Human Activity Recognition (HAR) is a pivotal area of research in machine learning and wearable technology. It has widespread applications in fields such as healthcare monitoring, fitness tracking, and smart environments. Despite significant advancements, existing research faces critical challenges. These include difficulties in recognizing an extended range of activities and achieving high classification accuracy. The challenges are especially pronounced in complex, multi-class activity datasets. Overfitting and the inability of single models to generalize effectively across diverse activities further exacerbate these issues. To overcome these limitations, this study introduces an ensemble learning approach. The method uses the strengths of multiple base models. It uses a meta-model trained to integrate the predictions of the base models optimally. By combining diverse perspectives, the proposed architecture enhances generalization and mitigates overfitting. It also improves robustness in multi-class activity recognition tasks. Experimental results validate the effectiveness of the proposed method. The approach demonstrates superior performance compared to traditional single-model methods. It offers a scalable and reliable solution for complex HAR applications. This contribution marks a significant step toward developing more accurate and adaptive HAR systems. These systems are better equipped to address real-world challenges.