Assessment of seismic damage to civil structures using statistical pattern recognition techniques and time series analysis

Abstract

The ability to estimate seismic induced damage to civil infrastructure is undoubtedly one of the most important challenges faced by structural engineers. In this research two complementary methods of damage estimation using either knowledge of the structure and earthquake or recorded structural responses were investigated. These methods gave different natured estimates, either prediction or detection, which are suitable for different applications. Firstly, damage to a structure was predicted based on analysis of structural and ground motion properties. Secondly, damage to a structure was detected and assessed by analysing the structural response under dynamic excitation. In the first approach, basic structural and ground motion properties were used to characterise a broad group of structures and earthquakes. These properties were used as inputs into a Back-Propagation (BP) Artificial Neural Network (ANN) and related to a damage index that quantified the extent of damage to the structure. A set of prior structural analyses was required to train the ANN before useful predictions could be made. Applied to 2D Reinforced Concrete (RC) frames, the method was capable of predicting with good accuracy damage to frames of varying stiffness, strength and topology whilst subjected to a range of ground motion severities. In the second approach, Autoregressive (AR) models were used to fit the acceleration time histories obtained when the structure was in both undamaged and damaged states. The AR coefficients were selected as damage sensitive features and statistical pattern recognition techniques were investigated for interpreting changes in the values of these features caused by damage. Initially, an offline damage detection method was developed in which BP ANNs were used for both classification and quantification tasks where the percentage remaining stiffness at a specific location was estimated. The method was applied to three experimental structures; a 3-storey bookshelf structure, the ASCE Phase II Experimental SHM Benchmark Structure and a RC column. In addition, for damage classification tasks only, the supervised classification techniques of Nearest Neighbour and Learning Vector Quantisation were found to be effective while Self-Organising Maps, an unsupervised classification method, showed promising results. Finally, an online damage detection method was developed based on recursive identification of the AR models using the forgetting factor and Kalman filter approaches. A linear 3-DOF model with time varying stiffness was investigated and the results showed that damage could be detected and quantified as it occurred. Nonlinear damage detection was addressed with the investigation of a 1-DOF bilinear oscillator and a 3- DOF Bouc-Wen hysteretic system. In both cases the on-set of nonlinearity was detected using Outlier analysis.