Model-based Damage Detection through Physics-guided Learning

for Dynamic Systems

Ali I. Ozdagli and Xenofon Koutsoukos

Department of Computer Science, Vanderbilt University, Nashville, TN 37212 USA

Ali.I.Ozdagli@vanderbilt.edu
Xenofon.Koutsoukos@vanderbilt.edu

A BSTRACT methods are proven to be successful in diagnosing the sys-

tem through characterization and localization of the damage
Data-driven learning approaches have gained a lot of inter-
(Bakhary et al., 2007). One of the obstacles for such methods
est in evaluating and validating complex dynamic systems.
is often the availability of sufficient training data (Z. Zhang
One of the main challenges for developing a reliable learn-
& Sun, 2020). More specifically, access to a complete train-
ing model is the lack of training data covering a large range
ing dataset covering a wide range of conditions is costly and
of various operational conditions. Extensive training data can
in some instances impossible without actually damaging the
be generated using a physics-based simulation model. On the
system prior to operation. This problem is a major roadblock
other hand, the learning algorithm should be still tested with
in developing efficient data-driven algorithms for diagnostics
experimental data obtained from the actual system. Modeling
of dynamic systems (Sadoughi & Hu, 2019).
errors may lead to a statistical divergence between the simu-
lation training data and the experimental testing data, causing For cases where training data captured from the field is lim-
poor performance, especially for domain-agnostic black-box ited, a data-driven black-box ML model could be trained with
learning methods. To close the gap between the simulation simulation data. In other words, to compensate for the lack of
and experimental domains, this paper proposes a physics- experimental training data, a representative analytical model
guided learning approach that combines the power of neural can simulate the behavior of the system physics to some de-
network with domain-specific physics knowledge. Specifi- gree. While physics-based analytical models are capable of
cally, the proposed architecture integrates physical parame- generating extensive training dataset, the resulting ML al-
ters extracted from physics-based simulation data into the in- gorithm should still be evaluated with experimental testing
termediate layers of the neural network to constrain the learn- data. Well-established analytical models are capable of sim-
ing process. To demonstrate the effectiveness of the proposed ulating the dynamic response of the target system (Teughels
approach, the architecture is adopted to a damage classifica- & De Roeck, 2005; Jaishi & Ren, 2006). On the other hand,
tion problem for a three-story structure. Our results show calibrating a large set of parameters for complex systems to
that the accuracy for localizing the damage correctly based achieve accurate physical behavior is often computationally
on experimental data improves significantly over black-box exhaustive and at times infeasible (R. Zhang et al., 2020).
models. We also use the physics-based intermediate layers to Eventually, the analytical representation inherits modeling er-
analyze the interpretability of the results. ror. In this case, it is expected that the ML algorithm will
fail to perform efficiently during testing since the simulation
1. I NTRODUCTION training data and experimental testing data are statistically
divergent (Gardner et al., 2020). To address this drawback
In the last decade, the use of machine learning (ML) algo-
of data-driven black-box algorithms, the inference should in-
rithms gained a lot of interest within the community of condi-
corporate domain-specific physical knowledge. The physics-
tion monitoring for dynamic systems (Widodo & Yang, 2007;
guided learning (PGL) which is essentially a hybrid approach
Farrar & Worden, 2012; Stetco et al., 2019). A majority of
aggregating data-driven inference with physical parameters
ML applications in this area exercise a data-driven black-box
has the potential to leverage the performance of the condition
approach that utilizes a large volume of experimental data ob-
monitoring further and to bridge the gap between simulation
tained directly from the actual dynamic system. Black-box
and experimental domains.
Ali I. Ozdagli et al. This is an open-access article distributed under the terms
of the Creative Commons Attribution 3.0 United States License, which per- In recent years, a number of PGL approaches have been pro-
mits unrestricted use, distribution, and reproduction in any medium, provided posed (Karpatne et al., 2017; Jia et al., 2018; Sadoughi & Hu,
the original author and source are credited.

