Machine Learning In Structural Engineering

9m ago
1.94 MB
12 Pages

Scientia Iranica A (2020) 27(6), 2645{2656Sharif University of TechnologyScientia IranicaTransactions A: Civil d/Review ArticleMachine learning in structural engineeringJ.P. Amezquita-Sancheza; , M. Valtierra-Rodrigueza , and H. Adeliba. ENAP-RG, CA Sistemas Din amicos, Faculty of Engineering, Departments of Electromechanical, and Biomedical Engineering,Autonomous University of Queretaro, Campus San Juan del Rio, Moctezuma 249, Col. San Cayetano, 76807, San Juan del Rio,Queretaro, Mexico.b. Department of Civil, Environmental, and Geodetic Engineering, The Ohio State University, 470 Hitchcock Hall, 2070 NeilAvenude, Columbus, OH 43220, USA.Received 14 November 2020; accepted 18 November 2020KEYWORDSCivil structures;Machine learning;Deep learning;Structuralengineering;System identi cation;Structural healthmonitoring;Vibration control;Structural design;Prediction.Abstract. This article presents a review of selected articles about structural engineeringapplications of Machine Learning (ML) in the past few years. It is divided into the followingareas: structural system identi cation, structural health monitoring, structural vibrationcontrol, structural design, and prediction applications. Deep neural network algorithmshave been the subject of a large number of articles in civil and structural engineering. Thereare, however, other ML algorithms with great potential in civil and structural engineeringthat are worth exploring. Four novel supervised ML algorithms developed recently by thesenior author and his associates with potential applications in civil/structural engineeringare reviewed in this paper. They are the Enhanced Probabilistic Neural Network (EPNN),the Neural Dynamic Classi cation (NDC) algorithm, the Finite Element Machine (FEMa),and the Dynamic Ensemble Learning (DEL) algorithm. 2020 Sharif University of Technology. All rights reserved.1. IntroductionMachine Learning (ML) is a key Arti cial Intelligence(AI) technology that is impacting almost every eldin a signi cant way from image recognition, e.g., pupildetection [1], multi-object tracking [2], video surveillance [3], multi-target regression [4], thermal infraredface identi cation [5], and human activity recognition [6], to various brain and neuroscience applications,e.g., building functional brain network [7], motor imagery brain-computer interface [8,9], mapping scalp tointracranial EEG [10], seizure detection [11], diagnosisof the Parkinson's disease [12], and characterization ofthe modulation of the hippocampal rhythms [13].*. Corresponding author. Tel.: 52 (1) 427 274 12 44E-mail address: [email protected] (J.P.Amezquita-Sanchez)doi :10.24200/sci.2020.22091In general, an ML system consists of threecomponents: inputs comprising a dataset of signals/images/features, the ML algorithm, and outputwhich is associated with the phenomenon studied (seeFigure 1). ML algorithms can be classi ed into threebroad categories:a) Supervised learning such as Support Vector Machine (SVM) [14], various neural network models,statistical regression, Random Forest (RF) [15],fuzzy classi ers [16], and Decision Trees (DTs);b) Unsupervised learning such as various clustering algorithms, e.g., k-means clustering and hierarchicalclustering [17], autoencoders, self-organizing maps,competitive learning [18], and deep Boltzmannmachine;c) Reinforcement learning such as Q-learning, Rlearning, and Temporal Di erence (TD) learning [19].

