Engineering Structures 318 (2024) 118703

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Structural damage identification by using physics-guided residual

neural networks
Ruhua Wang a, b , Jun Li b, * , Ling Li c, * , Senjian An c , Bradley Ezard c , Qilin Li c , Hong Hao b, d
a
College of Civil and Transportation Engineering, Shenzhen University, Shenzhen, 518060, PR China
b
Centre for Infrastructural Monitoring and Protection, School of Civil and Mechanical Engineering, Curtin University, Bentley, WA 6102, Australia
c
School of Electrical Engineering, Computing and Mathematical Sciences, Curtin University, Bentley, WA 6102, Australia
d
Earthquake Engineering Research and Test Center, Guangzhou University, Guangzhou, China

A R T I C L E I N F O A B S T R A C T

Keywords: In recent years, vibration-based structural damage identification has made significant progress by exploiting
Physics-guided neural networks data-driven deep learning techniques, which can efficiently extract damage-sensitive features from a large
Physics-informed learning amount of data. However, in some practical engineering applications, large volumes of measurement data are not
Deep learning
readily available. This paper proposes a novel physics-guided residual neural network (PhyResNet) framework to
ResNet
Structural damage identification
improve the robustness and accuracy of structural damage identification under data-scarce conditions. In
Structural health monitoring contrast to the state-of-the-art purely data-driven ResNet, the proposed method embedded available physics
Uncertainties knowledge (e.g., governing equations of dynamics) of structures into the feature learning process via a novel
physics-based loss function. The input-output relationship of the network is constrained to retain its physical
meaning implicitly while the demand for large amounts of labeled training data is reduced. Notably, even with
only 5 % of the dataset used for training, PhyResNet achieves a 13.1 % improvement in R-Value. The perfor-
mance of the proposed approach is evaluated through both numerical and experimental verifications. Results
demonstrate that damage localization and quantification are achieved with high accuracies and good robustness.

1. Introduction the current state of structural health can be determined by extracting

damage-sensitive features and patterns from measured vibration char-
Structural damage diagnosis has been widely conducted in the area acteristics. The majority of existing methods for structural damage
of structural health monitoring (SHM) [6,12] to identify the presence, identification focus on extracting such features and patterns effectively
location, and severity of damages in civil infrastructure at an early stage, from the measured data, which can be broadly divided into two cate-
and to forecast the remaining life of structures and assist critical de- gories: physics-based methods and data-driven methods.
cisions about maintenance. Through the advancement of SHM tech- Physics-based methods, either directly or indirectly, exploit the
nology, the industry can significantly reduce maintenance costs, as well physical laws governing the structural behavior in order to extract
as improve the safety and reliability of civil engineering structures. meaningful information about the damage and its evolution from the
Structural condition monitoring /damage identification as one of the measured sensor data. A common practice to evaluate structural con-
essential components in SHM has gained increasing attention from both ditions in physics-based methods is through finite element model
the civil engineering community and the wider asset management so- updating [8]. A representative physics-based model of the structure,
ciety. During the last decade, extensive researches have been conducted typically a finite element model (FEM), is first constructed as the base-
on vibration-based structural damage identification [5,13,38] tech- line. The FEM is then updated to determine unknown characteristics of
niques, and substantial progress has been made in this area [1]. These the structure, such as material properties and boundary conditions, by
techniques rely upon the fact that damage-induced changes in the reducing the discrepancy between the analytical predictions and the
structural properties (stiffness, mass and damping) will result in measurement data. Ultimately, the updated model can be regarded as a
detectable changes in the vibration characteristics (natural frequencies, calibrated numerical model for damage identification. Fritzen and Bohle
mode shapes and acceleration responses) of the structure. Consequently, [9] used this method to detect damage for the STEELQUAKE benchmark

* Corresponding authors.
E-mail addresses: [email protected] (J. Li), [email protected] (L. Li).

https://doi.org/10.1016/j.engstruct.2024.118703
Received 20 February 2024; Received in revised form 6 May 2024; Accepted 25 July 2024
Available online 31 July 2024
0141-0296/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
R. Wang et al. Engineering Structures 318 (2024) 118703

and showed satisfactory localization results. Xia and Hao [35] investi- neural networks in structural damage identification, lack of training
gated the uncertainty effects in experimental measurement and FEM for data (such as highly incompleteness of monitoring information or
damage detection using a perturbation method with two-stage model limited number of sensors) and low-quality data (such as noisy mea-
updating. A sub-structural damage identification method developed by surements) are two main challenges to reliable structural condition
Li et al. [21] also used model updating and has been validated through monitoring with uncertainties in data. Training a reliable machine
both numerical and experimental studies. learning model often requires sufficient data containing rich input-
Recently, Xu et al. [37] highlighted the challenges that traditional output relationships that typically cannot be provided by most engi-
methods face in identifying multi-damage locations and severities in neering tasks. Specifically, the quality of labelled data is a big factor in
large-span spatial structures with complex vibration modes. They training such models, which could have significant impact on the
introduced a structural damage identification (SDI) framework utilizing generalization performance in the external dataset (testing data).
the estimation of distribution algorithm (EDA) for single-layer cylin- Another problem for deep neural networks trained solely on data is that
drical latticed shells. Through numerical simulations and lab tests, the they are often perceived as uninterpretable black boxes that lack phys-
EDA-based approach achieved average identification accuracies of up to ical explicability for many engineering problems.
90 % using numerical data, showcasing its effectiveness in the damage To address the abovementioned challenges in deep learning
identification of these structures. Another notable contribution came methods, one promising solution is to integrate domain/physics
from Zhang et al. [44] who outlined a two-phased strategy for offshore knowledge into the deep learning process, which is also known as
wind turbine damage identification via FEM updating. Initially, their physics-informed learning. Recently, a rising number of researches have
method pinpoints potentially compromised components within the begun to focus on physics-informed learning for a variety of SHM tasks,
turbines, followed by a detailed FEM refinement to precisely determine ranging from dynamical system identification and dynamical response
and quantify the structural damage. These methods for structural simulation to structural damage identification, leveraging scientific
identification or damage identification require accurate initial finite knowledge of structures in various forms. Zhang et al. [41] introduced a
element models of structures for the further model updating to adjust the physics-informed deep learning framework for metamodeling of
parameters of structures, by using optimization algorithms and vibra- nonlinear structural systems. A deep long short-term memory (LSTM) is
tion data. One of significant challenges of physics-based methods is that integrated to two architectures with physical constraints included in the
an accurate initial finite element model is required to serve as baseline, loss function for learning the sequence-to-sequence features from scarce
and measurement noise and uncertainties may affect the performance of data. The model robustness is improved, resulting in more reliable
structural model updating or damage identification significantly. prediction compared with purely data-driven networks. The same au-
On the other hand, deep learning [11] algorithms make use of a large thors [41] also proposed a physics-guided CNN for predicting structural
amount of training data, instead of requiring accurate initial finite seismic responses. Similarly, they formulated available physics knowl-
element models of structures. Abstract features are learned from the edge into the loss function to provide constraints to the network outputs.
training data with different patterns and make predictions based on Lai et al. [20] proposed a physics-informed neural network (PINN)
learning relationship. These training datasets can consider scenarios approach to identify dynamical systems, using Neural Ordinary Differ-
with uncertainties in structures and measurement noise effect in struc- ential Equations (Neural ODEs). In their approach, the Neural ODEs are
tural responses, and output more robust and accurate damage identifi- structured as two-level representations involving a physics-informed
cation results under uncertainty effect. Data-driven methods primarily term and a discrepancy term. Experiments were conducted with a
focus on pattern recognition to determine not only the location and spring-mass model as a 4-degree-of-freedom dynamical system with
severity of damages but also offer significant potential for early damage cubic nonlinearity, along with noisy measured data. Liu et al. [23]
detection. These models are mainly trained on the patterns in the explored physics-based neural networks to identify dynamical systems
measured responses of undamaged and damaged structures. with some sort of parametric uncertainty. Their results demonstrated
Recently, deep learning [11], also known as deep neural networks, that the physics-based models essentially enforce a more structured and
has proven to be a powerful tool for extracting patterns from highly physically interpretable latent space, which can improve and generalize
complex datasets. Deep neural networks have attracted wide attention the predictive capabilities of deep learning-based models. Yu et al. [39]
and been applied in many fields such as computer vision [14], natural proposed a physics-guided machine learning method that directly
language processing [26], and recently a wide variety of engineering encoded underlying physics into recurrent neural network (RNN) [29]
tasks [2,7,10,28,31,32,34]. In particular, deep learning techniques have architecture for structural dynamics simulation. Compared with the
resulted in significant progress in structural health monitoring. For traditional physics-based models and finite element analysis methods,
example, the convolutional neural network (CNN) [19], and its variants, their method is more computational efficient. Another example of a
including the residual neural network (ResNet) [16] and densely con- physics-guided RNN was presented by Yuan et al. [40] in which two
nected neural network (DenseNet) [18] demonstrate exceptional capa- pieces of domain knowledge of wave propagation are employed for
bilities in detecting and identifying structural damages. Lin et al. [22] source localization in impact diagnosis. Eshkevari et al. [4] proposed a
developed a CNN model to automatically extract features and identify physics-based neural network architecture for full state response esti-
damage locations using low-level sensor data. Their results showed high mation of nonlinear systems. The RNN cell in the architecture is
accuracy and good robustness to noisy data. Zhang et al. [42] proposed a designed to resemble Newmark’s method [25] for nonlinear dynamic
one-dimensional CNN to detect tiny local structural stiffness and mass simulation. In extreme nonlinear cases, this physics-based neural
changes in structures by acceleration responses. Their proposed model network showed promising performance with smaller datasets and
has been validated on three real-world structures. Wang et al. proposed variable spaces. Moreover, two recent studies showed that embedding
two deep learning frameworks based on advanced CNN models, ResNet physics knowledge into machine learning is effective for structural
(2021) and DenseNet (2021), to identify stiffness reductions in struc- damage detection. Xu and Noh [36] presented a Physics-Informed
tures. These were crafted to discern stiffness reductions in structures and Multi-source Domain Adversarial Networks (PhyMDAN) for seismic
their validations across both numerical and experimental studies damage diagnosis by transferring the model learned from other build-
emphasized their robustness against uncertainties and noise. In a ings to the target building. In their work, a physics-guided loss function
comprehensive review of vibration-based CNN methodologies for SHM based on fuzzy physical knowledge of buildings is introduced to elimi-
by Avci et al. [1], it was underscored that certain methodologies have nate the uncertainties introduced by multiple source buildings. Zhang
even achieved a striking 100 % accuracy in structural damage detection and Sun [43] combined the data-driven and physics-based structural
applications. damage identification methods via physics-guided machine learning.
Despite the promising progress that has been achieved using deep The output of a physics-based method, i.e., finite element model

