Abstract: To address the issues of slow response speed, low efficiency, and difficulty in adapting to low-power devices in traditional melting pool monitoring methods during high-speed cladding, a lightweight YOLOv5 network is proposed. MobileNetV3 is used to replace the original backbone network structure for the dynamic capture and classification of melting pool states. Additionally, a custom dataset containing 933 melting pool images, created using high-speed cladding equipment, has been used for deep learning training, providing effective support for cladding quality control. Compared with the YOLOv5s baseline, the improved YOLOv5-MT model achieved a mAP@0.5 of 95.9%. While maintaining detection accuracy, the number of parameters was significantly reduced, and the detection speed increased to 288.8 f/s. Compared to mainstream detection models such as YOLOv8, this method offers faster response performance while ensuring accuracy, meeting the practical need for real-time monitoring of melting pool states in high-speed laser processing environments with limited computing resources.