2646J.P. Amezquita-Sanchez et al./Scientia Iranica, Transactions A: Civil Engineering 27 (2020) 2645{2656Figure 1. Components of an ML system.The rst journal article on civil engineering application of neural networks was published in 1989 [20].Amezquita-Sanchez et al. [21] presented a review ofresearch articles on neural networks in civil engineeringpublished from 2001 to 2016. This article presentsa review of selected articles on structural engineeringapplications of ML in recent years since 2017. It isdivided into the following topics where most of theML research in structural engineering is published:structural system identi cation, structural health monitoring, structural vibration control, structural design,and prediction applications. In addition, four novelsupervised ML algorithms developed recently by thesenior author and his associates with potential applications in civil/structural engineering are introduced.2. Structural system identi cationassociated with the structural behavior. AmezquitaSanchez et al. [25] presented a robust methodology foridenti cation of modal parameters of large smart structures based on the adroit integration of the multiplesignal classi cation algorithm, the empirical wavelettransform, and the Hilbert transform [26] and applied itto calculate the natural frequencies and damping ratiosof a 123-story super high-rise building structure, theLotte World Tower, the tallest building in Korea (seeFigure 2), subjected to ambient vibrations. The resultsshowed that the proposed approach could identify thenatural frequencies and damping ratios of large civilstructures with high accuracy.For a more robust SSI strategy capable of dealing with the inherent noise, nonlinearities, and uncertainties present in the acquired samples, PerezRamirez et al. [27] combined the empirical mode de-Structural System Identi cation (SSI) is an importanttopic in structural engineering as it allows constructinga mathematical model of a structural system from a setof input-output measurements generated by dynamictime series signals [22]. Perez-Ramirez et al. [23]presented a methodology for identi cation of modalparameters of structures using ambient vibrations andSynchrosqueezed Wavelet Transform (SWT). Jiang etal. [24] introduced a fuzzy stochastic neural networkmodel for nonparametric identi cation of civil structures using the nonlinear autoregressive moving average with exogenous inputs model through the combination of two computational intelligence techniques, i.e.,fuzzy logic and neural networks. The proposed modelwas validated using a 1:20 scaled model of a 38-storeyconcrete building and a benchmark 4-story 2 2 bay3D steel frame.Denoising a signal for an e ective SSI schemecan represent a challenging task because this processcan also inadvertently remove frequency componentsFigure 2. Lotte World Tower in Seoul, Korea.

J.P. Amezquita-Sanchez et al./Scientia Iranica, Transactions A: Civil Engineering 27 (2020) 2645{2656composition [28{30], a recurrent neural network model,Bayesian training [31], and mutual information forresponse prediction of civil structures subjected toextreme loadings. The e ectiveness of the proposedapproach was validated using the experimental dataof a 1:20-scaled 38-story high-rise building structuresubjected to arti cial seismic excitations and ambientvibrations and a ve-story steel frame subjected todi erent levels of the Kobe earthquake.Yao et al. [32] reported blind modal identi cationusing limited sensors and a modi ed sparse componentanalysis. Yuen and Huang [33] introduced a BayesianFrequency-domain substructure Identi cation. Yuenet al. [34] described self-calibrating Bayesian real-timesystem identi cation. Tian et al. [35] discussed systemidenti cation of pedestrian bridges using particleimage velocimetry.3. Structural Health Monitoring (SHM)Structural Health Monitoring (SHM) continues tobe the subject of intensive research in structuralengineering. It can be divided into two categoriesof image-based SHM employing the computer visiontechnology and vibration signal-based SHM basedon the signals obtained during dynamic events.The latter in turn can be divided into two generalapproaches: parametric system identi cation (modalparameters identi cation) and non-parametricsystem identi cation. ML algorithms have been usedextensively in both types of SHM.3.1. Vibration signal-based SHMSHM based on the non-parametric system identi cation approach consists of two main stages of feature extraction/selection and classi cation. The feature/patterns identi ed in the rst step are employedfor designing and training a machine learning algorithm2647with the goal of determining the health condition of thestructure in an automated manner.Kosti c and G ul [36] combined an autoregressivemodel with exogenous inputs with a Multi-Layer Perceptron Neural Network (MLPNN) for damage detection of a simulated footbridge structure at varyingtemperatures. Pan et al. [37] evaluated three timefrequency methods, Wavelet Transform (WT), HilbertHuang Transform (HHT), and Teager-Huang Transform (THT) for identifying features in measured signalsin combination with SVM using a simulated cablestayed bridge.Incipient or light damage represents a challengefor identi cation. Yanez-Borjas et al. [38] proposedthe fusion of statistical indices and a decision tree fordetecting damage due to corrosion in a 3D 9-bay and169-member truss-type bridge subjected to dynamicexcitations. The authors reported that the proposalcould identify light damage due to external corrosion,causing 1 mm reduction in the bar element diameter. Amezquita-Sanchez [39] integrated the Shannonentropy index with a decision tree for evaluating themeasured responses of a 1:20 scaled model of a 38storey concrete building structure under di erent levelsof damage produced by cracks.The aforementioned works have exhibited advances in SHM; however, they require a hand-craftedfeature extraction approach to e ective classi cation inthe subsequent step [40]. In recent years, deep learningalgorithms such as Convolutional Neural Networks(CNNs) have been employed for automatic feature extraction in SHM. In these methods, feature extractionand classi cation steps are performed in a single stepto avoid the exhaustive tests between features andclassi ers (see Figure 3) [41,42]. Abdeljaber et al. [43]explored a 1D-CNN for determining the condition of abenchmark 4-story 2 2 bay 3D steel frame subjectedto ambient vibrations. Krishnasamy and Arumulla [44]Figure 3. (a) Traditional machine learning and (b) deep learning.

