Electromechanical energy conversion systems, such as electric vehicles, wind turbines, or hydraulic units, are ubiquitous in many industries. A failure in any part of the system can incur considerable economical losses or might endanger the safety of the humans and environment. To minimize these risks, the condition of these systems is continuously monitored for prognosis and diagnosis.
The research project develops novel machine learning methods to automate the condition monitoring of these systems.
We will investigate high-dimensional non-linear models such as artificial deep neural networks, random forests or networked exponential families. These methods allow for a fine-grained fault analysis but are typically data-hungry, and require large amounts of training data. However, labeled training data for fault analysis is scarce since faults are typically rare. Moreover, the generation of synthetic data via simulations if often infeasible due to the high computational complexity incurred by standard methods for multi-body simulations.
This project studies two particular approaches that allow to train accurate non-linear models with small amounts of training data:
1) Transfer learning by pooling the measurements from many different powertrain configurations, which share some characteristics, e.g., using the same type of electrical motor, the same bearing structure, or similar gearbox configuration. Using the pooled data, a rough model is trained, e.g., corresponding to the first few layers of a deep neural network. This pre-trained rough model is then refined by training the last layers using data only from a particular subclass of machines and structures.
2) Data augmentation: Besides data obtained from laboratory experiments, we will use simulation models to obtain synthetic data for different type of faults. This procedure consists in a first stage to simulate the available measured operation cases, identify the difference between them, and then use the deviation to augment the simulation data. We will construct accurate models of the components of the proposed electromechanical systems and integrate them in a modular simulation environment. These models and the simulation environment can be used also for future research on system optimization and control design.
The project is highly multidisciplinary and involves ambitious national and international cooperation, which will lead to future innovations in the field of electrical energy systems.
Funding source
Academy of Finland
Coordinating organisation
Aalto University
People
Tatiana Minav
Associate Professor, Hybrid Drives![](/_next/image?url=https%3A%2F%2Fcontent-webapi.tuni.fi%2Fimage-style%2Fbase_square_lg%2Fproxy%2Fpublic%2Fprofile_pictures%2F5132.jpg%3Fitok%3DOFR3gSkN&w=1200&q=75)
Abid Abdul Azeez
Doctoral Researcher![](/_next/image?url=https%3A%2F%2Fcontent-webapi.tuni.fi%2Fimage-style%2Fbase_square_lg%2Fproxy%2Fpublic%2Fprofile_pictures%2F4622.jpg%3Fitok%3D44LjGwzD&w=1200&q=75)
Viacheslav Zakharov
Doctoral Researcher![](/_next/image?url=https%3A%2F%2Fcontent-webapi.tuni.fi%2Fimage-style%2Fbase_square_lg%2Fproxy%2Fpublic%2Fprofile_pictures%2F2392.jpg%3Fitok%3DM5LiCZXP&w=1200&q=75)
Partners
Contact persons
Tatiana Minav
Associate Professor, Hybrid Drives![](/_next/image?url=https%3A%2F%2Fcontent-webapi.tuni.fi%2Fimage-style%2Fbase_square_lg%2Fproxy%2Fpublic%2Fprofile_pictures%2F5132.jpg%3Fitok%3DOFR3gSkN&w=1200&q=75)