<p>This study is aimed at addressing the challenges of obtaining high-quality and consistent joints in refill friction stir spot welding (RFSSW) of 2195 Al-Li alloy thin-walled structures. A multi-sensor monitoring system was integrated into the RFSSW equipment to acquire real-time process force signals from the spindle, sleeve, and pin under various conditions, including the standard setting and two typical process anomalies (insufficient preload and sheet gap). The characteristic evolution of these force signals was found to be closely correlated with the four welding stages (preheating, plunging, refilling, and withdrawal), providing insights into the thermo-mechanical response and material flow behavior. A weld quality prediction model was subsequently developed using a sparse autoencoder (SAE) and fully connected neural network (FCNN), which accurately predicted the tensile-shear load of joints, achieving a test set <i>R</i>² of 0.771 with minimal performance drop from the training set (0.815). Furthermore, an adaptive weld quality control strategy was proposed, which would automatically identify non-conforming welds and trigger re-welding operations. Experimental results demonstrated that this strategy effectively restored the weld quality, with all re-welded joints meeting the qualification standards. This research establishes an integrated framework for online quality monitoring and adaptive control in RFSSW, providing a viable solution for ensuring high-quality and consistent welding of aerospace thin-walled structures.</p>