Data Analytics in Structural Engineering

The special collection on Data Analytics in Structural Engineering is available in the ASCE Library (https://ascelibrary.org/jsendh/data_analytics_structural).

The term analytics refers to the process of developing actionable decisions or recommendations for actions based on insights mainly generated from historical data. Supported by the fast developments in sensor technologies and data collection, storage, processing, and visualization capabilities, the field of data analytics has been experiencing rapid growth over the past two decades. This growth has resulted in the emergence of several game-changing technologies such as image, voice, text, trend, texture, and behavior recognition/prediction, to name a few. The impact of such technologies has been influencing multiple aspects of society, ranging from targeted marketing to personal computing and communication devices to artificial intelligence-based automation.

Unlike other civil engineering subdisciplines (e.g., water resource and environmental engineering) structural engineering has not taken full advantage of advancements in data analytics. In this respect, data analytics can be broadly categorized into descriptive, predictive, and prescriptive subfields. The tools within these subfields can benefit the structural engineering community in multiple ways. For example, descriptive analytics provide insight in situations where a large set of interacting parameters influence complex phenomena or systems. Predictive analytics provide validated data-driven models that predict challenging scenarios in lieu or in support of mechanics-based models. Such models can be particularly useful in solving structural problems that are too complex to be addressed using traditional techniques when the influence of ill-understood factors is key to accurately capturing system behaviors. Finally, prescriptive analytics provide key answers pertaining to optimum solutions involving multiple parameters and their uncertainties once predictions are made with acceptable accuracy. This can result, for example, in significant savings in structural rehabilitation costs or increased use of specific forms of structural system under certain conditions considering an array of interdependent factors.

I expect data-guided structural engineering to be among the most impactful research areas in our field within the coming decade. The 14 papers included in this Special Collection will accelerate the emergence of this area by showcasing recent advances, trends, and applications as well as highlighting some key challenges. More important, I hope the collection will inspire you to think of ways to harness the power of analytics regardless of your specialty.

The collection opens with a discussion of descriptive analytics by Ezzeldin and El-Dakhakhni (2020) that focuses on meta-researching the entire field of structural engineering using text analytics of more than 11,000 journal articles spanning more than $25\mathrm{years}$. The aim is to visually highlight gaps in structural engineering research, opportunities for collaborative research, and uncharted waters where major research breakthroughs are waiting to happen at the interface between structural engineering and emerging technologies.

Moving from descriptive to predictive analytics, four papers in the collection focus on reinforced concrete (RC) columns and shear walls. Mangalathu and Jeon (2019) compare the efficiencies of different machine learning models, including quadratic discriminant analysis, $k$-nearest neighbors, decision trees, random forests, naïve Bayes, and artificial neural networks, in predicting failure modes of circular bridge columns to facilitate bridge operation and recovery strategies following a seismic event. The models are evaluated using a randomly assigned test set assembled from 311 column specimens. The considered flexure, flexure-shear, and shear column failure modes are classified to higher than 90% accuracy using artificial neural networks. Kakavand et al. (2021) introduce data-driven models for predicting the maximum shear strength of rectangular and circular RC columns. The models are developed using two large experimental data sets for both column types that are randomly partitioned into calibration and validation sets. The developed predictive models are compared with those in standards and other documents, and showed higher accuracy and more superior performance than those achieved by currently available models. Feng et al. (2021) focus on predicting plastic hinge length—a key parameter in seismic design. The developed model, based on an ensemble machine learning algorithm, is trained using the results of 133 tests and its performance is assessed via 10-fold cross-validation. The model’s prediction accuracy is shown to be higher than that of models currently adopted in the literature and simulation using the developed model closely resembles the monotonic and cyclic column behavior observed in laboratory experiments. Finally, Gondia et al. (2020) develop a design expression for RC shear walls with boundary elements using an evolutionary computing technique, genetic programming, in conjunction with a data set created from 254 wall tests, to develop a mechanics-guided data-driven wall shear strength prediction expression. The authors’ analyses show that the developed expression provides predictions with significantly higher accuracy compared with shear strength prediction expressions available in relevant design standards and in the literature.

