In the cutting-edge fields of health monitoring, intelligent security, and human - computer interaction, the technology for recognizing human motion states plays a crucial role. However, there is an inherent contradiction between the continuity of human motion and the discrete data collection, which restricts the accuracy and operational efficiency of recognition algorithms. In light of this, this study innovatively introduces the functional data analysis method.Firstly, it transforms the human motion sequence into a functional form through the innovative application of functional data analysis, and precisely locates the starting point of the cycle based on the minimum extremum point, achieving efficient extraction of single - cycle function data. Then, the functional principal component basis expansion method is adopted to reasonably approximate the motion function, effectively capturing the key features of the data.Moreover, to address the challenge of selecting the number of principal components to truncate, this study innovatively introduces the Stacking ensemble learning strategy, constructing a two - layer model architecture: the output of the first - layer model serves as the input of the second - layer model, replacing the traditional model selection process through model aggregation, significantly enhancing the stability of the prediction model.The effectiveness of the proposed method is verified based on the WARD dataset. Compared with four common ensemble learning models such as principal component analysis combined with GBDT, the results of the Monte Carlo simulation experiments show that the method proposed in this paper demonstrates significant advantages in both recognition accuracy and stability, providing more reliable technical support for the practical application of human motion behavior recognition technology in multiple fields.