AbstractLithium-ion batteries have become increasingly popular, particularly in electromobility, due to their advantages, such as low costs, high energy/power densities, and low-self discharge rate. Nonetheless, like most electrochemical systems, lithium-ion batteries experience performance degradation over time during both usage and storage, highlighting the importance of assessing the battery's longevity and reliability under operation. Therefore, proper monitoring of lithium-ion battery capacity and precise prediction of battery degradation contributes to maintenance, safety, and asset optimisation and lays the foundation for the technical and economic analysis of potential second-life applications. However, accurately estimating and predicting the capacity of lithium-ion batteries is far from a simple task, as battery ageing is a complex nonlinear process involving various internal mechanisms that are extremely difficult to measure and model with precision. To address this need, this thesis concentrates on developing a framework that accurately estimates and predicts battery capacity by integrating advanced signal processing techniques with data-driven algorithms.
This thesis introduces a multi-domain feature time-frequency analysis and machine learning approach for lithium-ion battery capacity estimate and prediction. Time-frequency analysis is used to reveal hidden characteristics in the recorded nonlinear and nonstationary voltage signal. Specifically, Continuous wavelet transform (CWT) is used to obtain diagnostic features on lithium-ion battery deterioration by converting 1D terminal voltage signals into 2D pictures (wavelet energy concentration). The 2D deep learning convolutional neural network approach extracts battery voltage characteristics from produced images. Extracting characteristics predicts lithium-ion battery capacity. Due to its ability to extract electrochemical characteristics from the non-stationary and non-linear battery signal in both time and frequency domains, time-frequency analysis vividly visualised the battery deterioration process. The suggested technique achieves 95.60% prediction accuracy, showing strong potential for battery management system design.
To prevent battery failure, a general framework that estimates lithium-ion battery capacity regardless of operating parameters, such as cycle current profile and working temperature, must be developed. Thus, a unique generalised lithium-ion battery capacity estimation method based on the adaptive empirical wavelet transform (EWT) and long-short-term memory neural network (LSTM) has been developed. In the time-frequency domain, the adaptive EWT can analyse the charging-discharging terminal voltage signal with non-stationary and transient lithium-ion battery phenomena. 13 statistical characteristics are retrieved to see if capacity decline is correlated with them. After that, the LSTM model estimates LIB capacity. The suggested technique accurately calculates LIB capacity with an average root mean square error of 1.26% and a maximum of 2.74%.
Finally, to ensure safe, dependable, and low-cost operation, lithium-ion battery capacity deterioration must be predicted. Predicting lithium-ion battery capacity deterioration is difficult. Regeneration also affects capacity degradation trajectory prediction accuracy. This thesis presents three time-frequency analysis methods and the nonlinear autoregressive neural network algorithm to increase lithium-ion battery capacity deterioration trajectory forecast accuracy. The time-frequency analysis approach and nonlinear autoregressive neural network predicted with a 2.385% root mean square error and a maximum error of 1.6%.
|Date of Award||12 Jan 2022|
|Supervisor||Musbahu Muhammad (Supervisor), Michael Short (Supervisor) & Maher Al-Greer (Supervisor)|