Public defence, Electromechanics, MSc Md Masum Billah

Title of the thesis: Data Augmentation Methods for the Simulation of Energy Conversion Devices

Thesis defender: Md Masum Billah
Opponent: Prof. Pia Lindh, LUT University, Finland
Custos: Prof. Anouar Belahcen, Aalto University School of Electrical Engineering 

Electric motors are essential to modern industry, powering machinery used in manufacturing, mining, and many other applications. Unexpected motor failures can interrupt production, increase maintenance costs, and damage equipment. Machine learning can help detect faults before they lead to serious problems. However, developing reliable diagnostic systems requires data covering different operating conditions and types of faults. Collecting such data experimentally can be expensive and time-consuming, while computer simulations do not always accurately represent real-world measurements.

This thesis investigates methods for generating synthetic data to improve machine-learning-based fault diagnosis in induction machines, a widely used type of electric motor. The aim is to reduce the need for extensive experimental measurements and improve the ability of diagnostic models to recognize faults under operating conditions not represented in their training data.
The research combines computer simulations, experimental measurements, statistical methods, and artificial intelligence to generate additional training data. It examines how uncertainties in physical simulations affect fault diagnosis and develops methods for generating realistic fault-related features and electrical current signals.

The results indicate that carefully generated synthetic data can improve fault classification, particularly when diagnostic models are applied to operating conditions for which little or no experimental data are available. The research also shows how combining physical modelling with data-driven methods can help bridge the gap between simulated and measured data.

These findings help advance the development of reliable machine-learning systems that can detect faults in industrial equipment using less experimental data. The proposed methods could help researchers and industrial developers build more robust fault-detection systems with fewer experimental measurements. In the longer term, such systems could support earlier fault detection, better-informed maintenance decisions, and improved equipment reliability.

Key words: artificial intelligence, condition monitoring, data augmentation, fault diagnosis, induction machines, machine learning, uncertainty quantification

Thesis available for public display 7 days prior to the defence at Aalto University's public display page.

Contact:
www.linkedin.com/in/mdmasumbillah123