2
R. Wang et al. Engineering Structures 318 (2024) 118703

updating, is fed into the neural network as part of features for training. approach in the numerical and experimental studies, respectively; and
However, such features may not be the best representation of physics Conclusions are drawn in Section 5.
rules for structural damage evaluation.
These studies are still far from complete, as physics-informed 2. Methodology
learning in SHM is still in its early stages and is expected to blossom
in the coming years. Particularly, there is limited work incorporating Damage identification of structural systems aims to develop robust
scientific knowledge into deep learning for structural damage identifi- models that effectively capture the underlying nonlinear relationship
cation. One of the challenges in data-driven methods with pattern between vibration characteristics/dynamics responses, and the material
recognition and machine learning techniques pertains to the labeling and/or geometric properties of these structures. It has been reported in
condition of training data. While unsupervised methods have demon- recent studies that deep learning methods such as ResNet and DenseNet
strated some promising performance in detecting early damages, are able to model such relationships successfully, and can thus be used to
detailed damage detection and quantification often necessitate super- develop frameworks for structural damage prediction [30,33]. Deep
vised methods, where the training data comes with proper annotations learning models are generally trained solely on data, which may result in
or labels. Supervised learning usually needs a significant amount of data a poor performance due to data-related issues including lack of sufficient
for training and validation. However, in reality, data samples may be training data or poor quality data such as noise-contaminated data. By
limited, especially for the real data for different damage conditions of incorporating physic laws (e.g., governing equations of dynamics) into
structures. To address these challenges and the limitations in data- the training phase, it is possible to improve the performance and
driven structural damage identification methods, a novel and generic enhance the robustness and reliability of data-driven models. The
physics-guided deep learning framework is proposed in this paper. As embedded physics can inform and guide the training and specify a
illustrated in Fig. 2, the proposed PhyResNet integrates an encoder- feasible training space. For physics-guided deep learning for structural
decoder architecture where a ResNet serves as the encoder and a damage identification, physics is incorporated into a deep neural
physics-based model as the decoder. The encoder endeavors to extract network in the form of the motion equation for structural dynamics. The
damage-sensitive features from the input (e.g., acceleration responses) proposed approach will be presented in the following subsections.
and project them onto a low-dimensional space that signifies structural Firstly the structural equation of motion and damage index formulations
states (e.g., stiffness reductions). In contrast, the decoder reconstructs which represent the physics constraints between the structural states
the acceleration responses based on the learned stiffness reductions. The and dynamic responses are briefly presented. Next, a data-driven
decoder incorporates physics knowledge, framed as governing equations method based on ResNet, is presented. Finally, a physics-guided deep
of structural dynamic behavior. This is implemented as an updated finite learning method is proposed for structural damage identification based
element model where the response output is solved by the Newmark-β on structural dynamics formulations and ResNet.
method. A unique physics-based loss function narrows the gap between
the reconstructed responses and original input. A data-driven loss 2.1. Structural dynamics formulation
further diminishes the divergence between the estimated stiffness re-
ductions and the ground truth. Both these losses steer the optimization In the field of structural engineering, the dynamic behavior of a
of the PhyResNet. The prime goal is to yield an optimal encoder that structure is formulated by the following equation of motion
precisely identifies damage states, fortified by a physics-based decoder,
Mẍ(t) + Cẋ(t) + Kx(t) = F(t) (1)
especially when training data might be sparse or of sub-optimal quality.
The embedded physics knowledge during the training phase implicitly
where M, C and K are the mass, damping and stiffness matrices of the
constrains the ResNet’s input-output relationship, preserving its phys-
structure, respectively; ẍ(t), ẋ(t) and x(t) are respectively the nodal ac-
ical meanings, thereby augmenting its robustness compared to purely
celeration, velocity and displacement vectors of the structure; F(t) is the
data-driven models.
external force vector applied to the associated degrees of freedom (DOF)
The main contributions of this study are summarized below:
of the structure at the time instant t. The damping matrix C is pro-
portional to the combination of the mass and stiffness of the form C =
• A novel approach towards integrating physics knowledge into deep
a1 M + a2 K, where a1 and a2 are the Rayleigh damping coefficients,
learning using an encoder-decoder architecture for structural dam-
when Rayleigh damping is assumed. This equation represents a system
age identification, particularly in scenarios with limited data sam-
of the second-order linear differential equations. It is always possible to
ples. This end-to-end learning structure ensures that damage-
obtain a numerical solution to the second-order differential equations,
sensitive traits are explored and preserved from limited data vol-
by directly using the step-by-step time integration method, such as
umes, while respecting foundational physics laws.
Newmark-β method.
• A physics-based model combined with a numerical differential
Structural damage identification is usually based on the assumption
equation solver (e.g., Newmark-β method) encapsulates the physics
that damage will cause changes in the structural parameters such as
knowledge of structural dynamics. To address the computational
stiffness, mass and damping, which subsequently lead to changes in
inefficiencies inherent in gradient calculations of physics-based
vibration characteristics. In this study, stiffness degradation is used to
methods, a numerical differential equation solver is integrated
define damage for specific elements of structures. A set of scalar vari-
using an automatic differentiation technique, notably the PyTorch
ables zj (j = 1, 2, …, n) ∈ (0, 1) is used to represent the stiffness reduction
autograd engine, thereby significantly enhancing computational
ratio of a given structure, where j is the element number, and n is the
efficiency.
total number of elements. In other words, an elemental damage index is
• Both numerical and experimental studies demonstrate the accuracy
defined by reducing the elemental stiffness as
and performance of the proposed PhyResNet, by conducting studies
( )
on a numerical beam model and an experimental steel frame struc- Kd,j = 1 − zj Kin,j (2)
ture, respectively. Only a small amount of training data is used, with
both the measurement noise and modelling uncertainties taken into where Kin,j is the j-th elemental stiffness matrix under the intact state,
account. From both studies, the PhyResNet surpasses its data-driven and Kd,j is the j-th elemental stiffness matrix under the damaged state.
counterpart, such as the ResNet. Traditionally, these formulations have been used to construct
physics-based models of structures, such as a FEM. One of the most
The rest of the paper is organized as follows: Section 2 describes the popular physics-based methods for structural damage detection is to
methodology; Sections 3 and 4 present the evaluation of the proposed iteratively update the FEM by minimizing the discrepancy between the

