As the training sample data in deep learning is not always easy to obtain, and the labeling task of sample data is very labor-intensive, weakly supervised learning method as a branch of machine learning has increasingly garnered interest. The scale of training datasets for different actions in sensor-based human activity recognition (HAR) often varies widely. For some action categories with little sample data, it is difficult for conventional learning methods to learn their features, which will lead to poor model performance in practical applications. In weakly supervised learning methods, the model can still achieve excellent performance without the need for a good deal of various label data in the training stage. Therefore, this study will use weakly supervised learning methods to improve low model performance caused by unbalanced samples. Based on autoencoder, which is a weakly supervised learning method, this research proposed a framework for sensor-based HAR to ameliorate the negative effects of the unbalanced distribution of training samples on the model. Results from experiments on open datasets demonstrated that the proposed framework enhances the original model's low accuracy caused by sample imbalance.