor. The ASTUTE team studied the software’s capabilities and limitations and found relevant solutions to ensure its successful future use by the industry partner.
Reinforcement Learning: An Alternative to Traditional Machine Learning
Reinforcement Learning (RL) was identified as a way of improving the genetic algorithm-based optimisation process that was used in both YASA’s existing method, and the newly-developed machine learning-based alternative outlined above.
It is a paradigm of machine learning that characterises optimisation problems as an agent interacting with an environment to achieve a goal. The agent executes actions and the environment responds by presenting new states to the agent. It returns a scalar numerical reward signal representing the desirability of the current state, and the agent learns to optimise its received reward by exploring the state-action space.
The RL team provided an alternative optimisation methodology to traditional machine learning tools used for regression, such as xgboost and randomforest approach. This provided an excellent opportunity for both teams to compare their results and led to a significant reduction in computation time.
The Research into Electric Drive Unit System Optimisation resulted in mutual benefits for both partners. Significant knowledge was exchanged in both directions, especially regarding the technical details of electric powertrains for battery electric vehicles and artificial intelligence algorithms applied to modelling complex dynamic systems.
As a direct result of the collaborative project with ASTUTE 2020+, YASA has/are:
- Introduced 3 new manufacturing and prediction processes to their R&D department: machine learning tool, cost prediction tool, and modelling tool.
- In the process of recruiting 3 new full-time members of staff with roles linked to power converter topology optimisation
- In the process of registering 2 new product patents
- Contributed to the global automotive industry’s efforts to reduce CO2 emissions