PhD defence by Xin Sui
28.10.2021 kl. 13.00 - 16.00
Xin Sui, AAU Energy, will defend the thesis "Robust State of Health Estimation for Lithium-Ion Batteries Using Machine Learning"
Robust State of Health Estimation for Lithium-Ion Batteries Using Machine Learning
Professor Remus Teodorescu
Associate Professor Daniel-Ioan Stroe
Associate Professor Tamas Kerekes
Professor Huai Wang, Aalborg University, Denmark (Chairman)
Professor Ali Emadi, McMaster University
Professor Xiaosong Hu, Chongqing University
Machine learning (ML) technologies have attracted increasing attention for state of health (SOH) estimation of Lithium-ion (Li-ion) batteries due to their advantages in learning the behavior of nonlinear systems. ML models the battery degradation by mapping external features against the capacity loss, thereby avoiding the battery modeling processes. However, from the application perspective, there are still some challenges that need to be addressed, including the impact of data noise and data size on the estimation performance, the failure of features under variable operation conditions, the dependency on big data, and the difficulty of implementing complex algorithms in low-cost microprocessors. To cope with these issues, a systematic ML-based Li-ion battery SOH estimation framework is developed in this Ph.D. project, which has strong robustness to data size, data noise, and degradation conditions.
As batteries are complex electrochemical systems, their aging process is closely related to the operating conditions, and the SOH feature will be invalid under different conditions. Fuzzy entropy (FE) of voltage, from a short-term pulse test, is proposed as a novel SOH feature. Compared with the traditional sample entropy, the FE-based method is freer in parameter selection, more robust to noise and test conditions, and requires less training data. Furthermore, the interaction between the test conditions, entropy features, and estimation accuracy is analyzed. The results help to select the correct voltage datasets and improve the accuracy of entropy-based SOH estimation. For further improvement of entropy-based SOH estimation, various noise suppression methods are used before and after feature extraction. The experimental results prove the effectiveness of the smoothing step in improving the estimation accuracy and simulation speed of the ML model.
Another solution to avoid the failure of manual-extracted features is to use neural networks especially deep learning methods. This method allows high-dimensional input and automatically-extracted features through hidden layers, but at the same time, this method also relies on relatively large data. Therefore, a bagging-based ensemble method is proposed, which allows the model trained on limited data to achieve higher accuracy and better generalization performance. Finally, the effectiveness of the proposed methods in this project is verified by experiments including the cyclic aging test and calendar aging test under different temperature conditions.
THE DEFENCE will be IN ENGLISH - all are welcome.