Battery Usage-Profile Classification via Multimodal Vision Transformer with Feature-Level Fusion and Ensemble Method

Accurate characterization of battery usage profiles is vital for optimizing industrial processes and ensuring the safe operation of battery management systems. However, the nonlinear dynamics of electrochemical processes and multi-source battery data present major challenges for data-driven classification tasks. This paper introduces a multimodal Vision Transformer-based framework that leverages feature-level fusion, Improved Grey Wolf Optimization for feature selection, and an ensemble decision strategy integrating Support Vector Machine and Multi-Layer Perceptron classifiers. In the proposed approach, voltage, current, and temperature time-series data are transformed into image representations using Continuous Wavelet Transform, and modality-specific features are fused before classification. Experimental results demonstrate that the full-modality configuration achieves superior performance, with a test accuracy of 92.69%, F1-score of 91.27%, and recall of 90.70%. Also, for single-modality fusion scenarios the framework achieves up to 86.89% test accuracy, while dual-modality fusion configurations yield results up to 91.50%, confirming the consistent performance gains obtained as additional modalities are integrated. The application of Improved Grey Wolf Optimization for feature selection enhances validation and test accuracies by up to 2.81%. Furthermore, by applying the Synthetic Minority Oversampling Technique to balance the dataset against class imbalance, the final test accuracy reached 92.93%. SHAP-based interpretability analysis reveals that current-derived features contribute most significantly to class discrimination, improving both transparency and feature relevance. The framework maintains consistently high performance across single, dual, and full-modality scenarios, ensuring the robustness and practical potential for real-world battery diagnostics.