3
R. Wang et al. Engineering Structures 318 (2024) 118703

analytical calculation results and the measured responses. Intuitively, by including both weights w and biases b, where biases are typically
enforcing the structural responses (e.g., x(t), ẋ(t), and ẍ(t)) and physical considered constants. Specifically, N is the total number of training
states of the target structure (e.g., stiffness matrix K ) to satisfy the samples, ẍ is the acceleration responses that are used as input to the
governing equation descried in Eq. (1) during training of a neural ResNet, f(.) represents the ResNet which outputs the estimated stiffness
network, the performance and reliability of the learning can be reduction, and z is the true stiffness reduction; λ is the weighting of the
enhanced. L2 regularization term.

2.2. Data-driven deep learning method 2.3. Physics-guided ResNet (PhyResNet) for structural damage
identification
Deep neural networks are defined in a variety of architectures
depending on their applications. Given a sufficient amount of training For a structural damage identification problem, the ResNet intro-
data, deep neural networks can identify damage patterns in civil struc- duced in the previous section learns the input-output relationship (e.g.,
tures by mapping the unknown relationships between the dynamic re- fθ (ẍ)→z) between dynamic responses and the material and/or geometric
sponses and the stiffness reduction parameters. In this paper, one of the properties solely based on data. It can be unreliable when the data is
most prominent neural network architectures, the ResNet, has been noise-corrupted and highly scarce. Beyond the challenges associated
adopted as the backbone for extracting damage-sensitive features in the with data quality and quantity, a purely data-driven neural network
proposed physics-guided deep learning method. confronts an even more fundamental issue: alignment with known
ResNet was proposed by He et al. (2016) to solve the performance physical laws. There exists an inherent risk with data-driven models.
degradation problem exhibited during training of deep neural networks. They may make predictions that, the results while are mathematically
The core idea of ResNet is to explicitly let the layers fit a residual plausible given the input data, could violate foundational principles of
mapping between the input and output through the residual blocks. physics, especially when they are not exposed to a comprehensive rep-
Fig. 1 shows a general residual block. A shortcut connection is intro- resentation of the state space during training.
duced to skip one or more layers, which explicitly lets these layers Addressing the aforementioned challenges requires a more syner-
approximate a residual function F (X). Learning residual mapping has gistic approach. Instead of relying solely on data, there is an imperative
been shown to be relatively easier than directly estimating F (X) + X, to synergize it with known physical laws. Thus, a physics-guided deep
and it can also mitigate performance degradation in very deep archi- learning framework is proposed as illustrated in Fig. 2, which comprises
tectures. Additionally, skip connections provide an alternative path for two main components, namely, an encoder and a decoder.
gradient backpropagation (as shown by the dashed line in Fig. 1) and
alleviate the vanishing gradient issues. By stacking the residual blocks • The encoder uses a deep neural network, such as the ResNet, to
repeatedly, a deep residual network can be formed. In this study, a estimate the stiffness reduction of a target structure from measured
variant of the typical residual block, known as pre-activation residual responses (such as acceleration). This constitutes a forward mapping
block, where batch normalization and ReLU precede each convolution, fθ (ẍ)→z.
is used as the building block of the ResNet models. As experimentally • Conversely, the decoder uses a physics-based model and its solver,
shown in [15], the pre-activation residual block makes training easier such as the Newmark-β method. This enables the decoder to essen-
and improves generalization. tially perform an inverse mapping (g(z) = f − 1 (z)→ẍ), with the FEM
To compare the performance of structural damage identification solver interpreting stiffness reductions back into dynamic responses.
before and after incorporating physics laws into deep learning, a data- The decoder takes the stiffness reductions identified by the encoder
driven model, such as ResNet, is trained for extracting damage- and reconstructs how the structure would respond dynamically (e.g.,
sensitive features and predicting damages purely based on data. A its acceleration) with those damages, using the laws of physics (the
data-driven loss function is defined for optimizing the model governing equations of structural dynamics). This process effectively
simulates the dynamic behavior of the structure given the identified
N
1 ∑ 2 damages.
l D (θ) = ||z − fθ(ẍ)||2 + λ||w||22 (3)
N n=1
In a classical model, without the decoder, the forward mapping is
where the first term is the mean squared error between the estimated learned purely from data, without prior knowledge of structural dy-
stiffness reduction and the ground truth, and the second one is the L2 namics. However, in the proposed PhyResNet, the inverse mapping
weight regularization term. w represents the weights of the neural during training is hardcoded to follow the physics law described in Eq.
network, which are trainable parameters that influence the output based (1). This effectively ensures that the estimated ẍ
̂ is close to the actual
on the input data. θ denotes the entire set of model parameters, ẍ, making the forward mapping intuitively physics-interpretable. To
ensure that the estimated value of z adheres to appropriate physical
constraints, the proposed PhyResNet employs a sigmoid activation
function at the output layer. This confines that the values of z within the
[0,1] range, corresponding to the stiffness reduction range. While the
sigmoid activation serves as a primary regularization strategy, it’s worth
noting that other regularization methods, such as L1 and L2 regulari-
zation, can also be integrated to further enhance the model’s robustness.
These regularization techniques are instrumental in guiding the esti-
mation process, preventing extreme values that might deviate from
physical realities. It’s also important to highlight that the estimation
process for z is iterative in nature. As the PhyResNet is exposed to more
data or undergoes additional training iterations, the estimation of z is
continually refined, culminating in a more precise identification of the
actual damage state.
The primary goal is to optimize the network’s trainable weights and
Fig. 1. A general residual block. bias (θ) such that the PhyResNet can accurately interpret the measured

4
R. Wang et al. Engineering Structures 318 (2024) 118703

Fig. 2. Physics-guided deep learning framework for structural damage identification.

