Abstract
Lithium-ion batteries are widely used in electric vehicles (EVs) due to the advantage of high specific energy density, long lifespan, and low self-discharge characteristics. However, lithium-ion batteries experience irreversible degradation. This degradation leads to a reduced driving range, decline in power output and can even cause safety hazards. Therefore, it is essential to conduct battery health estimation to ensure efficient performance of battery management system (BMS).Accurate state of health (SOH) estimation for batteries remains a significant challenge due to the complexity of battery behaviour, variations in chemistry, and the limitations of traditional estimation methods. Current methods often rely on extensive charge-discharge cycles or lengthy data collection processes, which limit their ability to provide fast estimates. Furthermore, the existing methods struggle to handle variations in temperature and battery chemistry, leading to poor estimation performance in accuracy and robustness. These research gaps motivate the development of efficient and accurate
battery health estimation methods.
The initial research leverages impedance measurement-based techniques, which provide a fast alternative to traditional methods. An enhanced impedance measurement method is proposed to achieve a significant reduction in estimation time while maintaining accuracy. This method is verified and compared with two prevalent impedance measurement methods through experiments on two test platforms with two different battery chemistries, where it demonstrates effective health estimation performance and can be implemented nearly in real-time.
Further investigation into impedance measurement techniques leads to incorporation of advanced neural network methods to enhance SOH estimation accuracy. Specifically, distribution of relaxation time technique is introduced to interpret battery impedance measurements, which identifies novel feature types. Efficient neural networks including autoencoder and cascade feedforward neural networks further refine valuable features that are difficult to capture through traditional methods, which are then used for accurate SOH estimation.
The next step towards faster estimation is the introduction of meta-learning, which is tailored to a new BMS with limited data. By incorporating relaxation voltage profile, this method achieves rapid adaptation to different battery chemistries and conditions with only 10 samples, significantly improving development efficiency, with a low average mean absolute error of 0.0154 for NCA batteries and 0.0111 for NCM batteries.
Finally, an adversarial domain adaptation transfer learning method is proposed to overcome the shortage of sample labels, which promotes the efficiency of the method development. This method also enables effective adaption to variations in temperature and chemistry, maintaining estimation errors within ±5% for most validated batteries.
In conclusion, this research makes significant contributions to fast battery health estimation by providing innovative solutions focusing on improving efficiency and accuracy. These proposed solutions reduce the need for prolonged estimation time and extensive data collection, providing real-time and robust estimations that meet the demands of modern BMS.
| Date of Award | 9 Jun 2025 |
|---|---|
| Original language | English |
| Supervisor | Fengshou Gu (Main Supervisor), Louie Qin (Co-Supervisor) & Andrew Ball (Co-Supervisor) |