1
 A NNUAL C ONFERENCE OF THE P ROGNOSTICS AND H EALTH M ANAGEMENT S OCIETY 2021

2019; R. Zhang et al., 2020; Z. Zhang & Sun, 2020; Yao et al., has the potential to improve damage detection and localiza-
2020). However, the variety of applications implies that the tion for SHM applications and promises more accurate deci-
implementation of a proper PGL with domain-specific knowl- sions and prioritization for maintenance operations.
edge is non-trivial. Moreover, most of the existing work fo-
In summary, the major contributions of this paper are:
cuses on the prediction of system responses using PGL. In the
area of diagnostics, little research effort has been devoted to • A physics-guided learning architecture, PGL4SHM is
incorporating of physical knowledge into the data-driven ML proposed to generalize damage detection and location
and exploiting deep learning architectures for damage classi- prediction for dynamic systems.
fication. • The proposed architecture uses physics-constrained in-
In this paper, we propose a damage detection and local- termediate variable layers that rely on physical param-
ization architecture for dynamic systems, namely, physics- eters known to be statistically important features for
guided learning for structural health monitoring (PGL4SHM) damage detection such as natural frequencies and mode
that combines the power of neural networks with domain shapes.
specific physics knowledge. In particular, PGL4SHM is a • For proof of concept, the proposed method is eval-
multi-task deep learning architecture which (i) utilizes the uated by comparing damage localization performance
synthetic data simulated by a numerical physics-based rep- to black-box models for numerical and experimental
resentation of the target structure for training and (ii) in- cases.Results show that the new approach improves pre-
corporates domain-specific physical parameters derived from diction accuracy in the presence of modeling error.
this representation into the loss function. The multi-tasking • PGL4SHM improves the explainability of the results
PGL4SHM is trained with simulated structural responses in since the intermediate layers expose valuable informa-
time-series form which serve as the input to the deep network. tion that is highly relevant to the physics of the target
Additionally, during the learning phase, the physical param- structure.
eters such as natural frequencies and mode shapes that are
known to be structural damage indicators (Kim et al., 2003) 2. P ROBLEM F ORMULATION
are used for training the intermediate layers of PGL4SHM
(Muralidhar et al., 2019). The modal features (natural fre- Structural systems can experience damage throughout their
quencies and mode shapes) can be extracted directly from the life-cycle. It is essential to detect and locate the damage earl-
numerical model. Since the organic relationship between the on before it progresses to a bigger failure. In this context,
physics-based model representation, structural responses, and damage localization is a supervised classification problem.
damage state is embedded into the PGL4SHM, the architec- In this study, we consider black-box deep neural networks
ture is capable of generalizing damage detection compared to that label raw input data in the form of acceleration time-
black-box approaches. As PGL4SHM uses modal features series measurements according to the damage condition the
derived from numerical model, this embedding is physics- structure is experiencing. The implicit assumption for the
based rather empirical. black-box model is that the training data is available for ev-
To validate the PGL4SHM architecture, two case studies are ery expected damage condition. In reality, the training data
considered. The first study is purely analytical and inves- is only available for no damage condition. To obtain training
tigates the performance and efficiency of the proposed ap- data for other damage conditions, the structure should be de-
proach under ideal conditions since every aspect of the sys- liberately tarnished which is not practical. We can create a
tem is simulated. The second case study considers the ex- physics-based model of the structure and generate simulated
perimental setup of a small-scale three-story structure tested data for various damage scenarios of interest. Accordingly,
at Los Alamos National Lab. The results from both the an- the black-box model can be trained with the simulated data
alytical and experiential case studies show that the proposed and tested with experimental data after deployment. How-
architecture has better generalization capability in localizing ever, this approach is often not feasible since the physics-
the damage compared to black-box models in the presence based simulation often has intrinsic modeling error. Due to
of modeling errors. The performance gain is more evident the deviation between simulated training and experimental
when the numerical representation deviates from the actual testing data, the black-box model will become ineffective in
structure further. Lastly, we evaluated the explainability of labeling the input correctly.
the results by analyzing the relationship between structural The problem considered in this paper is localizing the damage
responses, damage state, and the integrated physical parame- accurately in the presence of modeling errors. To address this
ters. problem, three challenges should be resolved. First, we need
Overall, PGL4SHM combines data-driven machine learning to create a physics-based representation of the target system
with knowledge of physics. As a result of this, PGL4SHM based on the available data. This model should be used to
generate simulated data. Secondly, we should design a deep

