Machine learning-supported and data-driven enhancements of distributed infrastructure seismic resilience
Thesis DisciplineCivil Engineering
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
This dissertation presents four research elements that aim to improve the seismic resilience of communities. Two of these elements are dedicated to directly improve the resilience of buried infrastructure networks (in particular, the water supply networks), whereas the two others are devoted to ground motion records analysis and immediate regional ground motion intensity estimation.
Utilizing pipe repair data collected following the 22 February 2011 Mw, the ﬁrst research element introduces new empirical fragility functions for buried pipelines. These new functions integrate a quantiﬁable soil liquefaction susceptibility metric into their functional form. This enables the fragility assessment of water supply network assets in liquefaction-prone areas without having to estimate the liquefaction severity.
Complemented by additional datasets, the same pipe repair data is used in the second research element to infer the historical recovery of the water supply network, and to develop an optimization method to accelerate its post-earthquake functional recovery. The historical recovery inference provides both temporal and geographical insights on the eﬃciency of the inspection and repair operations. The optimization method is based on a mixed integer linear programming whose solution is approached using a genetic algorithm. Results show that the recovery was carried out eﬃciently but could have been optimized by using the proposed optimization method.
To enhance the immediate earthquake source identiﬁcation and estimation of regional ground motion intensity, the third research element provides a framework to compare recorded ground motions at instrument stations with physics-based ground motion simulations from the Cybershake New Zealand v18.6 program. This framework is composed of two main parts based on machine learning algorithms: (1) an earthquake source discriminator based on a gradient boosting machine, and (2) a ground motion map generator based on a predictive generative network. Applied to the recent and complex 14 November 2016 Mw 7.8 Kaikōura earthquake, this method delivers immediate results that are superior to contemporary counterparts, and could provide valuable insights for precise source inversion.
The last research element proposes an automated quality classiﬁcation of ground motion records. The method utilizes a neural network that non-linearly combines ground motion records quality metrics such as the shape of the Fourier spectrum or the several signal-to-noise ratios and is trained on manually classiﬁed data from two New Zealand regions. The automated method achieves a human- comparable classiﬁcation quality and its eﬀect on some intensity measures is thoroughly analyzed to ensure that it does not introduce any biases.
In aggregate, the work presented in this dissertation provides new methods that could beneﬁt communities in the pursuit of seismic resilience and provide a solid incentive to foster the application of machine learning technique in the earthquake engineering research ﬁeld.