Safety Warning of Lithium-Ion Battery Energy Storage Cabin by Image Recogonition

Abstract

Lithium-ion battery will emit gas-liquid escapes from the safety valve when it gets in an accident. The escapes contains a large amount of visible white vaporized electrolyte and some colorless gas. Effective identification of the white vaporized electrolyte and an early warning can greatly reduce the risk of fire, even an explosion in the energy storage power stations. In this paper, an early warning method of lithium-ion battery fire is proposed, which is based on gas-liquid escape image recognition. Firstly, an image recogonition algorithm based on the YOLOv3 is proposed. The original Darknet53 feature extraction network in the algorithm is replaced with a lightweight ReXNet feature extraction network, considering the safety requirements of fast and accurate identification of the storage. In addition, the K-means clustering algorithm is used to obtain appropriate initialized anchor boxes to speed up the convergence of the model. Finally, the multi-scale feature fusion is combined with the path aggregation network to improve the detection accuracy of the model, so that the model can achieve good recognition of both large and small targets. The results show that the method shows a good effect in identifying the vaporized electrolyte of the actual lithium-ion battery storage. The model prediction speed tested on the GTX1650 graphics card can reach 65 frames per second, and the average accuracy is 84.35%. It basically meets the needs of practical applications. The research in this paper can further improve the safety of lithium-ion battery energy storage power stations and promote the healthy development of electrochemical energy storage.