2
 A NNUAL C ONFERENCE OF THE P ROGNOSTICS AND H EALTH M ANAGEMENT S OCIETY 2021

learning architecture that is trained with simulated data but (Kim et al., 2003; S.-Q. Wang & Li, 2012). On the other
can generalize well for experimental data compared to the hand, majority of the aforementioned approaches depends on
black-box model. Thirdly, we need to establish the physi- the existence of full-range experimental data. In this paper,
cal parameters most relevant to the damage condition of the to overcome the limitations of black-box models, we propose
structure to integrated into intermediate layers of the architec- PGL4SHM where the domain-specific knowledge is embed-
ture during inference. Ultimately, successful implementation ded into the deep learning architecture through intermediate
of PGL4SHM should generalize the prediction well for the layers inspired from Muralidhar et al. (2019).
cases where field data is limited and physics-based simula-
Here, we assume that the input is structural response mea-
tion has some modeling error.
surements in the time domain obtained from a physics-based
simulated model, the output is the damage condition associ-
3. P HYSICS - GUIDED L EARNING
ated with the input data. The intermediate layers utilize phys-
For a given set of structural response measurements in the ical variables to improve supervised learning to enable a rich
time domain, we are interested in predicting the damage con- and generalized representation of the target system and to im-
dition of the structure. Such a predictor can be trained with prove supervised learning. The physics-based model is devel-
a supervised learning approach since for each input, x there oped as a representative finite element model (FEM) based on
is a label, y corresponding to the damage state. One way to the available experimental data obtained from the undamaged
learn the mapping from x to y is by training a black-box feed- structure. The simulated training data is generated using this
forward neural network. By utilizing nonlinear activation FEM for various damage conditions of interest, including no
functions within neurons, this network allows us to expose damage case. The modal parameters f and ϕ can be extracted
the complex relationship between the structural responses and directly from the FEM. A simplified layout of the architec-
the damage conditions. While the black-box networks are ture for training is given in Figure 1. In this architecture,
capable of learning the latent feature space, they can fail to
generalize the predictions for unseen observations.
For many dynamic systems, the data labeled as normal is of-
ten available when they are deployed. However, access to
data relevant to damage conditions is limited without harm-
ing the system. The absence of experimental data can be com-
pensated by simulating damage on a finite element model of
the system and obtaining new input/output pairs. However,
due to the poor generalization of black-box models, the pre-
dictor will suffer from the presence of modeling errors and Figure 1. Simplified layout for training PGL4SHM
label the given inputs incorrectly. To address the limitation
of the black-box models, this paper presents the physics- the input later takes the simulated time-series data obtained
guided learning for structural health monitoring (PGL4SHM) from FEM. Each piece of simulated data is associated with a
which integrates the physical knowledge regarding the dy- label designating the damage condition. The feature extrac-
namic characteristics of the target structure into the deep tion layers are a set of layers designed as convolutional neural
learning architecture. networks (CNN). Additionally, there are two individual inter-
mediate variable layers in parallel. The output of each inter-
3.1. Overview of PGL4SHM Architecture mediate layer corresponds to a modal parameter (f and ϕ).
Damage occurring in a load-carrying member changes the dy- While the modal parameters are simply extracted from the
namic characteristics of the structure (Balageas et al., 2010). FEM using eigenvalue problem (Craig Jr & Kurdila, 2006),
Fundamental dynamic characteristics of a system can be de- it can be also derived from time series using domain-specific
scribed in terms of its modal parameters such as natural fre- frequency-domain analysis processes (Ghanem & Shinozuka,
quencies, f , and mode shapes, ϕ. These parameters can be 1995). The intermediate layers are physics-guided and are di-
obtained from time series data using frequency domain anal- rectly associated with physically meaningful modal parame-
ysis techniques (Brincker et al., 2000). Prior research has ters which are known to be good damage indicators. In this
shown that supervised black-box algorithms utilizing modal regard, this architecture exploits the feature extraction as a
parameters in the input layer can predict damage detection modal analysis step to compute intermediate variables which
and localization with success (Z. Wang et al., 1997; Hakim essentially blends domain-specific knowledge with the learn-
& Razak, 2014). A number of literature specifically focused ing process. For this study, we assumed the intermediate vari-
on the use modal parameters such as natural frequencies and able layers are densely connected following a flattening layer
mode shapes to predict damages in a more refined manner after CNN based feature extraction layers. Next, the label
prediction layers are tasked to extract features from modal

3
 A NNUAL C ONFERENCE OF THE P ROGNOSTICS AND H EALTH M ANAGEMENT S OCIETY 2021

properties to determine the damage condition for the given in- learning process as domain-specific knowledge is a promising
put. For training, PGL4SHM requires simulated time-series tool for generalizing damage detection compared to black-
data, associated modal parameters, and labels. During test- box approaches.
ing, the architecture needs only experimental data to the input
layer and predicts the relevant damage condition accordingly. 3.1.2. Learning Process
A black-box architecture is very similar to the PGL in nature This network is typically trained with structural response data
with a main difference. Since black-box does not utilize in- obtained from a representative FEM. Additionally, the ar-
termediate layers and physical parameters associated with it, chitecture utilizes physics-based modal parameters also ob-
these layers are simply not implemented. tained from FEM. Accordingly, the empirical loss function
that needs to be minimized during learning can be formalized
3.1.1. Physics-based Modeling as follows:
A physics-based presentation of the structure can be often Loss = LossDMG + λPGL LossPGL (3)
achieved by modeling the target structure using a finite ele-
ment model. In FEMs, the structural systems are modeled Eq. 3 implies that the network utilizes a multi-task learn-
as a set of discrete elements (known as finite elements) that ing scheme, where LossDMG corresponds to the categorical
are related to the physical properties of the structure such as cross-entropy loss between the actual damage condition, y
stiffness, area of the member section, etc. A finite element and predicted label, ŷ; LossPGL represents mean square er-
model can be idealized as a set of mass (M ), stiffness (K), ror (MSE) for the physics-guided learning parameters; and
and damping matrix (C) which can be written in terms of λP GL is the trade-off parameter. Please note that a black-box
equations-of-motions (EOMs). model will only use LossDMG for the training and disregard
the physics related loss, LossPGL .
M ẍ + C ẋ + Kx = F (1)
The physics-guided loss, LossPGL given in Eq. 3 can be de-
where F is the input excitation such as ground motion, x, scribed as:
ẋ, and x are the acceleration, velocity, and displacement ob-
tained from the system, respectively. A finite element model LossPGL = LossPGL (f, fˆ) + LossPGL (ϕ, ϕ̂) (4)
can be generated either manually by representing each struc-
Here, LossPGL (f, fˆ) is the MSE between the actual nat-
tural element faithful to its physical properties or by extract-
ural frequencies, f and predicted frequencies, fˆ; and
ing the EOM matrices from the experimental data (Fritzen,
LossPGL (ϕ, ϕ̂) is the MSE between the actual mode shapes,
1986; Chen et al., 1996). Complex FEMs involve large di-
mensional matrices which complicate the modeling and make ϕ and predicted ones, ϕ̂. Since LossPGL is a regression loss,
the calibration process cumbersome. As a result, the model- we assume the neurons of the intermediate layers are linearly
ing errors are inevitable but often acceptable for many engi- activated.
neering applications.
4. E VALUATION
Once the matrices are obtained, the training data can be sim-
ulated using Eq 1. By modifying M or K depending on the For this study, we evaluated the PGL4SHM architecture by
damage type, various damage conditions can be simulated. comparing it to the black-box model performance. We have
For every damage type, an eigenvalue problem can be applied considered two case studies. The first case study focuses on
to extract modal parameters as follows: a finite element model of a simply supported beam, and the
second case investigates experimental testing of a three-story
λM ϕ = Kϕ (2) structure.

