AM-MFF: A multi-feature fusion framework based on attention mechanism for robust and interpretable lithium-ion battery state of health estimation

State of health (SOH) is a crucial parameter in a battery management system (BMS). Using multiple sources of data effectively improves the end-to-end SOH estimation performance. However, existing multidimensional feature-based methods fail to fully utilise the intrinsic relationships between multiple sources of data. Meanwhile, most of these methods are not interpretable and overlook the detrimental effects of noise. This study proposes a multi-feature fusion framework based on attention mechanism (AM-MFF) to achieve a robust and interpretable SOH estimation. The AM-MFF integrates the benefits of convolutional neural networks (CNNs) and attention mechanism (AM) to efficiently extract and fuse health features, achieving a comprehensive perception of ageing information. The data from two operational stages are used as inputs and their features are automatically extracted using two independent CNN modules. This design of the AM-MFF overcomes the differences in data distribution and noise levels across inputs. Subsequently, an AM-based feature fusion (AMFF) module is adopted to explicitly capture the intrinsic relationship between features, facilitating efficient multi-feature fusion. The attention scores from the AMFF module elucidate the contributions of each input to the SOH estimation. The experimental results on 130 cells demonstrated the superiority of the proposed AM-MFF over existing single-input and multidimensional feature-based models in terms of accuracy, stability, and noise immunity. The AM-MFF maintained excellent SOH estimation performance even when the data was noisy or the estimation of an input failed. The interpretability results elucidated the rationale behind the excellent performance of the AM-MFF. The reasons for the model anomalies were identified by combining attention scores with battery ageing knowledge.