Online Detection of Molten Pool State in Flame Brazing Based on YOLOv12 and Kalman Filter

Welding state detection is essential for ensuring welding quality. Manual inspection suffers from limitations such as subjectivity, poor real-time performance, and difficulty in quantitative evaluation, necessitating algorithmic assistance. To address the highly variable molten pool morphology, intense flame and smoke interference, and strict online detection requirements in flame brazing, this paper proposes a welding state detection method based on YOLOv12 and a Kalman filter. For data acquisition, three filter bands—380–460 nm, 400–480 nm band-pass, and 450 nm center wavelength—were compared, and a 400–480 nm band-pass filter was selected to exploit the molten pool’s blue-violet spectral characteristics and suppress flame and stray-light interference. Algorithmically, a molten pool state detector based on YOLOv12 is constructed, and a constant-velocity Kalman filter temporal enhancement module is further developed to smooth frame-wise detections and compensate for missed detections through prediction-update cycles, thereby improving temporal stability. A narrow-band-filtered molten pool dataset was built, and systematic comparison experiments against representative methods were conducted. Results show that YOLOv12 achieves competitive accuracy, attaining mAP@0.5 of 0.9886 and 0.9548, and mAP@0.5:0.95 of 0.7670 and 0.7525, on the two constructed datasets, respectively. Under the sweeping-occlusion scenario, the proposed method attains a frame detection rate of 99.1%, versus 86.5%–95.5% for the compared detectors without the Kalman filter, demonstrating that the Kalman filter effectively improves temporal stability and robustness without sacrificing accuracy, thereby providing a real-time-capable detection method for online molten pool state detection in welding under the reported laboratory hardware conditions.