data (e.g., ẍ) while conforming to the physics law in Eq. (1). To this end, calculate the gradient of any mathematical or non-mathematical oper-
a differentiable solver of FEM is implemented in Pytorch. This provides ation of a system. The differentiable FEM solver is implemented through
an efficient way to calculate the derivative/gradients with respect to the the PyTorch [27] autograd engine and the decoder can be automatically
physics-guided loss differentiated in reverse mode to calculate the gradients starting from
N
the output. The physics-based loss only applies during training, when
1 ∑ 2 acceleration responses (ẍ) and forces (F(t)) are needed to satisfy the
l P (θ) = ||ẍ − g(f(ẍ))||2 (4)
N n=1 physics law. After the proposed PhyResNet is optimized, only the
encoder is used to estimate the stiffness reduction. Therefore, only the
employing a mean squared error. measured acceleration is needed for testing, making it more realistic
The overall loss function of the physics-guided ResNet is hence since external forces are usually difficult to obtain. During the training
written as phase, both acceleration responses and forces are utilized to ensure the
l (θ) = αl D (θ) + γl P (θ) (5) model aligns with real-world physics. However, for real-world applica-
tion tests, only acceleration data is required, making it a practical
where α and γ are the hyper-parameters that adjust the weights of data- approach, especially when force data is not readily available. In sum-
driven loss and physics-based loss, respectively. These hyperparameters mary, the integration of traditional physics principles with advanced
enable an optimal balance between the data and physics contributions to data-driven techniques aims to enhance the capability to detect damages
the learning process, ensuring that the proposed PhyResNet is both in structures, especially when only limited dataset samples are available.
robust and grounded in physical realities. All experiments in this study are conducted on a server running Python
3.9 and PyTorch 2.0, equipped with dual Intel Xeon 5218 CPUs, eight
NVIDIA Titan RTX GPUs, and 512 GB RAM.
2.4. Optimization and Implementation
2.5. Merits of the proposed approach
For damage identification, a gradient descent method is usually used
in a neural network to find the optimal solution for θ = The merit of the proposed approach is to conduct the structural
argminfθ (X) → Y, by minimizing the discrepancy between the true damage identification and improve its accuracy given very few available
θ
data, but under significant uncertainties and measurement noise. The
damages Y and estimated damages Y. ̂ To optimize the PhyResNet, the
classical model-based method has several drawbacks. For example, it
gradients of the overall loss described in Eq. (5) are computed. The
needs an accurate initial finite element model. A minor inaccuracy in the
gradients of the first term (e.g., data-based loss) can be easily obtained
finite element model, which is inevitable in reality, may lead to signif-
using PyTorch’s auto differentiation engine. For the physics-based loss,
icant damage identification errors. The model-based method could also
the gradients can be obtained via chain rule, for example,
be vulnerable to modelling uncertainties and measurement noise in the
∂l P ∂l P ∂g(f(xi ) ) ∂f(xi ) vibration data. However, the proposed approach learns the relationship
= • • (6)
∂xi ∂g(f(xi ) ) ∂f(xi ) ∂xi between measured vibration responses and structural damages based on
direct structural response measurements. Therefore the proposed
where xi is the ith input sample. However, the gradients (boxed part) of approach has a much better performance than the classical model based
a complex physics simulator such as FEM are not readily available. method in detecting structural damage. Compared with existing deep
Although one way is to use a finite difference to approximate the de- learning-based methods for structural damage identification, another
rivatives, the computation during training the network can be ineffi- important advancement of the proposed approach is that it can provide
cient, and it is difficult to choose an optimal step size. In order to solve good damage identification results with a significantly smaller number
these limitations, a newly emerging idea called differential program- of samples in the training dataset, by incorporating physics knowledge
ming [3,17,31,32] is explored, which uses automatic differentiation to into the purely data-driven neural networks. These will be demonstrated

5
R. Wang et al. Engineering Structures 318 (2024) 118703

in the following numerical and experimental studies. Table 1

Data simulation scenarios.
3. Numerical studies Scenario Noise type Description

Measurement Uncertanity
In this section, a simply supported beam as shown in Fig. 3 is studied Noise
numerically to validate the performance of the proposed PhyResNet, in
1 No No Clean/Baseline dataset.
comparison with ResNet, for structural damage identification. It consists 2 10 % / 20 % No Measurement noise is added to the
of 10 Euler–Bernoulli beam elements with 11 nodes. Each node has two acceleration responses.
degrees of freedom, namely, a vertical transitional displacement and a 3 No 1–3 % Uncertainty is considered in
rotational displacement. Nodes 1 and 11 of this beam are fixed, and its structural finite element modelling.
4 10 % / 20 % 1–3 % Both Measurement noise and
length, width and height are respectively 20 m, 0.6 m, and 1 m. The
Uncertainty are included.
mass density, Young’s modulus and cross-sectional moment of inertia of
the beam are 2500 kg/m3, 3.3 × 104 MPa and 0.05 m4, respectively.
The first five natural frequencies of the intact beam are 4.12 Hz, that the physics-based loss is employed in training the proposed PhyR-
16.48 Hz, 37.10 Hz, 66.04 Hz and 103.43 Hz. Acceleration responses in esNet for obtaining the optimal network parameters with physics-guided
the vertical direction measured at four nodes 2, 5, 8, and 9 are taken for information. Then the testing datasets are used with the tuned PhyR-
numerical studies. The acceleration responses are generated using an esNet to demonstrate the performance and improvement against the
external force applied at Node 5. Random impact force is applied in this counterpart ResNet. It is noted that the applied force is not required
study, with a Gaussian distribution with a mean of 8000 N and a stan- when using the trained PhyResNet and testing datasets to test the per-
dard deviation of 50. formance of the proposed approach.

3.1. Data simulation 3.2. Model architecture

The acceleration responses are measured along the vertical direction The acceleration responses collected from the four nodes are
at nodes 2, 5, 8, and 9 with a sampling rate of 100 Hz. Both intact and formulated as the input of dimension 100 × 4. A generic architecture is
damaged cases are included during simulating the measurement. To proposed for all the experiments, as shown in Table 2. With only 10
simulate damaged cases, stiffness reductions from 0 % to 30 % with an stiffness elements in this beam structure, a ResNet architecture of a
interval of 0.5 % are considered. Each sample has a ground truth of ten relatively small size is developed for the numerical studies. It should be
variables that represent the stiffness reduction in its corresponding noted that this ResNet is used as an encoder in the framework of the
element. Four simulation scenarios are studied as shown in Table 1. Four proposed approach as shown in Fig. 2. The ResNet architecture with
different scenarios are considered by assuming different uncertainties residual blocks is chosen to effectively capture complex dependencies
and measurement noise levels in the numerical simulations. For each within the structural data while ensuring model generalizability. The
scenario, different damage cases with different locations and extents of number of layers and filters are optimized through trial-and-error based
structural damage are considered to generate the training datasets. It on the available data to achieve the best balance between accuracy and
should be noted that in the numerical studies, for each scenario, 70 % of robustness. A novel nonlinear activation function Mish [24] is used for
training data are used for training, 15 % for validation and 15 for all convolutional layers. It should be noted that sigmoid is selected as the
testing. Therefore, different generated data samples are used to validate activation function of the last fully-connected layer, since its output
the accuracy and performance of the proposed approach for those four range corresponds to the appropriate range of stiffness reduction ratio.
scenarios listed in Table 1. The finite element modelling errors (uncer- The total number of trainable parameters of the used networks is 62,400.
tainty) and noises in the measurement data are considered in different
scenarios in the simulation. Uncertainty is introduced into the FEM by
varying the stiffness parameters of all the elements. The uncertainty
level is quantified by applying a predefined range of variability (e.g.,
1 % to 3 %) to the stiffness parameter values. Acceleration responses are
then obtained by using the FEM including uncertainties in stiffness pa- Table 2
rameters. Additionally, measurement noise is introduced by adding The architecture of ResNet for numerical studies.
10 % or 20 % Gaussian noise into the acceleration responses. These Layer type [Kernel size, # of filters] Output shape
responses are then used as input to the proposed PhyResNet. This study Input - 100 × 4
aims to demonstrate the advantages of incorporating physics laws in the Conv [10 × 1, 32] 20 × 4
[ ]
data-driven method, when relatively limited training data is available. ResBlock1 5 × 2, 32 10 × 2
(BN-Conv-BN-Conv) 5 × 2, 32
Only 2305 samples for each scenario are generated, one sample for each [ ]
ResBlock2 5 × 2, 64 5× 1
damage case. For monitoring the structure, the force should be applied (BN-Conv-BN-Conv) 5 × 2, 64
at the same location in data generation. Each sample is generated with a Average Pooling - 64
randomly selected force from a force vector matrix with a varying Fully-connected 10 10
magnitude following a normal distribution, and the applied impact force
Note: BN - batch normalization, Conv – convolutional layer.
is used only for data generation and not for testing. It should be noted

Fig. 3. A simply supported beam.

