Synthetic Data Generation Methodologies for Addressing Bias in AI-Driven Battery Thermal Protection Systems
Author(s): Vijayachandar Sanikal
Publication #: 2507023
Date of Publication: 27.07.2025
Country: United States
Pages: 1-11
Published In: Volume 11 Issue 4 July-2025
DOI: https://doi.org/10.5281/zenodo.16500834
Abstract
Bias present in AI models can be a considerable safety and performance concern for battery thermal protection systems in electric vehicles. This paper presents several new synthetic data creation techniques to solve the problem of underrepresentation in critical thermal events. Our data generation methodologies are hybrid, and use modified SMOTE (Synthetic Minority Over-sampling Technique), as well as physics-informed constraints to produce reasonable data for thermal anomalies, while maintaining the statistical features of real-world battery thermal behavior. Our approach to developing synthetic thermal data employs existing knowledge about battery thermal dynamics to sufficiently constrain behavior, ensuring the validity of synthetic data in an array of operating conditions. Experiments completed in a high-fidelity digital twin environment show that AI models developed using our synthetic data changes detection accuracy of thermal runaway precursors by 37% and the time to respond is 42% faster than models developed using only available physical test data. In addition, the approach reduces false positives by 28% in extreme ambient test conditions. Collectively, these results enable greater safety margins, reliability and other improvements to battery thermal protection systems during edge cases which can involve rapid temperatures changes during Direct Current Fast Charging. The methodology provided demonstrates a generalizable framework for overcoming data bias in safety-critical systems for the automotive context while demonstrating reductions in costly physical testing. This research contributes to accelerating the development cycle of robust AI-driven thermal protection systems for electric vehicles.
Keywords: Synthetic data generation, battery thermal protection, AI bias mitigation, electric vehicles, SMOTE, digital twin, thermal anomaly detection, Electric Vehicle, battery.
Download/View Count: 144
Share this Article