SLI_FRM: Multi-scale real-time detection of lithium battery swelling and early smoke based on YOLOv8

With the growing concern for safety performance in applications of lithium-ion batteries such as electric vehicles and laptops, batteries often exhibit dangerous signals like SWELLING and early smoke under abnormal conditions such as high temperature, overcharging, or over-discharging. If these signals are not detected in time, they can lead to thermal runaway or explosions, causing severe safety accidents. Addressing the limitations of existing visual detection methods in accuracy and speed, this paper proposes the SLI_FRM model based on YOLOv8 (You Only Look Once version 8). The design integrates three targeted modules: HEM (Hierarchical Enhancement Module) for multi-scale enhancement, Slimneck for lightweight deep–shallow fusion via VoVGSCSP (Visual Object Volumetric Generalized Spatial Cross Stage Partial) and GSConv (Ghost-Shuffle Convolution), and FRMHead (Feature Refinement Module Head) for channel–space joint optimization, collectively improving detection of small, low-contrast, and blurred-edge defects. On a self-constructed bulge and smoke dataset, SLI_FRM achieves a Precision of 0.944(+9.6%), Recall of 0.960(+3.8%), mAP@0.5 of 0.956(+2.9%), 100 FPS (frames per second) inference speed, and 0.7 PPI (Power Performance Index) energy efficiency; on the AOD-4(Airborne Object Detection) airborne object detection dataset, it attains 0.919 Precision and 133 FPS, showing strong cross-domain generalization. The proposed unified joint detection framework meets stringent industrial requirements (“inference speed > capture cycle”) and supports zero-modification deployment in YOLOv8-based production lines, offering a robust and high-precision solution for real-time battery safety monitoring.