Extending beyond structural components to structural systems, the collection includes four papers focused on building, bridge, and dam system-level responses. Na et al. (2020) develop a method for automated reconnaissance of building damage from seismic events using smartphones through adopting interstory drift ratio as the damage indicator. Buildings can subsequently be digitally tagged in terms of their damage immediately following a seismic event. The proposed method addresses noise reduction, sliding detection, data fusion, and double integration errors. The authors show that multiple smartphone records are needed to reduce stochastic errors but that current smartphone technology imposes limitations that are expected to be alleviated as the quality of smartphone accelerometers improves. Sun et al. (2020) present a comprehensive review of bridge structural health monitoring (SHM) using data analytics. They highlight computing techniques employed to build data-oriented SHM frameworks and to address computing problems as well as data analysis methods and their applications. They also review deep learning applications in SHM, especially those related to processing unstructured data for visual inspection and structured data for structural damage detection. Pang et al. (2021) introduce a data-driven approach to rapid bridge fragility assessment. Using a Gaussian process regression model, they link bridge seismic damage to parameters including system design and earthquake characteristics. They also adopt a uniform design approach to optimize selection of the model’s training data set. The developed model is validated using data of damaged bridges from the 2008 Wenchuan earthquake and other earthquakes not included in the training data set. Fragility curves for different bridge types are subsequently derived using the developed model and compared with available fragility curves in the literature, demonstrating the accuracy of the authors’ approach. Remaining with the theme of fragility but applied to dams, Segura et al. (2020) offer an alternative to costly simulations of gravity dams under seismic events considering the probabilistic nature of model parameters. By developing a data-driven technique to develop a meta-model that efficiently approximates the seismic response of dams, they generate multivariate fragility functions that offer efficiency while considering the most critical model parameter variation influencing dam response. The model is applied to a dam in northeastern Canada, and the results are compared to current safety guidelines in order to establish a range of model parameter values.

Shifting from seismic to wind hazard, two papers in the collection focus on extreme wind speed and associated vibration predictions. Cui et al. (2021) develop a windstorm identification algorithm based on feature extraction and generalization. Using data from three meteorological stations on the southeast coast of China for the model’s single station and regional-scale predictions, they evaluate the performance of six machine learning algorithms in predicting extreme wind speed under different conditions. They compare the results with those obtained from conventional methods. Finally, the authors quantify the effects on structural design for different return periods. Arul et al. (2020) propose a framework to handle the massive amount of data generated by tall building-monitoring systems, with a focus on wind vortex–induced vibrations (VIVs) and their subsequent impact on component fatigue life. To demonstrate the application of the proposed framework, the authors assess the crosswind fatigue life of the pinnacle of Burj Khalifa, adopting an unsupervised machine learning technique to effectively identify and extract VIV data from the data of other responses recorded by the building’s monitoring system. The extracted data are used to evaluate expected fatigue damage using conventional solutions, demonstrating the applicability of data-driven approaches in other systems’ SHM applications.

In addition to textual and numerical data, the collection includes two papers that employ image data for damage assessment. Mangalathu and Jeon (2020) present a methodology for predicting structural damage through the wavelet transform of ground motions and for assessing the damage state of a structure using convolutional neural networks. They demonstrate the methodology on a low-rise nonductile concrete building and a concrete box girder bridge in California, and damage states are identified with an accuracy of more than 75% in both cases. Keeping with the same theme, Gao and Mosalam (2020) propose a general framework for automated vision-based damage detection that includes eight benchmark classification tasks based on domain knowledge and experience. They conduct benchmark experiments with various data-driven models and training strategies, and apply their framework in postdisaster damage assessment following the 1999 Chi-Chi earthquake to demonstrate the potential for computer vision analysis in SHM applications.

The collection concludes with Dutta and Gandomi’s (2020) use of prescriptive analytics to tackle numerical simulations of large-scale problems involving millions of degrees of freedom and model parameter uncertainty. Their bilevel data-driven modeling framework combines proper orthogonal decomposition and polynomial chaos expansion metamodels as well as a heuristic particle swarm optimization (PSO) technique. The authors demonstrate the effectiveness of their metamodel-PSO combination with two large-scale problems involving structural optimization under uncertainty.

In closure, I view this collection as the seed for future research to harness the power of data analytics in advancing structural engineering across all its current subdisciplines (Ezzeldin and El-Dakhakhni 2020). There are also tremendous opportunities at the interface of structural engineering and numerous technologies currently changing the world around us, where data analytics can play a key linkage role. Whether you choose to learn about and adopt data analytics in your research or venture into a new area, I hope this collection is not only informative but also intriguing and inspiring.