Abstract:

The evaluation of the state of health (SOH) for real-vehicle batteries is challenging owing to poor data quality, inconsistent operating conditions, and limited data utilization. This paper presents a multisource feature extraction and SOH estimation model specifically designed for step-rate charging conditions. First, charging segments are obtained through data cleaning, segmenting, and filling processes. Next, capacity is calculated using data from various current stages, achieving a raw data utilization rate of 96.9%. Compared to methods that calculate capacity within a restricted state of charge (SOC) range, this approach reduces error by over 48.1%. Finally, health factors are extracted based on current operating conditions and historical data accumulation. For current operating condition feature values, dual screening is performed using grey correlation analysis and random forest importance analysis to manage interference. For historical cumulative feature values, Spearman correlation analysis and Kernel Principal Component Analysis (KPCA) are employed to reduce information redundancy. Finally, an attention mechanism and Runge-Kutta optimizer (RUN) are integrated into the Gated Recurrent Unit (GRU) network model. The performance of this optimized model is then compared with five existing models using an actual vehicle operation dataset. The experimental results demonstrate that the optimized model achieves superior estimation accuracy, with an error margin of no more than 0.0086, regardless of whether the test samples include single-stage or multi-stage currents. Additionally, the model shows excellent error convergence as the number of charging cycles increases and effectively predicts the trend of SOH fluctuations.

Key words: real-vehicle battery, step rate charging, SOH estimation, multisource features extraction, Runge Kutta optimization algorithm, machine learning