PREDICTIVE MODELS FOR PARAMETERS OF SEISMIC SIGNALS BASED ON ADVANCED MACHINE LEARNING

Abstract

In this thesis, advanced machine learning algorithms are used to develop predictive models for forecasting ground motion parameters. The machine learning algorithms used are extreme learning machines (ELM), support vector regression (SVR) and its three variations, namely ε-SVR, ν-SVR and Ls-SVR, decision trees and hybrid algorithm ANFIS (adaptive neuro fuzzy inference system). In this thesis, a novel neuro fuzzy algorithm, RANFIS (randomized ANFIS) is also proposed for forecasting ground motion parameters. This advanced learning machine integrates the explicit knowledge of the fuzzy systems with the learning capabilities of neural networks, as in the case of conventional adaptive neuro fuzzy inference system (ANFIS). In RANFIS, to accelerate the learning speed without compromising the generalization capability, the fuzzy layer parameters are not tuned. The three time domain ground motion parameters which are predicted by the developed predictive models are peak ground acceleration (PGA), peak ground velocity (PGV) and peak ground displacement (PGD). Each ground motion parameter is related to mainly to four seismic parameters, namely earthquake magnitude, faulting mechanism, source to site distance and average soil shear wave velocity. The model is developed using real earthquake records obtained from the database released by PEER (Pacific Earthquake Engineering Research Center) Conventionally, the ground motion parameters are estimated using strong ground motion prediction equations which are also known as attenuation equations. Ground motion prediction equations (GMPEs) are equations that related the ground motion parameter PGA, PGV, PGD to independent parameters like earthquake magnitude, source to site distance and site conditions. They are developed using the traditional regression analysis method. The development of GMPEs involves highly complex computation because of the high nonlinearity and inhomogeneous dependencies among the parameters. The regression analysis is applied for the computation after reducing the complexities by including assumptions. Incorporating the simplified assumptions into modelling leads to very large errors. Thus, there is a huge need for the modelling of ground motion parameters using newer techniques so as to reduce the existing complexities. These overheads are minimized by using advanced learning machines. ii The predictive models for forecasting ground motion parameter, developed using advanced learning machines have many advantages. For modelling using machine learning, it is not required to assume linear dependencies among the variables. Thus, there are no assumptions made and no irrelevant coefficients are required. This makes the predictive models developed using advanced machine learning computationally faster. Moreover, using the advanced learning machines, efficient predictive models with higher precision and lesser error measure is obtained. In this study, all the developed prediction models based on advanced machine learning, are compared to the existing GMPEs as well as the existing benchmark models. The existing GMPE models are Ambraseys et al model [6], Campbell and Bozorgnia model [29] and Smit et al model [142]. The existing benchmark models are ANN/SA model by Alavi and Gandomi, GP/OLS model by Gandomi et al, MEP model by Alavi et al [5] and GP/SA model by Mohammadnejad et al [100]. The quantitative and the qualitative analysis of all the proposed prediction models based on advanced machine learning algorithm shows that the developed prediction models have a good prediction accuracy for the forecasting of ground motion parameter. The significance of the proposed work in this thesis is the application of advance machine learning for faster and easy prediction of the ground motion parameters. The ground motion parameters are the most relevant criteria required for designing any earthquake resistant infrastructure. With growing urbanization, there is tremendous increase in the population density in earthquake prone areas, which in turn is increasing the demand for earthquake resistant structure. All the developed models are tested on the real earthquake data. The database used for modelling is the database known as NGA WEST 1 compiled and systematized by Pacific Earthquake Engineering Research Center (PEER) in 2003 as a part of a project named PEERNGA project. The database file is termed as NGA flatfile V 7.3. The predictive models are trained on 2252 earthquake records and tested on 563 earthquake records. To further validate the efficacy of the proposed models, the models are tested on another set of 140 earthquake records. In this study, the different types of learning methods used are namely neural network learning, kernel method learning, hybrid models and decision tree learning. The hybrid models used in this study are neuro fuzzy techniques which combine the fuzzy logic and neural networks. iii In this study, six different prediction models are proposed. The ground motion parameter prediction model developed based on neural network learning are ANN model, and ELM (extreme learning machines) model. The ground motion parameter prediction model developed based on kernel method learning are ε-support vector regression model, ν-support vector regression and Ls-SVR (least square support vector regression) model. The ground motion parameter prediction model developed based on hybrid models are ANFIS (adaptive neuro fuzzy inference system) model and the novel neuro fuzzy technique, RANFIS (randomized ANFIS) model. The ground motion parameter prediction model developed based decision tree learning is a regression tree model. In this study, a further comparative study of all the developed models is done to deduce the best prediction model. Furthermore a comparative study of the learning effectiveness of each algorithm is done in terms of measure of ‘overfitness’. The overfitness measure is a comparison of the training error with the testing error. This comparative analysis further highlights the advantages and drawbacks of each advanced machine learning algorithm. In this study all the comparisons and conclusions are well validated, as the models are based on real earthquake data, rather than the synthetic data. Furthermore, it is observed that the proposed novel neuro fuzzy technique RANFIS proves to be promising prediction algorithm for forecasting ground motion parameters and hence could be applied to other prediction problem in various domains.