Online parameterization of battery models using transfer learning of a pre-trained neural network with included online anomaly detection (AILERON)

Context

Electrochemical battery storage systems are a core component for implementing alternative drive systems in the mobility sector. To investigate battery behavior in vehicle integration, models supplement resource-intensive and time-consuming test setups, which are also limited in the possible number of scenarios that can be mapped. In addition, physical models often form the basis for estimation algorithms for various battery condition parameters.
The correct parameterization of the individual elements is crucial for a sufficiently accurate battery model. By using vehicle-specific operating data collected online to determine the model parameters, individual variances within a battery generation are taken into account. As there is also no need for time-consuming preliminary tests in a test environment, the online approach contributes to the efficiency of development, especially in an environment of rapidly changing battery technologies.
Deep learning models can be used for this purpose, which estimate the correct parameter set for the respective battery online. These can be pre-trained using available simulation data and then adapted to the series in the vehicle in an extremely data-efficient manner. For this purpose, a modern training strategy and architecture for the neural network is to be used, which also allows potential malfunctions in the battery to be detected at an early stage and thus also provides a safeguard for the ML component and the energy storage system itself.

Goals

The aim of the thesis is the online parameterization of an electric battery model based on simulated and/or real data. This requires the creation of an ML model with a suitable training strategy, which estimates the electrical parameters of the physical representation of the battery. Subsequently, integration and testing of the anomaly detection will take place and the methodology will be evaluated for multivariate regression.

Requirements

• Good programming skills in a high-level language (Python)
• Knowledge of Tensorflow/Pytorch would be desirable
• Very good knowledge of German or English

In addition, knowledge in one of the following areas is beneficial

• Battery modeling
• Modeling of electric drive systems
• machine learning
• Neural networks
• Multivariate time series modeling
• Ability to work independently
• Team and contact skills
• Conscientiousness and a confident demeanor