2648J.P. Amezquita-Sanchez et al./Scientia Iranica, Transactions A: Civil Engineering 27 (2020) 2645{2656combined a second-order blind identi cation with WTand autoregressive time series models for detectingminor incipient damage such as subtle cracks in a beamsubjected to dynamic excitations.The e ective training of supervised ML approaches requires a large set of data from healthyand damaged structures. To overcome this limitation,unsupervised ML-based methods have been proposedrecently because they do not require labeling the training data from di erent damage scenarios. Ra ei andAdeli [45] presented a novel unsupervised deep learningmodel for global and local health condition assessmentof structures using ambient vibration response throughintegration of SWT, fast Fourier transform, and deeprestricted Boltzmann machine. The model extractsfeatures from the frequency domain of the recordedsignals automatically.Ibrahim et al. [46] compared the classi cation performance of three machine learning algorithms, SVM,K-Nearest Neighbor (KNN), and CNN, for evaluatingthe health condition of two simulated four- and eightstory building structures subjected to earthquakes.They reported that CNN outperformed SVM and KNNin terms of accuracy for damage detection. Zhanget al. [47] discussed vibration-based structural stateidenti cation by a one-dimensional CNN. Huang etal. [48] presented a multitask sparse Bayesian learningfor SHM applications.Wang et al. [49] described shear loading detectionof through bolts in bridges using a percussion-basedone-dimensional memory-augmented CNN. NaranjoPerez et al. [50] presented a collaborative machinelearning-optimization algorithm to improve the niteelement model updating of structures. Their proposalconsists of the harmonic search and active-set algorithms, multilayer perceptron neural networks, and theprincipal component analysis, where advantages suchas the computation time, robustness and e ectivenessof an actual steel footbridge model are obtained. Fromthis work, it is observed that the combination of severalmachine-learning algorithms and other mathematicaltools can lead to more powerful solution methods.Wang and Cha [51] combined a deep autoencoder, an unsupervised deep learning method, with aone-class SVM for vibration-based health monitoring ofa laboratory-scaled steel bridge. The authors reportedan accuracy rate of 91% for light damage detection.Sajedi and Liang [52] discussed the vibration-basedsemantic damage segmentation for SHM.3.2. Image-based SHMCha et al. [53] presented the autonomous structuralvisual inspection using region-based deep learningfor detecting di erent types of damage. Gao andMosalam [54] employed a deep transfer learning forimage-based structural damage recognition. Zhang etal. [55] described a context-aware deep convolutionalsemantic segmentation network for detecting cracks instructures. Wu et al. [56] discussed pruning CNNsfor e cient edge computing in health condition assessment of structures. Nayyeri et al. [57] described aforeground-background separation technique for bridgecrack detection.Deng et al. [58] employed CNN for concrete crackdetection with handwriting script interferences. Panand Yang [59] described a post-disaster image-baseddamage detection of reinforced concrete buildings usingdual CNNs. Liu et al. [60] reported an image-basedcrack assessment of bridge piers employing UnmannedAerial Vehicles (UAVs) and 3D scene reconstruction.Jiang and Zhang [61] also discussed a real-time crackassessment using deep neural networks and wallclimbing UAVs.Athanasiou et al. [62] outlined a machine learningapproach to crack assessment of reinforced concreteshells using multifractal analysis as a feature extractor.4. Vibration control of structuresDynamic loadings such as tra c, wind, and seismicactivity generate vibrational responses that can negatively a ect the integrity of a structure.