where λ is the diagonal eigenvalue matrix and can be written 4.1. Implementation
also as λ = diag(2πf 2 ). It should be noted that for every
combination of K and M pair, a unique pair of f and ϕ can The FEM (simply supported beam) and experimental struc-
be generated. ture (three-story structure) are excited with white noise un-
der various damage conditions and the resulting dynamic re-
In addition to eigenvalue analysis, the modal parameters can sponses are collected from all available sensors in terms of ac-
be obtained from structural responses using sophisticated celerations for some amount of time. Then, the accelerations
time and frequency analysis techniques (Ghanem & Shi- are divided into 1-second chunks and each of these chunks
nozuka, 1995). A clear relationship between physics-based are categorized according to the relevant damage state. The
EOM matrices, structural responses, and modal features is data obtained from this process is the reference data and used
obvious as all of them are related to the dynamic characteris- for testing.
tics of the structure. Integration of modal parameters into the

4
 A NNUAL C ONFERENCE OF THE P ROGNOSTICS AND H EALTH M ANAGEMENT S OCIETY 2021

Figure 2. Black-box architecture adopted from Lin, Nie, & Figure 3. PGL4SHM architecture
Ma, 2017

parameters. The intermediate layers are densely connected

In parallel, for each case, another FEM model is developed and the neurons are linearly activated. Both black-box model
to replicate the original structure. This model is intentionally and PGL4SHM architectures are trained using Keras running
misrepresented to some degree in order to introduce model- on TensorFlow 2.0 in Python 3.7. The performance of both
ing errors that occasionally occur during the design process. architectures is evaluated by computing the classification ac-
Using this FEM, the structural responses and corresponding curacy.
damage labels are generated. In addition, the modal param-
eters are extracted from this FEM and vectorized. This data 4.2. Case 1: Analytical Example
is then divided into training, validation, and testing with a
This case focuses on the effectiveness of the proposed model
ratio of 0.6 : 0.2 : 0.2, respectively. The training and vali-
where modeling errors relevant to environmental, opera-
dation data is used during the training phase of PGL4SHM.
tional, and material uncertainties are controlled more pre-
The testing data and the reference data are used for perfor-
cisely. Here, we consider a simply supported beam studied by
mance evaluation. Before training, all available data is stan-
Lin et al. (2017). The beam has a span length of L = 10.0 m
dardized by removing the mean and scaling to unit variance
and a rectangular section with 0.1m width and 0.25m (see
with respect to training data. All FEM and experimental data
Figure 4). The beam is assumed to be made of steel with the
is standardized with scikit-learn toolbox.
elastic modulus of 206 GP a and density of 7, 900 kg/m3 .
Next, two neural network models are trained for each case. The damping is simulated with classic Rayleigh damping
The first neural network is a black-box model that learns end- where mass matrix (M ) proportional factor, α is 1.0 s−1 and
to-end relationship between the time series input and the dam- stiffness matrix (K) proportional factor, β is 1.15 × 10−6 s.
age condition (see Figure 2). The network is structured as The beam is modeled using FEM tool, Open System for
prescribed in Lin et al. (2017). This model does not utilize Earthquake Engineering Simulation - OpenSees (McKenna
physics-guded variables, mode shapes and frequencies at all. et al., 2010). The beam is discretized into ten equally long
The dimension of the input depends on the number of the members that have linear elastic-beam column element prop-
sensors and the sampling number. The feature extractor and erties. Excluding support nodes, the beam has 9 nodes. To
label prediction layers are CNN and DNN, respectively. All generate acceleration responses, the beam is excited at each
neurons have leaky RELU activation functions. The size of of the nine nodes vertically with a random noise. This ex-
the output layer changes with respect to the number of dam- citation has a Gaussian distribution with a mean of 200 N
age conditions considered for the case study. The neurons of and standard deviation. To simulate finite features of the en-
this layer are activated with softmax. To generalize the pre- vironmental noise, the random excitation is filtered with an
dictions and mitigate the internal covariate shift, batch nor- eighth-order Butterworth filter that has a cutoff of 512 Hz.
malization layers are also inserted to the black-box model. The sampling rate for the simulation is selected 8192 Hz. To
Lastly, to reduce the number of trainable parameters, every reduce the volume of the data, the simulation data is down-
batch normalization layer is followed by a max pool layer. sampled to 1024 Hz and only vertical accelerations at nine
The second model, PGL4SHM architecture is trained with nodes are considered. For each loading case, the size of one
the training data to minimize the loss function given in Eq. 3 simulation instance is (9 nodes × 1024 Hz). To simulate
(see Figure 3). This network structure follows an arrangement damage conditions, each of the ten members is damaged in-
similar to the black-box model with the addition of interme- dividually by reducing the member stiffness by 10% to 50%
diate value layers which employs the physics-guided modal with 10% increments. Including no damage state, 11 damage

