As a direct reflection of the interaction between the stirring tool and the base metal in the friction stir welding process, the force signal is an important means to characterize welding quality. In this paper, the variation mechanism of three-dimensional force and its relation with welding quality were explored. The acquired signals were subject to interference from high-frequency noise, so mean filtering and variational mode decomposition were applied to obtain the real signals. The denoised signals were analyzed and the results showed that the traverse force was ahead of the lateral force by a ratio of π /4, while the phase difference between the axial force and the other two forces changed with the process parameters. Through application of the least square method and polynomial fitting, the empirical formulas of three-dimensional force were obtained, and these were applicable regardless of tunnel defects. The minimum value of the lateral force increased several times more than that of traverse force when the welding speed increased from 80 mm/min to 240 mm/min. When the pole radiuses of most data points had a value greater than 4, tunnel defects were highly likely to generate. In order to predict welding quality more accurately, a prediction model based on long short-term memory was constructed. The model recognized the various modes of good welds and tunnel defects with 100% accuracy. The identification ability for large and small defects was relatively poor, and the average accuracy of classifying the three categories of welding quality was 84.67%.