6
R. Wang et al. Engineering Structures 318 (2024) 118703

3.3. Performance evaluation metrics measurement noise or uncertainty effect. The performance for this sce-
nario is shown in the “Baseline” column in Table 4. Notably, with only
Two metrics, namely mean squared error (MSE) and regression value very limited amount of data for training, the proposed PhyResNet ach-
R, are chosen to evaluate the overall performance on structural damage ieves an excellent damage identification performance, with R-Value of
identification. MSE measures the distance between the ground truth and 0.890 and MSE value of 1.2 × 10− 4 , while the R-Value and MSE values
the predicted outputs from the trained models. It is considered as an obtained from ResNet are only 0.806 and.
excellent general purpose error metric for numerical predictions, and 2.0 × 10− 4 , respectively, indicating a significant performance
the default loss function for regression tasks in machine learning. improvement by the proposed PhyResNet, especially on the R-Value.
Another evaluation metric, regression value R between 0 and 1, is also Fig. 5(a)-(d) depict a couple of typical damage examples selected from
employed to assess the quality of trained models. Specifically, R-Value the testing dataset to visually demonstrate the performance of the
represents the correlation coefficient, which quantifies the linear cor- trained models for structural damage identification. These samples
relation between the predicted outputs and the ground truth. Generally, include single- and multiple-element damages at various severity levels.
the smaller the MSE or the higher the R-Value, the more accurate the It is evident that the proposed PhyResNet can detect all the damage
trained models are. locations, while the ResNet can only detect part of them. The purely
data-driven ResNet fails to recognize some true damages, and gives some
false positive predictions. With the guidance of the physics laws adopted
3.4. Results
in the model, the PhyResNet corrects these false predictions. With a
limited number of training data samples, i.e., only 5 % generated data
In this section, the performance of the proposed PhyResNet for
samples are used for training, 5 % validation and 90 % testing, the
damage identification is demonstrated. A PhyResNet with the backbone
identified stiffness reductions by the proposed PhyResNet are good and
architecture outlined in Table 2 is constructed to evaluate the effec-
very close to true values.
tiveness of the proposed method. A pure data-driven model, ResNet with
Scenario 2 considers datasets with measurement noises. Two levels
the same configurations, is used as the base model for comparison to
(10 % and 20 %) of measurement noise are included in the dataset. A
evaluate and demonstrate the improved performance of PhyResNet for
similar performance pattern is shown in this scenario that the proposed
different scenarios. To demonstrate the accuracy of the proposed
approach improves the overall performance significantly against the
PhyResNet for damage identification, a situation where training data is
ResNet, in terms of the MSE and R-Values. Using data with 10 % mea-
relatively sufficient is first investigated. A random selection of 70 % of
surement noise, PhyResNet improves the R-Value from 0.789 to 0.870.
the baseline dataset is used for training, and the remaining 30 % is
Even when the noise level reaches 20 %, the R-Value obtained from
divided into two subsets for validation and testing. Afterwards, an
PhyResNet still maintains a decent level, i.e., 0.808, while the R-Value
extreme situation where training data is scarce is investigated to
from ResNet drops drastically to 0.669. Fig. 6(a)-(d) illustrate several
demonstrate the superiority of the proposed method. To simulate a sit-
typical damage cases. In Fig. 6(b)-(d), it can be observed that ResNet
uation with limited training data, the datasets are randomly split into a
fails to predict introduced damage locations accurately, at elements 5
ratio of 5 %:5 %:90 % for training, validation and testing, respectively,
and 10. Meanwhile, PhyResNet reliably identifies damage locations and
i.e., only 5 % of the available data is used for training. Table 3 shows the
levels for all cases considered. Similar results are shown in Fig. 7(a)-(d),
common hyper-parameters for both the data-driven model (ResNet) and
demonstrating that PhyResNet generally produces better damage iden-
physics-guided model (PhyResNet), selected according to the validation
tification results compared to ResNet.
loss and optimized using random search. Two specific hyper-parameters
Scenario 3 considers datasets with uncertainty effect, which is used to
for PhyResNet, such as the weights of data-driven loss (α) and physics-
simulate the modelling error. Based on Table 4, it can be seen that the
based loss (γ) are tuned via validation set as well. In this study, α = 1
ResNet is robust to the uncertainty effect in comparison with the base-
is used for all scenarios. γ is selected between 0.01 and 0.1 for different
line scenario, with an MSE and R-Value of 2.0 × 10− 4 and 0.805,
scenarios. With 70 % of training data available, both PhyResNet and
respectively. Nevertheless, with the guidance of physics law in training,
ResNet models achieve accurate damage identification results. The R-
PhyResNet can still improve the performance further, with a lower MSE
Value and MSE of the proposed PhyResNet are 0.984 and 4.9 × 10− 5 ,
value (1.6 × 10− 4 ) and a higher R-Value (0.850). Four damage examples
respectively, whereas ResNet performs slightly worse, with an R-Value
are shown in Fig. 8(a)-(d) to illustrate the performance of both models.
of 0.983 and MSE of 5.1 × 10− 5 . To visually demonstrate the effective-
Again, the proposed PhyResNet detects all damages accurately whereas
ness of the trained models for identifying structural damage, Fig. 4(a)-
ResNet can only identify some of them, generally with less accurate
(d) present several typical damage examples selected from the testing
levels of severity.
dataset. It is observed that the proposed PhyResNet can identify all
Scenario 4 considers datasets with both measurement noise and un-
damages accurately in terms of both locations and severity, while the
certainty effects, which represents the most challenging task in the nu-
purely data-driven ResNet fails to recognize the true severity of damage
merical study. The datasets include 1 % to 3 % uncertainty, and 10 %
in Fig. 4(d), giving a false positive prediction.
and 20 % measurement noise, respectively. As shown in the last two
With only 5 % training data available, the proposed PhyResNet
columns in Table 4, PhyResNet once again outperforms ResNet in terms
shows a significant improvement over the ResNet. Table 4 provides the
of both the MSE values and R-Values in this most challenging scenario.
damage identification performance comparison between the proposed
With 20 % noise level, the R-Value obtained by the ResNet decreases
method and ResNet. Four different scenarios, as described in Section 3.1,
dramatically to 0.663, the lowest in all scenarios. However, the pro-
are evaluated in terms of MSE values and R-Values on testing dataset.
posed PhyResNet is still able to produce a R-Value of 0.794, comfortably
Scenario 1 is the simplest case in which the datasets do not include
at an acceptable level. Figs. 9 and 10 provide further evidence to support
these observations. Compared with ResNet, PhyResNet identifies dam-
Table 3 ages at the correct locations with severities closer to the true values, with
Common hyper-parameters for training models in numerical studies. much less false positive and false negative predictions. The comparative
Hyper-parameters results further highlight the merits of the deep learning method in
Batch size 128 structural damage identification when integrated with physical laws.
Epoch 200 In summary, the proposed physics-guided PhyResNet is superior to
Optimizer Ranger the purely data-driven ResNet for structural damage identification. It is
Initial learning rate 0.0004
demonstrated that incorporating the physical laws into data-driven
Learning rate scheduling Cosine annealing warm restarts

7
R. Wang et al. Engineering Structures 318 (2024) 118703

Fig. 4. Damage identification results using 70 % data for training.

Table 4
Performance comparison between the proposed method and ResNet under different scenarios.
Baseline Measurement Noise Uncertainty Measurement Noise + Uncertainty

10 % 20 % 10 % 20 %

ResNet MSE 2.0 × 10 − 4

2.2 × 10− 4
3.3 × 10− 4
2.0 × 10− 4
2.2 × 10− 4
3.4 × 10− 4

R-Value 0.806 0.789 0.669 0.805 0.784 0.663

PhyResNet MSE 1.2 × 10− 4
1.4 × 10− 4
2.1 × 10− 4
1.6 × 10− 4
1.4 × 10− 4
2.0 × 10− 4

R-Value 0.890 0.870 0.808 0.850 0.868 0.794

neural networks generally improves performance of structural damage used to validate the effectiveness and robustness of the trained
identification, especially with a limited amount of training data, and networks.
measurement noise and uncertainties are considered in the datasets.