5
 A NNUAL C ONFERENCE OF THE P ROGNOSTICS AND H EALTH M ANAGEMENT S OCIETY 2021

conditions are simulated at nine loading positions across ten cent. For the cases ME 10%, ME 15%, and ME 20%, the per-
members. In addition to time-series data, for each data sim- formance of both PGL4SHM architectures succeeds black-
ulation, first 5 natural frequencies and 27 three modes shape box significantly. Overall, the improvement of prediction ac-
points (3 modes × 9 nodes) are extracted from the FEM an- curacy increases progressively with the modeling error.
alytically using OpenSees. The resulting data is categorized
In general, the black-box model is a good choice when the
according to the damage location independent of the magni-
modeling fidelity is ensured. Both black-box and extended
tude of the damage and the location of the excitation. All the
PGL4SHM have about the same amount of trainable parame-
data generated so far constitutes the reference data for testing.
ters, and their prediction accuracies are similar. The extended
In parallel, another set of simulation data are generated with
PGL4SHM is successful for almost every case except the case
ME 5%, however, compared to black-box, the performance
loss is negligible. When the modeling error is small, com-
pared to the extended PGL4SHM, the regular PGL4SHM is,
in general, less effective due to the small number of trainable
parameters. The results clearly show that especially when the
numerical model does not represent the actual system prop-
Figure 4. Simply supported beam model used for analytical erly, blending physical parameters with data-driven machine
case learning has a positive impact in improving the damage local-
ization.
an inaccurate FEM model. To account for environmental, op-
erational, and material uncertainties, for each data instance,
the stiffness of the inaccurate model is perturbed with a log-
normal distribution. Four inaccurate models are developed
where the maximum error of all sampled elastic modulus
varies between 5% and 20% with 5% increments. This data
is used for training, validation, and testing of PGL4SHM. In
addition to time-series data, for each data simulation, the first
5 natural frequencies and 3 modes shapes are extracted from
the FEM analytically using OpenSees.
Following the deep-learning architecture provided by Lin et
al. (2017), the black-box model and PGL4SHM counterpart
are developed, yielding about 1,072,267 and 621,739 train-
able parameters to optimize, respectively. The PGL4SHM Figure 5. Visualization of classification accuracy for analyti-
has intermediate layers between feature extraction and label cal case
prediction layers. These intermediate layers act as a choke
point, decreasing the number of trainable parameters. To
make up for the capacity of the PGL4SHM, two more con- 4.3. Case 2: Experimental Example
volutional layers (a regular convolutional layer and one with
The performance of the PGL4SHM approach is also eval-
batch normalization and max pooling) are added before the
uated on a small-scale three-story structure tested by
flattening. This model, namely PGL4SHM - Extended, has
Figueiredo et al. (2009). An electromagnetic shaker is
1,097,387 trainable parameters.
attached to the base of the structure (see Figure 6). The
Two versions of PGL4SHM (regular and extended) are com- structure was excited with a band-limited white noise and the
pared to the black-box architecture. Table 1 summarizes the resulting horizontal acceleration responses and the excitation
classification accuracy and the improvement over black-box force were measured at a sampling rate of 320 Hz for about
architecture with respect to the maximum modeling error in 25 s. For this study, including undamaged state, four damage
percentage. In addition, Figure 5 visualized the accuracy of conditions are considered. The damage states are established
all architectures. For no modeling error (ME 0%), while by reducing the stiffness of one or two columns at each floor
black-box outperforms the regular PGL4SHM, the perfor- by 87.5 percent. Each response data instance is categorized
mance of extended PGL4SHM surpasses all of them. When according to its respective damage condition. After the input
there is a small modeling error (ME 5%), black-box is the best force is removed from the measurements, time series data
among the three, resulting in to 94 percent accuracy. On the are divided into 1-second chunks. Each chunk is catego-
other hand, the difference between black-box and extended rized according to its respective damage condition. The data
PGL4SHM (84.98 vs 84.55 percent) is negligible. The power collected in this phase is the reference data for testing. In
of PGL4SHM shines when the modeling error is above 5 per- addition, a high-fidelity lumped-mass model is generated

