In the fields of image recognition and natural language processing, foundation models built with extremely large datasets are being applied to various downstream tasks. However, in the field of sensor-based human activity recognition (HAR), constructing extremely large datasets is challenging, and the sensors used for each task differ, leading to significant variations in the input data formats. These unique challenges make it difficult to realize foundation models in HAR. In this study, I developed a model architecture that can uniformly handle diverse data formats, aiming to create a foundation for pre-training across multiple datasets in the future. I proposed an input combination strategy, a parameter sharing strategy, and an output combination strategy for HAR using multiple sensors, achieving a model capable of handling various inputs. Furthermore, I proposed dynamic grouping and evaluated its effectiveness using a public HAR dataset PAMAP2, revealing a model with improved estimation accuracy and appropriate parameter count due to dynamic grouping.