Abstract: The temperature of lithium battery fire rises rapidly, which is easy to cause the surrounding objects to burn and expand the fire range; At the same time, flammable gases produced by lithium battery fire are prone to explosion risk, which increases the risk of fire. The timely detection of lithium battery fires and the implementation of emergency rescue measures are crucial in preventing chain accidents. This study enhanced the feature extraction ability of the YOLOv5s algorithm by incorporating the Coordinate Attention mechanism, and improved the model's generalization ability through the use of the Mosaic-9 data augmentation algorithm. Additionally, we introduced the CIoU loss function to enhance small target flame detection accuracy. A lithium battery fire identification model based on the improved YOLOv5s algorithm is established, and the loss function and evaluation index robustness of the algorithm model before and after the improvement are analyzed based on the multi-interference lithium battery fire dataset training. The results indicate that the loss value of the model after the improvement has better convergence, and the loss value of the improved algorithm model is lower; compared with the original algorithm model, the accuracy of the improved algorithm model increased by 2.25%, the recall rate increased by 2.11%, the mAP increased by 2.98%, and the F1 score increased by 4.14%. The improved algorithm model achieves the detection speed of 46 frame per second while maintaining a good recognition effect. The establishment of this model has a reference value for the research of intelligent identification of lithium battery fire.

Key words: lithium battery fire, fire identification, YOLOv5s algorithm, CA mechanism, Mosaic-9 data augmentation, CIoU loss function