6
 A NNUAL C ONFERENCE OF THE P ROGNOSTICS AND H EALTH M ANAGEMENT S OCIETY 2021

Black-box PGL4SHM PGL4SHM - Extended

Modeling Error Accuracy (%) Accuracy (%) Improvement (%) Accuracy (%) Improvement (%)
ME 0% 93.75 91.90 -1.97 93.89 0.15
ME 5% 84.98 83.88 -1.29 84.55 -0.50
ME 10% 67.71 70.78 4.55 70.52 4.15
ME 15% 52.98 56.67 6.96 56.29 6.24
ME 20% 42.75 47.18 10.35 48.60 13.68
Table 1. Classification accuracy of black-box and PGL4SHM for analytical case

Black-box PGL4SHM
Modeling Error Acc. (%) Acc. (%) Impr. (%)
ME 0% 96.18 90.74 -5.66
ME 10% 38.06 70.82 86.07

Table 2. Classification accuracy of black-box and PGL4SHM

for experimental case

Additionally, Table 2 presents the averaged F1-scores for the

experimental case. The results and improvements are in par-
allel with the classification accuracies.
Black-box PGL4SHM
Modeling Error F1 (%) F1 (%) Impr. (%)
ME 0% 96.21 90.48 -5.96
ME 10% 31.09 67.36 116.66
Table 3. Averaged F1-score of black-box and PGL4SHM for
experimental case

Figure 6. Three-story structure used for experimental case

Lastly, Fig. 7 illustrates averaged ROC for the experimen-
tal case. Blackbox and PGL4SHM have similar ROC per-
in the form of mass-stiffness-damping matrices using the formance when there is no modeling error. Under moderate
parameters provided by Hernandez-Garcia et al. (2010) and level of modeling error, PGL4SHM has a better classification
Sun & Betti (2015). A 10 percent error is introduced into performance compared to the blackbox model.
the stiffness matrix to simulate the modeling uncertainties.
Using this imperfect model, data for all damage conditions
are produced. In addition, three natural frequencies and 9
mode shapes points (3 modes × 3 stories) are obtained using
this model. The data from the imperfect FEM model is used
for training and validation of PGL4SHM. Black-box and
PGL4SHM architectures have 707,844 and 557,459 train-
able parameters, respectively. No further layers are added
to PGL4SHM to extend the capacity of PGL4SHM. We
also ensured there is no overfitting by validating the models
against their respective numerical datasets. The performance
of trained architectures and the improvement of PGL4SHM
over black-box architecture for the corresponding modeling
error are provided in Table 2. The classification performance
Figure 7. Averaged ROC curves for experimental case
of the black-box for no modeling error (96.18%) is greater
than that of PGL4SHM (90.74%) and the performance loss
reaches up to 6 percent. For a moderate level of model- From the results, it is evident that the black-box overfits the
ing error (ME 10%), the black-box model yields a poor numerical data such that the latent features of the experimen-
performance (38.06%) compared to PGL4SHM (70.82%). tal data cannot be perceived. As a result, without the inte-

7
 A NNUAL C ONFERENCE OF THE P ROGNOSTICS AND H EALTH M ANAGEMENT S OCIETY 2021

gration of the physical parameters, the data-driven black-box

architecture fails to predict the damage classes correctly. The
results indicate that in the presence of modeling error, the
generalization of PGL4SHM is much more successful.

4.4. Effect of Hyper-parameters

Here, we investigated the effect of the trade-off parameter,
λP GL on the prediction accuracy to understand the general-
ization of PGL4SHM under no modeling error. Table 4 sum-
marizes the performance of both models and the weights for
classification loss and physics-based loss for no modeling er-
ror. Trade-off parameter, in general, does not affect the per-
formance of PGL4SHM, except for the case (λP GL = 2.0)
where the weight for physical parameter loss is larger than the
one for the classification loss. PGL4SHM with no weights Figure 8. Interpretability of intermediate layer outputs
to the physical parameters (λP GL = 0.0) is similar to the
black-box model, but it still contains the intermediate layers.
It is clear from the results that the introduction of interme- by reducing the stiffness matrix by 10 percent which causes
diate layers degrades the performance of PGL4SHM when the simulated frequencies to decrease. In general, the pre-
there is no modeling error. The small dimension of inter- dicted frequencies range from simulated to experimental val-
mediate layers after the label prediction layer (see Figure 1) ues.
causes the learning to be under-complete leading to decrease During training, the simulated modal parameters are used for
in accuracy. For larger models, the number of physical pa- physics-based loss function. On the other hand, integration
rameters can be increased and the label prediction layer will of physics-based parameters into the training also constrains
have a more complete basis for learning. For general pur- the inference such that PGL4SHM favors to predict the modal
poses, weighting the losses equally (λP GL = 1.0) is a good parameters towards the experimental true counterparts. There
starting point in training the PGL4SHM. are some cases where the predicted values do not distribute
uniformly between experimental and simulated values. The
λP GL Accuracy (%)
distortion is substantial especially for the second modes ( 50
0.0 86.26 Hz) of damage class 1 and 2. This error causes some of the
0.5 89.24 intermediate value outputs from class 1 and 2 to overlap with
1.0 87.48
2.0 68.32 the damage class 0 (no damage class) leading to mislabel-
ing. Due to the explainability of results, such problematic
Table 4. Effect of hyper-parameters on the classification ac- instances can be in theory captured algorithmically and cor-
curacy under no modeling error (ME 0%) rected at testing time.

