Abstract: During the butt-welding process of thin-plate aluminum alloys, the base metal was affected by fluctuations in the joint gap and uneven heating and heat dissipation. As a result, the weld was usually defective. A passive vision sensing method was selected to capture the front-side welding images which contained dynamic changes in the molten pool. A weld formation image database for the thin-plate aluminum alloy with a thickness of 3 mm using butt tungsten inert gas (TIG) was established. A double-layer tandem weld formation prediction network model was proposed to predict the weld formation under conditions such as incomplete penetration, normal penetration, over-penetration, burn-through, left misalignment, and right misalignment. The first-layer weld formation prediction network predicted three types of irregular weld formations, such as burn-through, left misalignment, and right misalignment, as well as normal weld formation. The second-layer weld penetration prediction network further classified the images of normal weld formations into incomplete penetration, normal penetration, and over-penetration. Different datasets were used to train the model, respectively. The image-enhanced dataset showed the most excellent performance, and the overall prediction accuracy of the model could reach 95%. Welding tests with varying joint gap fluctuations, heat dissipation, and misalignments were carried out, verifying that the model could accurately classify six types of different weld formations in image sequences.