X.W. Ye, T. Jin and C.B. Yun
effectively grasp the structural service condition and the
characteristics of the long-term deterioration of the target
structure, it will also promptly issue warning information as
well as make decisions regarding inspection, repair and
strengthening (Min et al. 2015, Feng and Feng 2018).
The artificial neural network (ANN) algorithm is a
classical machine learning method, and has been applied to
civil engineering since 1989 (Adeli and Yeh 1989). Early
ANNs were perceptrons with one or two hidden layers, and
had a limited capacity for non-linearity abstraction (Wu et
al. 1992, Szewczyk and Hajela 1994, Yun and Bahng 2000).
Meanwhile, the application frameworks were realized based
on the general-purpose computing languages such as
FORTRAN, MATLAB or C language (Adeli 2001, Ni et al.
2002). Later studies employed the hand-crafted algorithms
to extract features from the data and applied ANNs with a
limited number of hidden layers for classification (Ceylan et
al. 2014); while the capacity of autonomous feature
learning from raw data was not available before the training
of a deep neural network (DNN) (Hinton et al. 2006). In
recent years, along with the significant improvement of
network architecture and computing capacity, deep learning
algorithms, e.g., convolutional neural networks (CNNs),
recurrent neural networks (RNNs), etc., have experienced
rapid growth, and have been applied to automatically
process all kinds of data, especially image data (Dong et al.
2016). Many kinds of DNN frameworks and datasets have
been developed to deal with various data processing
scenarios and to satisfy different types of industrial
demands (Vodrahalli and Bhowmik 2017).
Much research has been carried out to explore the
application of deep learning-based approaches in the field
of the SHM of civil infrastructures (Spencer et al. 2019).
This paper aims to address a review on deep learning-based
SHM of civil infrastructures, and is organized as follows:
Section 2 briefly summarizes the history of the development
of deep learning with incidents of milestones. Section 3
presents the applications of deep learning-based approaches
for SHM on various kinds of civil infrastructures. Section 4
discusses the current key challenges and future trends of the
deep learning-based SHM strategy. Section 5 gives some
conclusions of issues dealt with in the paper.
2. A brief history of deep learning research
2.1 Significant contributions to deep learning
Nowadays, deep learning-based approaches have played
an increasingly important role in the field of image
recognition, natural language processing, recommendation
systems, etc., to execute automated, time-saving and low-
cost operations (Schmidhuber 2015, Goodfellow et al.
2016, Silver et al. 2016). Deep learning is a kind of
representational learning method, which enables a network
architecture to autonomously learn highly abstract features
from raw data to fulfill recognition or classification tasks
(Hinton and Salakhutdinov 2006, LeCun et al. 2015). It is a
branch of machine learning, which belongs to a part of AI.
Machine learning is a process of enabling a computer to
learn hidden patterns among extracted features and targets
for classification or prediction (Lake et al. 2015). Machine
learning algorithms contain ANNs, support vector machines
(SVMs), random forests, decision trees, Bayesian inference,
etc. (Bishop 2006). AI is a system that is able to
demonstrate the intelligence by machines, similar to but not
the same as the natural intelligence of human beings
(Russell and Norvig 2016, Silver et al. 2017), which
contains computer vision, machine learning, robotics,
speech recognition, expert systems, etc. The relationship
among AI, machine learning and deep learning is shown in
Fig. 1.
The development of deep learning has mainly evolved
from the ANN. The basic element of an ANN was called the
neural cell, and has not been changed much since the first
neural cell model, i.e., the MP model, was proposed in 1943
by McCulloch and Pitts (1943). A neural cell with three
input elements and one output element is shown in Fig. 2.
The input elements, i.e., x
1
, x
2
and x
3
, are multiplied by
weights, i.e., w
1
, w
2
and w
3
, for summation, and a bias, b, is
added for modification. An activation function, f(x),
implements nonlinear transformation to generate an output.
Rosenblatt (1958) proposed a single layer perceptron
structure that consisted of multiple neural cells, which could
learn through perceptron convergence algorithms to
improve the capacity for classification. Rumelhart et al.
(1986) applied a back-propagation algorithm to train multi-
layer neural networks, enabling the hidden layers to
construct useful features for classification.