Time-frequency-statistical multi-domain feature fusion for lithium-ion battery capacity estimation in electric vehicles

Accurate estimation of battery state of health (SOH) of power batteries is crucial for reliable assessment of driving range, safety, and service life in electric vehicles (EVs). Since SOH is commonly defined as the ratio of the current capacity to the rated capacity, accurate capacity estimation provides a practical approach for SOH assessment. However, many existing methods rely on a single type of feature, such as microscopic time-series signals or macroscopic statistical indicators, which limits their ability to capture complementary degradation information. To address this issue, this study proposes a time-frequency-statistical multi-domain feature fusion framework for battery capacity estimation using real-world EV data. Time-domain features are extracted from voltage and current sequences using a convolutional neural network-long short-term memory (CNN-LSTM) module, while time–frequency features are obtained through multi-level wavelet decomposition (MLWD). In addition, statistical features describing battery usage patterns are processed using a CNN. These features are fused through an attention mechanism and used to predict battery capacity. Experiments on data from 184 retired EV batteries show that the proposed time-frequency-statistical feature fusion method outperforms baseline approaches. The method effectively overcomes the limitations of single-feature-based capacity estimation and provides a promising solution for battery aging assessment.