Development of a Data Preprocessing and Feature Engineering Pipeline for Shear Strength of Shear Wall

Concrete shear wall plays a crucial role in effective distribution of seismic forces, thereby adding significant value to earthquake resistance behavior for a structural frame. Hence, understanding failure patterns and deriving shear strength of shear wall has been a significant point of attention for researchers to understand and develop an effective system for structures located especially in seismically active zones. Current experimental dataset, which includes geometrical and material parameters that are usually limited in size, is found to be insufficient for prediction. Hence, in this study, experimental data from two sources, one comprising 252 shear wall data and other 182, have been combined and augmented using mathematical methods, including quantile jittering and noise injection. This is followed by data outlier treatment via quantile capping, winzoraization, and Interquartile Range (IQR) methods. The resulting dataset now comprises 1736 data points, one output variable, and 13 predictors, which is available for subsequent feature engineering. Web thickness of wall tw is the most influential parameter in feature engineering conducted using univariate feature selection and RFE (regressive feature elimination with random forest). The contribution of tw is found to be significant in Shapley Additive Expansion (SHAP) as well. The 13 predictors have been further reduced to 10 orthogonal components using Principal Component Analysis (PCA). This extensive dataset provides a strong basis for machine learning applications and model development that aim to forecast shear strength with greater precision and generalizability. A reliable tool for creating precise predictive models to understand shear wall behavior has been provided by this dataset.