Noise-Resilient Deep Learning for Post-Earthquake Strain Prediction in Steel Frames with Setbacks

Abstract

A systematic engineering framework is presented that translates established machine-learning techniques into a noise-resilient deep neural network (NR-DNN) for post-earthquake building assessment using noisy strain sensor data. The proposed model is built upon four major mechanisms: (1) noise‑aware training using a multiplicative synthetic noise model (calibration error, thermal drift, random perturbations), (2) dropout, (3) Bayesian hyperparameter tuning with K‑fold cross‑validation (CV), and (4) ensemble averaging. An internal ablation study is performed to show that simultaneous incorporation of these mechanisms yields the best results. Random Forest (RF) is used to identify the best locations for strain monitoring. The Performance of the model is investigated on two SMRF case studies using nonlinear time history analyses (NTHAs) of 58 ground motion records at immediate occupancy (IO) and life safety (LS) performance levels, plus 43 out-of-sample collapse level records. The model predicts full field strains with a limited number of strain sensors under highly nonlinear structural responses caused by unseen out-of-sample earthquake excitations. Compared with a conventional DNN, the proposed model reduces root mean squared error (RMSE) by up to 90 % under noisy conditions and maintains high damage state classification accuracy. The sensitivity analyses validate the framework's stability both under varying noise levels and under threshold variations in damage state classification. The results confirm that, unlike a DNN trained without appropriate noise-handling mechanisms, the proposed NR-DNN model remains numerically stable under noisy conditions. This validation is numerical, based on a synthetic noise model; experimental field validation is left for future work.