Integrated feature selection and multi-model ensemble learning for comprehensive battery health state estimation

The nonlinear degradation behaviour, different operating conditions, and measurement uncertainties make it difficult to accurately estimate the battery health indicators in electric vehicles. The traditional single-model methods tend to perform inconsistently on various targets of prediction. The paper will suggest a unified ensemble learning framework in the concurrent prediction of state of health (SoH), remaining useful life (RUL), internal resistance (IR) and capacity with a common feature representation. The framework incorporates six heterogeneous base learners such as recurrent neural networks, convolutional models, and kernel-based regressors to learn various patterns of degradation. Multi-objective optimisation approach using NSGA-II/III is used to identify the best fusion weights within a Pareto set of non-dominated solutions, which makes the accuracy of all indicators balanced. To improve transparency, a feature level analysis that combines both correlation measures and Random Forest importance is added to find physically relevant drivers of degradation. The method is proposed and tested on real-world data gathered on an electric two-wheeler platform, and time-aware validation scheme is used to ensure there is no data leakage. The experimental data indicate that it has a substantial increase in prediction accuracy as well as decreased bias in all outputs in comparison with individual models. Though tested on one platform, the physics-informed feature usage and a data-driven structure enable flexibility to a variety of battery systems and operating conditions.