5. C ONCLUSION
4.5. Interpretability of Intermediate Layer Outputs
In this paper, we have presented a physics-based deep learn-
We evaluated the explainability of the PGL4SHM by analyz- ing architecture, PGL4SHM to detect and localize the dam-
ing the relationship between the damage condition and in- age in mechanical systems. The proposed approach incor-
termediate layer outputs. Specifically, we focused on the porates physical parameters such as natural frequencies and
interpretability of natural frequency, as it is more human- mode shapes, which are known to be statistically meaningful
comprehensible and easier to visualize. Figure 8 illustrates features for damage detection, into the intermediate layers of
predicted natural frequencies from intermediate layers, along deep neural networks. To accommodate the intermediate lay-
with the experimental (true) and simulated (training) coun- ers, the architecture introduced physics-based loss into em-
terparts for four damage cases, where the modeling error is pirical loss function. To evaluate the proposed approach, we
%10. Here, PGL4SHM is evaluated with experimental data. considered analytical and experimental cases. Both examples
For each damage case, the intermediate layers in PGL4SHM show that physics-guided learning improves the accuracy of
predict three natural frequencies around 30, 55, and 70 Hz the damage localization compared to black-box models in the
with some variance. Compared to the experimental true fre- presence of modeling errors. Our empirical study shows that
quencies of the structure (square markers), simulated values weighting the classification and physical loss equally is an
extracted from FEM (cross markers) always undershoot. This effective starting point for training. Lastly, we discussed the
is expected since the modeling error is introduced to the FEM interpretability of intermediate layer output by analyzing the

