<p>Understanding the dynamic behavior of complex multi-structure systems often requires identifying the dynamic characteristics of their individual components, or substructures. This study introduces a data-driven method for the estimation of substructure frequency response functions (FRFs) using only the responses measured during operation. The approach involves measuring operational responses at joints and target points, and establishing their relationships between them. Deep learning neural networks, enhanced with data augmentation schemes, are used to identify these relationships and, in conjunction with known joint stiffness values, estimate the FRFs. The proposed method is validated using a car body structure and real two-plate structure. Operational responses are calculated numerically or measured experimentally to create a training dataset. Two typical neural networks are trained with these datasets and used to estimate substructure FRFs. The results demonstrate that the proposed method has higher accuracy and more stable performance in estimating substructure FRFs than the conventional pseudo-inverse method even when the contribution information of each channel in the dataset is minimal.</p>