Understanding the results of partial penetration laser keyhole welding requires analyzing the relationship between energy absorptance and the keyhole’s complex dynamic behavior. To facilitate such analysis, we propose a deep learning–based absorptance monitoring system that uses only melt pool images, specifically tailored for partial penetration laser welding. The system integrates an object detection model and two convolutional neural network (CNN) regressors. From a single melt pool image, the object detection model estimates the keyhole aperture shape, while the first CNN regressor simultaneously predicts the penetration depth. These two features are then combined to form a two-channel keyhole representation, which serves as the input to the second CNN regressor to predict laser beam absorptance. To evaluate the proposed model, fiber laser welding experiments were conducted on aluminum alloy 1050P-H16 under three different process conditions. While the time-averaged absorptance remained within a consistent range, three distinct deviation patterns were observed, each directly associated with different welding outcomes. Previous absorptance monitoring approaches using melt pool images have been limited to full penetration keyhole welding. In contrast, our study demonstrates that reliable absorptance prediction is also achievable for partial penetration welding. By leveraging visual information from the melt pool, the proposed system enables real-time, image-based monitoring of keyhole absorptance, offering new possibilities for improving process control and weld quality in partial penetration applications.