Abstract: Data and features are the foundation of intelligent technologies, but the existing literature on structural intelligent computation rarely covers data-side related studies. Therefore, the research on feature engineering of intelligent computation in civil engineering was carried out, which automatically nondimensionalize the raw data of structural problems and transform them into effective features, thus improving the performance of the model. A feature engineering architecture independent of the downstream intelligent computation models was established. The input structural features were automatically nondimensionalized by introducing a dimensionless preprocessing net based on dimensional analysis and using logarithmic activation functions. On this basis, an algorithm was proposed to interpret the physical meaning of the dimensionless parameters obtained by model training, which could perform the factor analysis on the input data and enhance the physical interpretability of the model. In order to validate the proposed model and algorithm, numerical experiments regarding the biaxial bending of reinforced concrete columns were conducted. Compared with the control model without feature engineering, this architecture could speed up the model convergence rate by 4~5 times and improve the prediction accuracy rate by 20%~50%. At the same time, the dimensionless parameters reproduced by the physical meaning interpretation algorithm were highly consistent with the classical theoretical analysis conclusions, which proved that the feature engineering architecture successfully captured the influencing factors closely related to the target problem.