A digital twin framework for industrial battery fleet management: A unified multi-signal ensemble approach with uncertainty quantification

Electric vehicle battery state estimation for fleet management has recently caught attention due to advancements in storage technology. However, it faces numerous challenges due to dynamic operating conditions, driving behaviors, and environmental factors, as well as the need for real-time, dynamic state estimation algorithms across multi-pack batteries. To overcome these issues, the present study develops a unified digital twin framework integrating multi-signal ensemble estimation via feature engineering, adaptive learning through embedded estimators, and fleet coordination via holistic enterprise control methodology. The proposed model integrates a novel feature engineering mechanism to extract a 13-dimensional signal feature vector from voltage, current, state of charge, and temperature profiles. Then, rigorous mathematical models are developed and analytically combined to merge three estimators, such as Multi-Linear Regression, Polynomial Regression, and Locally Weighted Regression, via an uncertainty-aware ensemble with dynamically adaptive weight assignments and online and offline parameter tuning via gradient descent and convex optimization. Moreover, to account for uncertainty, a quantification mechanism through stochastic dynamics and covariance propagation is implemented to ensure accurate state estimation with guaranteed coverage. The novelty of the proposed mechanism lies in integrating a parallel-processing capability to enable near-linear behavior, supporting time-critical operations for large fleets. The digital twin synchronizes virtual and physical battery states with Lyapunov stability guarantees, while distributed optimization coordinates fleet-level performance. Experiments are conducted, and results are compared with other state-of-the-art algorithms, on battery datasets, demonstrating superior performance, with max., state-of-charge estimation errors of 0.049-0.050%, voltage errors of 0.066-0.0632%, and current errors of 2.264-2.208%, achieving 95% uncertainty coverage and cycle times below 25 ms.