8
 A NNUAL C ONFERENCE OF THE P ROGNOSTICS AND H EALTH M ANAGEMENT S OCIETY 2021

relationship between predicted modal parameters and classifi- networks-a review. Smart Structures and Systems, 14(2),
cation performance. Our findings indicate that the misclassi- 159–189.
fied instances could be explained through the characterization Hernandez-Garcia, M. R., Masri, S. F., Ghanem, R.,
of predicted natural frequencies. Figueiredo, E., & Farrar, C. R. (2010). An experimen-
tal investigation of change detection in uncertain chain-like
ACKNOWLEDGMENT systems. Journal of Sound and Vibration, 329(12), 2395–
This work is supported in part by DARPA through contract 2409.
number FA8750-20-C-0537. Any opinions, findings, and Jaishi, B., & Ren, W.-X. (2006). Damage detection by finite
conclusions or recommendations expressed are those of the element model updating using modal flexibility residual.
authors and do not necessarily reflect the views of the spon- Journal of sound and vibration, 290(1-2), 369–387.
sor.
Jia, X., Karpatne, A., Willard, J., Steinbach, M., Read, J.,
Hanson, P. C., . . . Kumar, V. (2018). Physics guided re-
R EFERENCES current neural networks for modeling dynamical systems:
Application to monitoring water temperature and quality in
Bakhary, N., Hao, H., & Deeks, A. J. (2007). Damage de- lakes. arXiv preprint arXiv:1810.02880.
tection using artificial neural network with consideration of
uncertainties. Engineering Structures, 29(11), 2806–2815. Karpatne, A., Watkins, W., Read, J., & Kumar, V. (2017).
Physics-guided neural networks (pgnn): An applica-
Balageas, D., Fritzen, C.-P., & Güemes, A. (2010). Structural
tion in lake temperature modeling. arXiv preprint
health monitoring (Vol. 90). John Wiley & Sons.
arXiv:1710.11431.
Brincker, R., Zhang, L., & Andersen, P. (2000). Modal iden-
tification from ambient responses using frequency domain Kim, J.-T., Ryu, Y.-S., Cho, H.-M., & Stubbs, N. (2003).
decomposition. In Proc. of the 18*‘international modal Damage identification in beam-type structures: frequency-
analysis conference (imac), san antonio, texas. based method vs mode-shape-based method. Engineering
structures, 25(1), 57–67.
Chen, S., Ju, M.-S., & Tsuei, Y. (1996, 01). Estimation of
mass, stiffness and damping matrices from frequency re- Lin, Y.-z., Nie, Z.-h., & Ma, H.-w. (2017). Structural damage
sponse functions. Journal of Vibration, Acoustics, Stress, detection with automatic feature-extraction through deep
and Reliability in Design, 118(1), 78-82. Retrieved from learning. Computer-Aided Civil and Infrastructure Engi-
https://doi .org/10 .1115/1 .2889638 doi: neering, 32(12), 1025–1046.
10.1115/1.2889638 McKenna, F., Scott, M. H., & Fenves, G. L. (2010).
Craig Jr, R. R., & Kurdila, A. J. (2006). Fundamentals of Nonlinear finite-element analysis software architecture
structural dynamics. John Wiley & Sons. using object composition. Journal of Computing
Farrar, C. R., & Worden, K. (2012). Structural health mon- in Civil Engineering, 24(1), 95-107. Retrieved
itoring: a machine learning perspective. John Wiley & from https :// ascelibrary .org / doi / abs /
Sons. 10.1061/%28ASCE%29CP.1943-5487.0000002
doi: 10.1061/(ASCE)CP.1943-5487.0000002
Figueiredo, E., Park, G., Figueiras, J., Farrar, C., & Worden,
K. (2009). Structural health monitoring algorithm com- Muralidhar, N., Bu, J., Cao, Z., He, L., Ramakrishnan, N.,
parisons using standard data sets. Los Alamos National Tafti, D., & Karpatne, A. (2019). Physics-guided design
Laboratory, Los Alamos, NM, Report No. LA-14393. and learning of neural networks for predicting drag force
on particle suspensions in moving fluids. arXiv preprint
Fritzen, C.-P. (1986, 01). Identification of Mass, Damping,
arXiv:1911.04240.
and Stiffness Matrices of Mechanical Systems. Journal
of Vibration, Acoustics, Stress, and Reliability in Design, Sadoughi, M., & Hu, C. (2019). Physics-based convolutional
108(1), 9-16. Retrieved from https://doi.org/10 neural network for fault diagnosis of rolling element bear-
.1115/1.3269310 doi: 10.1115/1.3269310 ings. IEEE Sensors Journal, 19(11), 4181–4192.
Gardner, P., Liu, X., & Worden, K. (2020). On the applica- Stetco, A., Dinmohammadi, F., Zhao, X., Robu, V., Flynn,
tion of domain adaptation in structural health monitoring. D., Barnes, M., . . . Nenadic, G. (2019). Machine learning
Mechanical Systems and Signal Processing, 138, 106550. methods for wind turbine condition monitoring: A review.
Ghanem, R., & Shinozuka, M. (1995). Structural-system Renewable energy, 133, 620–635.
identification. i: Theory. Journal of Engineering Mechan- Sun, H., & Betti, R. (2015). A hybrid optimization algo-
ics, 121(2), 255–264. rithm with bayesian inference for probabilistic model up-
Hakim, S., & Razak, H. A. (2014). Modal parameters dating. Computer-Aided Civil and Infrastructure Engineer-
based structural damage detection using artificial neural ing, 30(8), 602–619.

9
 A NNUAL C ONFERENCE OF THE P ROGNOSTICS AND H EALTH M ANAGEMENT S OCIETY 2021

Teughels, A., & De Roeck, G. (2005). Damage detection and ology integrating pattern recognition with finite ele-
parameter identification by finite element model updating. ment model updating. Structural Health Monitoring,
Revue européenne de génie civil, 9(1-2), 109–158. 1475921720927488.
Wang, S.-Q., & Li, H.-J. (2012). Assessment of structural
damage using natural frequency changes. Acta Mechanica
Sinica, 28(1), 118–127.
Wang, Z., Lin, R., & Lim, M. (1997). Structural damage
detection using measured frf data. Computer methods in
applied mechanics and engineering, 147(1-2), 187–197. B IOGRAPHIES
Widodo, A., & Yang, B.-S. (2007). Support vector ma- Ali I. Ozdagli is a graduate student in the Department of
chine in machine condition monitoring and fault diagnosis. Electrical Engineering and Computer Science at Vanderbilt
Mechanical systems and signal processing, 21(6), 2560– University. Prior to this, he got his M.S. in civil engineer-
ing from the University of Notre Dame and his Ph.D. from
2574. Purdue University. His research focuses on adapting machine
Yao, H., Gao, Y., & Liu, Y. (2020). Fea-net: A physics- learning approaches to structural health monitoring applica-
guided data-driven model for efficient mechanical response tions.
prediction. Computer Methods in Applied Mechanics and Xenofon Koutsoukos is a professor with the Department of
Engineering, 363, 112892. Electrical Engineering and Computer Science and a senior
research scientist with the Institute for Software Integrated
Zhang, R., Liu, Y., & Sun, H. (2020). Physics-guided con- Systems, Vanderbilt University. His research work is in the
volutional neural network (phycnn) for data-driven seismic area of cyber-physical systems with an emphasis on security
response modeling. Engineering Structures, 215, 110704. and resilience, control, diagnosis and fault tolerance, formal
methods, and adaptive resource management. He received the
Zhang, Z., & Sun, C. (2020). Structural damage identi- Ph.D. degree in electrical engineering from the University of
fication via physics-guided machine learning: a method- Notre Dame. He is a Fellow of the IEEE.

10