4.1. Data generation

4. Experimental validation

The training datasets for the intact case and damaged cases are
In this section, experimental studies are carried out to further eval-
generated from an updated FEM as described above. A random impact
uate the performance of the proposed PhyResNet against the purely
force is applied at node 44 in the x direction to generate acceleration
data-driven method. A steel frame structure has been fabricated in the
responses, as illustrated in Fig. 11(d). In order to simulate the variations
laboratory and used for the performance evaluation of applying the
in the applied impact force, a force vector is generated with 1–2 %
proposal approach for structural damage identification. In this study, the
variance of the force measured in the experimental tests. During this
acceleration measurements collected from this frame structure in the
simulation, nine nodes are used to measure the acceleration response at
laboratory with measurement and environmental noises are used to test
a sampling rate of 1000 Hz. The sensor locations for the single-element
the performance of the trained models for structural damage identifi-
damage case are 4x, 7x, 9x, 11x, 17x, 47x, 50x, 53x, and 56x, where x
cation. Fig. 11(a)-(c) show the steel frame structure and its experimental
denotes the acceleration response measured in the x-direction. For the
setup for vibration tests. An updated finite element model of the frame
two-element damage case, sensor locations at 4x, 5x, 11x, 14x, 19x, 50x,
structure is developed with 70 elements as shown in Fig. 11(d). Addi-
53x, 56x, and 59x are chosen.
tional details about the model updating are provided in Ref. 9. During
Training samples are generated by first reducing stiffness in the
experimental testings, two damaged cases are conducted to collect the
structural elements randomly from 0 % to 30 %, then the acceleration
experimental data. As shown in Fig. 12, a 12.5 % stiffness reduction is
responses from the selected sensors are obtained. A total of 4131 sam-
initially introduced in the structure to simulate a single-element damage
ples are generated for the single-element damage case and 12,400
case at Element 6 in the steel frame structure. After that, another stiff-
samples for the two-element damage case, respectively. Two sets of
ness reduction with the same ratio introduced in addition to the first
numerical data are obtained to optimize the neural networks for each
damage, at Element 12 in the steel frame structure, providing a two-
case study. To simulate real world situations in which large amounts of
element damage case. Numerical data generated by the FEM are used
training data are unavailable, the neural networks are first trained with
to train and optimize the neural networks, while experimental data are
a small subset of the training data. Furthermore, an ablation study is

8
R. Wang et al. Engineering Structures 318 (2024) 118703

Fig. 5. Damage identification results for Scenario 1: Clean data.

Fig. 6. Damage identification results for Scenario 2: 10 % of measurement noise.

conducted to demonstrate the benefits of the physics-informed learning training set of a different size. Despite the same number of sensors being
mechanism, as implemented in PhyResNet, over conventional data- used in both cases, their locations are different. As a result, two case
driven methods, and to determine under which conditions the PhyR- studies using different datasets are conducted separately. The acceler-
esNet is most effective at improving performance over the purely data- ation response data, collected from laboratory tests, are solely employed
driven ResNet. Numerical data samples are split into three subsets for as testing datasets to provide experimental verifications of the accuracy
training, validation and testing, respectively. The validation set (15 %) of the proposed PhyResNet for structural damage identification.
and the testing set (15 %) are fixed to compare the performance with a

9
R. Wang et al. Engineering Structures 318 (2024) 118703

Fig. 7. Damage identification results for Scenario 2: 20 % of measurement noise.

Fig. 8. Damage identification results for Scenario 3: Uncertainty.

4.2. Model architecture structure. As described in Section 3.2, Mish activation is used for con-
volutional layers, and sigmoid activation is used for the last fully-
The acceleration responses collected from nine nodes are formulated connected layer. The total number of trainable parameters of the used
as the input with a dimension of 500 × 9. It is noted that the trained networks is 3036,610 (~3 M).
neural networks for both case studies share the same architecture as
presented in Table 5. A ResNet architecture with three residual blocks is
developed for experimental validations with a more complex frame

10
R. Wang et al. Engineering Structures 318 (2024) 118703

Fig. 9. Damage identification results for Scenario 4: uncertainty + 10 % of measurement noise.

Fig. 10. Damage identification results for Scenario 4: uncertainty + 20 % of measurement noise.

4.3. Results parameters for the trained models, which are determined based on the
validation loss and optimized through random search. The weight of
This section presents the experimental validation results with two physics loss γ, varies between 0.01 and 0.1, depending on the case study.
case studies. The same metrics as used in the numerical studies are used The weight of data-driven loss α is set as 1.
to evaluate the overall performance. Both the physics-guided PhyResNet
and the data-driven ResNet are trained with the same architecture as 4.3.1. Single-element damage case
shown in Table 5 for comparison. Table 6 shows the common hyper- A comprehensive investigation is conducted with four models

11
R. Wang et al. Engineering Structures 318 (2024) 118703

Fig. 11. Experimental testing model and setup: (a) a steel frame structure; (b) data acquisition system; (c) fixed bottom support; and (d) the finite element model.

Table 5
The architecture of ResNet for experimental studies.
Layer type [Kernel size, # of filters] Output shape

Input - 500 × 9
Conv [10 × 1,32] 100 × 9
[ ]
ResBlock1 5 × 2, 32 100 × 9
× 12
(BN-Conv-BN-Conv) 5 × 2, 32
[ ]
ResBlock2 5 × 2, 64 50 × 5
× 12
(BN-Conv-BN-Conv) 5 × 2, 64
[ ]
ResBlock3 5 × 2, 128 25 × 3
× 14
(BN-Conv-BN-Conv) 5 × 2, 128
Average Pooling - 128
Fully-connected 70 70

Note: BN - batch normalization, Conv – convolutional layer.

Table 6
Common hyper-parameters for training models in the experimental studies.
Hyper-parameters Value

Batch size 128

Epoch 600
Optimizer Ranger
Initial learning rate [0.0002, 0.0004]
Learning rate scheduling Cosine annealing warm restarts

Fig. 12. Damage cases in the frame structure - left: single damage; right:
two damages. Table 7
Performance comparison between PhyResNet and ResNet for the single-element
damage case.
trained for both the PhyResNet and ResNet using four varying sizes of
Networks Metrics 20 % 30 % 40 % 50 %
training data. In the first comparison, only 20 % of the available data (e.
g., 20 % of the initial training dataset of 4131 samples) are used for ResNet MSE 4.4 × 10− 4
2.4 × 10− 4 9.7 × 10− 5
7.1 × 10− 5

training, simulating the absence of large amounts of labelled data in R-Value 0.089 0.683 0.882 0.911
PhyResNet MSE 3.7 × 10− 4
1.4 × 10− 4
8.4 × 10− 5
5.0 × 10− 5
practice. After that, the training size is increased (e.g., 30 %, 40 % and
R-Value 0.401 0.824 0.902 0.932
50 %) to further demonstrate the effectiveness of the proposed

