Auto-Regressive Integrated Moving-Average Machine Learning for Damage Identification of Steel Frames

Abstract

Auto-regressive (AR) time series (TS) models are useful for structural damage detection in vibration-based structural health monitoring (SHM). However, certain limitations, e.g., non-stationarity and subjective feature selection, have reduced its wide-spread use. With increasing trends in machine learning (ML) technologies, automated structural damage recognition is becoming popular and attracting many researchers. In this paper, we combined TS modeling and ML classification to automatically extract damage features and overcome the limitation of non-stationarity. We propose a two-stage framework, namely auto-regressive integrated moving-average machine learning (ARIMA-ML) with modules for pre-processing, model parameter determination, feature extraction, and classification. Based on shaking table tests of a space steel frame, floor acceleration data were collected and labeled according to experimental observations and records. Subsequently, we designed three damage classification tasks for: (1) global damage detection, (2) local damage detection, and (3) local damage pattern recognition. The results from these three tasks indicated the robustness and accuracy of the proposed framework where 97%, 98%, and 80% average segment accuracy were achieved, respectively. The confusion matrix results showed the unbiased model performance even under an imbalanced-class distribution. In summary, the presented study revealed the high potential of the proposed ARIMA-ML framework in vibration-based SHM.