Data Driven Fast Battery Health Diagnostics

Abstract

Abstract This thesis presents two semi-supervised deep learning diagnostic systems based on Gaussian Process Regression (GPR) combined with Multi-Layer Perceptron (MLP), Bidirectional Gated Recurrent Unit (BiGRU) and Autoencoder (AE). The primary objective is to assess battery health, which is a critical aspect of electric vehicles (EVs) and renewable energy systems. Traditional methods for evaluating battery health are often time-consuming, costly, and not suitable for immediate application. This study aims to overcome these challenges by adopting novel approaches. To address these challenges, the research leverages recent developments in data analytics and machine learning to offer a scalable and effective approach for diagnosing battery health. A key achievement of this study is the use of semi-supervised learning techniques, which effectively utilize a small amount of labeled data together with a large amount of unlabeled data, significantly improving the accuracy of battery capacity estimates. Through the analysis of extensive data from battery usage cycles, the research has developed quick, automated protocols that provide accurate, real-time monitoring of battery health. These systems enhance the management of battery usage in electric vehicles and renewable energy installations while supporting the broader adoption of sustainable energy practices. The performance of both algorithms was evaluated using datasets from NASA and the Technical University of Munich (TUM). The experimental results demonstrate that the BiGRU+MLP+GPR algorithm achieves a Root Mean Square Error (RMSE) of 0.30 on the NASA dataset and 0.10 on the TUM dataset. In contrast, the AE+CNN+GPR algorithm yields RMSE values of 0.52 for the NASA dataset and 0.09 for the TUM dataset. Overall, the BiGRU+MLP+GPR algorithm exhibits greater robustness and stability. In addition to comparing the performance of these algorithms, this study provides a detailed analysis of their specific hyperparameter settings and investigates the impact of dataset availability on algorithm performance.