12
R. Wang et al. Engineering Structures 318 (2024) 118703

PhyResNet. Table 7 presents the evaluation results on numerical testing Table 8

data in terms of MSE and R-Value metrics, and the predictions on Performance comparison between PhyResNet and ResNet for the two-element
experimental testing data are shown in Fig. 13. It should be noted that damage case.
the finite element model of the frame structure has 70 elements. How- Networks Metrics 5% 10 % 20 % 30 % 40 %
ever, the introduced structural damages are only present in the 6th and ResNet MSE 8.3 × 7.0 × 4.4 × 3.5 × 3.0 ×
12th elements. Therefore, to show the identification results more 10− 4 10− 4 10− 4 10− 4 10− 4
clearly, only the damage identification results in the first 25 elements R- 0.261 0.467 0.702 0.778 0.816
are shown herein. The R-Value obtained from PhyResNet compared to Value
ResNet is significantly improved from 0.089 to 0.401. The improvement PhyResNet MSE 7.5 × 5.3 × 4.1 × 3.4 × 2.9 ×
10− 4 10− 4 10− 4 10− 4 10− 4
is more significant, when considering that only a very small amount of
R- 0.413 0.640 0.733 0.784 0.821
training data is used. A consistent improvement of performances can be Value
observed with other training sizes, as shown in Table 7. The improve-
ment in performance can also be clearly observed in the testing cases
shown in Fig. 13, demonstrating that physics guidance is extremely data, though the improvement gap decreases with the increment of
useful especially when only limited amount of data is available. For training sizes. In the example cases shown in Fig. 14, it is observed that
example, in Fig. 13(a), the ResNet model fails to predict the true damage the ResNet models cannot locate the damage at Element 6 when the
at Element 12. Rather it makes a significant false positive prediction at training size does not exceed 20 %. PhyResNet, on the other hand,
Element 19. Meanwhile the proposed PhyResNet accurately identifies located both damages precisely, even when the training size is only 5 %.
the location of the true damage, although the severity level prediction is Furthermore, PhyResNet produces a better performance on the overall
not very accurate. When the training size is increased, the ResNet model damage quantification, and results in fewer false predictions for un-
shows some improvement in performance, but PhyResNet still out- damaged elements.
performs ResNet consistently with better identification of locations and
severities of damages, with much less false positive or false negative 4.3.3. Intact case
predictions. It is clearly demonstrated in Table 7 and Fig. 13 that both To further assess the effectiveness of the proposed method, an intact
methods perform better with more training samples, but the physics- case from the test set is presented, as depicted in Fig. 15. In the specific
guided PhyResNet shows a consistent superiority over the purely data- instance where the training size is 5 %, as illustrated in Fig. 15(a), it
driven ResNet, especially when a limited amount of training data is becomes evident that the ResNet model tends to predict more minor
available. damages at undamaged elements. In contrast, PhyResNet provides more
accurate predictions, with false predictions being less than 0.01 %.
4.3.2. Two-element damage case Notably, when operating with limited training data, as depicted in
This is a more challenging case with two actual damages introduced Fig. 15(a)-(c) with 5 %, 10 %, and 20 % of the training data, the pro-
in the frame structure, which is a better reflection of the real world posed PhyResNet consistently surpasses the performance of the ResNet
challenges since structures usually have multiple damages. The same model, resulting in fewer mispredictions and false identifications for the
level (12.5 %) of stiffness reductions is introduced into the 6th and 12th intact case.
elements of the structure. In this case study, even fewer available data
are used for training, ranging from 5 % to 40 %, to evaluate the effec- 5. Conclusions
tiveness and robustness of the proposed PhyResNet in comparison with
ResNet. The obtained MSE values and R-Values are listed in Table 8, This study attempts to improve the robustness and accuracy of deep
showing a similar trend to the one observed in the single-element learning models by incorporating physics knowledge into the purely
damage cases. The proposed PhyResNet consistently outperforms the data-driven neural networks for structural damage identification. A
ResNet over the five cases with 5 %, 10 %, 20 %, 30 % and 40 % training novel physics-guided residual neural network, namely PhyResNet, is
proposed. Physics integration is realized with a physics-based loss that

Fig. 13. Damage identification results for the single-element damage case of the structure using 20–50 % training data.

13
R. Wang et al. Engineering Structures 318 (2024) 118703

Fig. 14. Damage identification results for the two-element damage case of the structure using 5–40 % training data.

Fig. 15. Damage identification results for the intact case of the structure using 5–40 % training data.

14
R. Wang et al. Engineering Structures 318 (2024) 118703

implicitly enforces the input-output mappings following the physics law Declaration of Competing Interest
of motion. To evaluate the performance of the proposed approach, a
numerical case study using a beam structure and an experimental case The authors declare that they have no known competing financial
study using a seven-storey frame structure are conducted. The results interests or personal relationships that could have appeared to influence
from both the numerical and experimental studies demonstrate that the the work reported in this paper.
proposed PhyResNet can improve the robustness and accuracy of
structural damage identification, especially when a limited number of Data availability
training data samples are available. In detail, a simply supported beam is
investigated in the numerical study considering both measurement noise Data will be made available on request.
and uncertainty effects. The proposed PhyResNet greatly reduces the
MSE values and improves the R-Values, with only 5 % of the dataset Acknowledgments
used for training. For instance, it achieves a 13.1 % improvement in R-
Value in the most challenging scenario, which includes 20 % measure- The work described in this paper was supported by Australian
ment noise combined with uncertainty effect. In the experimental study, Research Council Discovery Projects DP210103631, “AI Assisted Prob-
a seven-storey steel frame structure tested in the laboratory is employed, abilistic Structural Health Monitoring with Uncertain Data”.
where structural damages are introduced through reducing stiffness at
certain elements. The physics-guided PhyResNet outperforms the data- References
driven ResNet by being able to accurately locate the true damages and
quantifying the stiffness reductions of them at a decent level, even when [1] Avci O, Abdeljaber O, Kiranyaz S, Hussein M, Gabbouj M, Inman DJ. A review of
vibration-based damage detection in civil structures: From traditional methods to
very small amount of data is used for training. Notably, it achieves a Machine Learning and Deep Learning applications. Mech Syst Signal Process 2021;
31.2 % improvement in R-Value when utilizing only 20 % of the dataset 147:107077.
for training. Both the numerical and experimental studies demonstrate [2] Banan A, Nasiri A, Taheri-Garavand A. Deep learning-based appearance features
extraction for automated carp species identification. Aquac Eng 2020;89:102053.
the superior performance of the proposed PhyResNet especially when
[3] Baydin AG, Pearlmutter BA, Radul AA, Siskind JM. Automatic differentiation in
the amount of available training data is limited. In summary, this paper machine learning: a survey. J Marchine Learn Res 2018;18:1–43.
proposes a novel method to incorporate physics into pattern recognition [4] Eshkevari SS, Takáč M, Pakzad SN, Jahani M. DynNet: Physics-based neural
architecture design for nonlinear structural response modeling and prediction. Eng
for structural damage identification, especially for real-word tasks that
Struct 2021;229:111582.
may suffer from noise, uncertainty and data scarcity. [5] Farrar CR, Doebling SW, Nix DA. Vibration–based structural damage identification.
It is important to acknowledge that currently the validation focuses Philos Trans R Soc Lond Ser A: Math, Phys Eng Sci 2001;359(1778):131–49.
on numerical simulation and laboratory structure to elaborate the [6] Farrar CR, Worden K. An introduction to structural health monitoring. Philosophical
Transactions of the Royal Society of London A: Mathematical. Phys Eng Sci 2007;365
fundamental principles of the proposed PhyResNet. Further application (1851):303–15.
and validation of the proposed approach to more complex real-world [7] Fink O, Wang Q, Svensen M, Dersin P, Lee W-J, Ducoffe M. Potential, challenges
structures will be conducted in the future. To support this, it is recom- and future directions for deep learning in prognostics and health management
applications. Eng Appl Artif Intell 2020;92:103678.
mended to explore diverse data acquisition methods, including [8] Friswell M, Mottershead JE. Finite element model updating in structural dynamics,
enhanced simulation techniques, robust internet-based monitoring sys- Vol. 38. Springer Science & Business Media,; 1995.
tems, and targeted field experiments. Such approaches would enable the [9] Fritzen C-P, Bohle K. Global damage identification of the ’STEELQUAKE’ structure
using modal data. Mech Syst Signal Process 2003;17(1):111–7.
gathering of extensive data under various real-world operational con- [10] Gao Z, Cecati C, Ding SX. A Survey of Fault Diagnosis and Fault-Tolerant
ditions, enriching the validation process. While this research provides Techniques—Part I: Fault Diagnosis With Model-Based and Signal-Based
pivotal insights, its current scope does not fully address the complexities Approaches. IEEE Trans Ind Electron 2015;62(6):3757–67.
[11] Goodfellow I, Bengio Y, Courville A, Bengio Y. Deep learning, Vol. 1. Cambridge:
of large-scale structures. Future research are anticipated to focus on
MIT Press,; 2016.
deploying PhyResNet for more expansive structures with higher- [12] Hao H, Bi K, Chen W, Pham TM, Li J. Towards next generation design of
dimensional parameter spaces. In the current study, the measurement sustainable, durable, multi-hazard resistant, resilient, and smart civil engineering
structures. Eng Struct 2023;277:115477.
points and excitation location are selected arbitrarily. It is reasonable to
[13] Hao H, Xia Y. Vibration-based damage detection of structures by genetic algorithm.
assume that fewer measurement points could be used for successful J Comput Civ Eng 2002;16(3):222–9.
structural damage identification, since physical constraint has now been [14] Hassaballah M, Awad AI. Deep Learning in Computer Vision: Principles and
introduced. Study on the selection of optimal sensor locations and Applications. CRC Press,; 2020.
[15] He K, Zhang X, Ren S, Sun J. Identity mappings in deep residual networks. Eur Conf
minimum number of sensors required could be conducted in future Comput Vis 2016:630–45.
studies. Moreover, while the primary focus here is on structural damage [16] He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. Proc
identification, the physics-informed learning strategy has potential IEEE Conf Comput Vis Pattern Recognit 2016:770–8.
[17] Hernández A, Amigó JM. Differentiable programming and its applications to
applicability across various SHM tasks, such as structural response dynamical systems. ArXiv Prepr ArXiv 2019;1912:08168.
prediction and anomaly detection in sensor signals. [18] Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected
convolutional networks. Proc - 30th IEEE Conf Comput Vis Pattern Recognit, CVPR
2017, 2017-Janua 2017:2261–9. https://doi.org/10.1109/CVPR.2017.243.
CRediT authorship contribution statement [19] Krizhevsky A, Sutskever I, Hinton GE. ImageNet Classification with Deep
Convolutional Neural Networks. In: Pereira F, Burges CJC, Bottou L,
Qilin Li: Writing – review & editing, Methodology, Investigation. Weinberger KQ, editors. Advances in Neural Information Processing Systems 25.
Curran Associates, Inc; 2012. p. 1097–105.
Hong Hao: Writing – review & editing, Supervision, Methodology,
[20] Lai Z, Mylonas C, Nagarajaiah S, Chatzi E. Structural identification with physics-
Investigation, Conceptualization. Senjian An: Writing – review & edit- informed neural ordinary differential equations. J Sound Vib 2021;508:116196.
ing, Supervision, Methodology, Investigation, Funding acquisition, [21] Li J, Law SS, Ding Y. Substructure damage identification based on response
reconstruction in frequency domain and model updating. Eng Struct 2012;41:
Conceptualization. Bradley Ezard: Methodology, Investigation, Formal
270–84.
analysis. Jun Li: Writing – review & editing, Supervision, Methodology, [22] Lin YZ, Nie ZH, Ma HW. Structural Damage Detection with Automatic Feature-
Investigation, Funding acquisition, Data curation, Conceptualization. Extraction through Deep Learning. Comput-Aided Civ Infrastruct Eng 2017;32(12):
Ling Li: Writing – review & editing, Supervision, Methodology, Inves- 1025–46.
[23] Liu W, Lai Z, Bacsa K, Chatzi E. Physics-guided deep Markov models for learning
tigation, Funding acquisition, Conceptualization. Ruhua Wang: Writing nonlinear dynamical systems with uncertainty. Mech Syst Signal Process 2022;178:
– original draft, Visualization, Validation, Methodology, Investigation, 109276.
Formal analysis, Conceptualization. [24] Misra D. Mish: A self regularized non-monotonic neural activation function. ArXiv
Prepr ArXiv:1908 08681 2019;4(2):10–48550.
[25] Newmark NM. A method of computation for structural dynamics. J Eng Mech Div
1959;85(3):67–94.

