<p>Keyhole plasma arc welding (K-PAW) is a high-efficiency welding method that relies on the formation and maintenance of a fully penetrated keyhole to ensure process stability and weld quality. However, direct observation of backside keyhole geometry is often infeasible in industrial environments due to harsh imaging conditions and spatial limitations. To address this challenge, this study proposes a vision-based deep learning framework for temporal regression prediction of backside keyhole feature variables from topside weld pool image sequences. The keyhole area (KA) and keyhole deviation distance (KDD) are defined as quantitative geometric indicators of penetration status, which are calibrated from backside images using a trained adaptive semantic segmentation model. To extract and integrate spatial–temporal relationships, a CNN-LSTM architecture is designed with a MobileNetV3 backbone and two-layer LSTM units. Independent regression heads are constructed for KA and KDD to improve learning stability. Experimental evaluation shows that the proposed CNN-LSTM model reduces the MAE and improves <i>R</i>² compared to CNN-only baselines, highlighting the benefit of temporal modeling for stable KDD variable. To bridge the gap between algorithm development and real-world application, an online deployment system is designed and implemented using a lightweight inference engine. This system achieves real-time prediction of keyhole status and geometric variables with high accuracy and low latency. Experimental results demonstrate that the proposed approach enables stable and accurate online prediction based solely on topside weld pool images, offering a practical solution for data-driven monitoring and control in K-PAW applications without the need for backside visual access.</p>