Percepción infrarroja en tiempo real para el diagnóstico térmico de baterías de iones de litio

Accurate thermal fault localization in densely packed lithium-ion battery packs is significantly challenged by mutual cell occlusion, hindering traditional single-perspective monitoring and thermal runaway prevention. To overcome this, we propose a novel two-stage thermal characterization and localization framework. Its core innovation uses real-time camera rotation to capture multi-perspective surface thermal images, effectively circumventing occlusion. Stage 1 uses a detection transformer (DETR) for coarse target characterization via video instance query and frame tracking. Stage 2 achieves precise cell-level fault localization using a 3D incoherent region detector combined with point quadtree-based mask fine-adjustment. Tested on Ansys Fluent-generated battery models under varying ambient temperatures, the system provides robust, accurate end-to-end thermal fault cell positioning under occlusion, achieving 0.78 mAP, 0.75 mAR, and 8.0 FPS inference speed with a 4.5G-parameter model. The refinement of the segmentation boundary in this work makes it clearer and provides a new paradigm for battery fault location using thermal imaging.