15
R. Wang et al. Engineering Structures 318 (2024) 118703

[26] Otter DW, Medina JR, Kalita JK. A survey of the usages of deep learning for natural [35] Xia Y, Hao H. Statistical damage identification of structures with frequency
language processing. IEEE Trans Neural Netw Learn Syst 2020;32(2):604–24. changes. J Sound Vib 2003;263(4):853–70.
[27] Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, Killeen T, Lin Z, [36] Xu S, Noh HY. PhyMDAN: Physics-informed knowledge transfer between buildings
Gimelshein N, Antiga L, Desmaison A, Kopf A, Yang E, DeVito Z, Raison M, for seismic damage diagnosis through adversarial learning. Mech Syst Signal
Tejani A, Chilamkurthy S, Steiner B, Fang L, Chintala S. PyTorch: An Imperative Process 2021;151:107374.
Style, High-Performance Deep Learning Library. In: Wallach H, Larochelle H, [37] Xu Y, Pan Y, Wang Y, Deng D, Han Q. Damage identification of single-layer
Beygelzimer A, d\textquotesingle Alché-Buc F, Fox E, Garnett R, editors. Advances cylindrical latticed shells based on the model updating technique. J Civ Struct
in Neural Information Processing Systems, 32. Curran Associates, Inc; 2019. Health Monit 2022;12(2):289–303.
p. 8024–35. 〈http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-hi [38] Yan YJ, Cheng L, Wu ZY, Yam LH. Development in vibration-based structural
gh-performance-deep-learning-library.pdf〉. damage detection technique. Mech Syst Signal Process 2007;21(5):2198–211.
[28] Pathirage CSN, Li J, Li L, Hao H, Liu W, Wang R. Development and application of a [39] Yu Y, Yao H, Liu Y. Structural dynamics simulation using a novel physics-guided
deep learning–based sparse autoencoder framework for structural damage machine learning method. Eng Appl Artif Intell 2020;96:103947.
identification. Struct Health Monit 2019;18(1):103–22. [40] Yuan F-G, Zargar SA, Chen Q, Wang S. Machine learning for structural health
[29] Sherstinsky A. Fundamentals of recurrent neural network (RNN) and long short- monitoring: challenges and opportunities. Sens Smart Struct Technol Civ, Mech,
term memory (LSTM) network. Phys D: Nonlinear Phenom 2020;404:132306. Aerosp Syst 2020;2020(11379):1137903.
[30] Wang R, Chencho An, S, Li J, Li L, Hao H, Liu W. Deep residual network framework [41] Zhang R, Liu Y, Sun H. Physics-guided convolutional neural network (PhyCNN) for
for structural health monitoring. Struct Health Monit 2021;20(4):1443–61. data-driven seismic response modeling. Eng Struct 2020;215:110704.
[31] Wang F, Decker J, Wu X, Essertel G, Rompf T. Backpropagation with callbacks: [42] Zhang Y, Miyamori Y, Mikami S, Saito T. Vibration-based structural state
Foundations for efficient and expressive differentiable programming. Adv Neural identification by a 1-dimensional convolutional neural network. Comput-Aided Civ
Inf Process Syst 2018:31. Infrastruct Eng 2019;34(9):822–39.
[32] Wang R, Li L, Li J. A novel parallel auto-encoder framework for multi-scale data in [43] Zhang Z, Sun C. Structural damage identification via physics-guided machine
civil structural health monitoring. Algorithms 2018;11(8):1–15. learning: a methodology integrating pattern recognition with finite element model
[33] Wang R, Li J, An S, Hao H, Liu W, Li L, others. Densely connected convolutional updating. Struct Health Monit 2021;20(4):1675–88.
networks for vibration based structural damage identification. Eng Struct 2021; [44] Zhang Z, Sun C, Jahangiri V. Structural damage identification of offshore wind
245:112871. turbines: A two-step strategy via FE model updating. Struct Control Health Monit
[34] Wei X, Yang Z, Liu Y, Wei D, Jia L, Li Y. Railway track fastener defect detection 2022;29(2):e2872.
based on image processing and deep learning techniques: A comparative study. Eng
Appl Artif Intell 2019;80:66–81.

16