Open access dataset, code library and benchmarking deep learning approaches for state-of-health estimation of lithium-ion batteries

Abstract

Great progress has been made in deep learning (DL) based state-of-health (SOH) estimation of lithium-ion batteries, which helps to provide recommendations for predictive maintenance and replacement of lithium-ion batteries. However, despite the abundance of articles, few open-source codes are publicly available. While there are several public datasets, they tend to be more oriented toward simulating laboratory environments rather than real-world usage scenarios. Moreover, they solely provide raw data without any corresponding preprocessing codes, resulting in inconsistencies in preprocessing methods across different papers. These reasons lead to unfair comparisons and ineffective improvements. In response to these problems, this paper publishes a large-scale lithium-ion battery run-to-failure dataset, consisting of 55 batteries, and provides a unified data preprocessing method. Besides, we comprehensively evaluate 5 well-known DL-based models to provide benchmark research. To be specific, first, the existing DL-based SOH estimation methods are reviewed in detail. Second, we provide a comprehensive evaluation of DL-based models on 2 large-scale datasets, including 100 batteries, with 3 input types and 3 normalization methods. Third, we make the complete evaluation codes and dataset publicly available for better comparison and model improvement. Fourth, we discuss future DL-based SOH estimation, including unsupervised learning, transfer learning, interpretability, and physics-informed machine learning. We emphasize the importance of open-source code, provide baseline estimation errors (error upper bounds), and discuss existing issues in this field. The code library is available at: https://github.com/wang-fujin/SOHbenchmark.