Abstract: The molten penetration state is a critical indicator for assessing welding quality and is closely related to the morphology of the molten pool. To address the impact of fluctuating operating conditions on welding quality, this paper establishes a quality monitoring system for the laser welding process based on a set of molten pool profile points. First, a laser welding experimental platform was developed that integrates welding processing, coaxial visual monitoring, and edge computing. Second, to balance operational efficiency, a hybrid algorithm utilizing maximum inter-class variance multi-threshold segmentation based on histogram pyramids and edge detection was developed. This algorithm accurately extracts features from the high-temperature region and the molten pool profile. Additionally, we innovatively employ a multi-level representation based on polar coordinates to characterize the profile, resulting in a set of molten pool profile points. Finally, we developed a deep convolutional network model called VCAS-Net that efficiently and accurately maps molten pool profile point sets to the molten penetration state. The results indicate that this model achieves an identification accuracy of 95.83%, with a processing time of 13.23 ms per image. This performance satisfies the requirements for both high accuracy and real-time monitoring, making it suitable for industrial applications.