A digital twin framework for civil engineering structures

Matteo Torzoni∗1 , Marco Tezzele†2 , Stefano Mariani‡1 , Andrea Manzoni§3 , and

Karen E. Willcox¶2
1
Dipartimento di Ingegneria Civile e Ambientale, Politecnico di Milano, Piazza L.
da Vinci 32, 20133 Milan, Italy
2
Oden Institute for Computational Engineering and Sciences, University of Texas
at Austin, Austin, 78712, TX, United States
3
MOX, Dipartimento di Matematica, Politecnico di Milano, Piazza L. da Vinci 32,
arXiv:2308.01445v2 [math.NA] 31 Oct 2023

20133 Milan, Italy

November 1, 2023

Abstract
The digital twin concept represents an appealing opportunity to advance condition-based
and predictive maintenance paradigms for civil engineering systems, thus allowing reduced
lifecycle costs, increased system safety, and increased system availability. This work proposes
a predictive digital twin approach to the health monitoring, maintenance, and management
planning of civil engineering structures. The asset-twin coupled dynamical system is encoded
employing a probabilistic graphical model, which allows all relevant sources of uncertainty to be
taken into account. In particular, the time-repeating observations-to-decisions flow is modeled
using a dynamic Bayesian network. Real-time structural health diagnostics are provided by
assimilating sensed data with deep learning models. The digital twin state is continually
updated in a sequential Bayesian inference fashion. This is then exploited to inform the
optimal planning of maintenance and management actions within a dynamic decision-making
framework. A preliminary offline phase involves the population of training datasets through a
reduced-order numerical model and the computation of a health-dependent control policy. The
strategy is assessed on two synthetic case studies, involving a cantilever beam and a railway
bridge, demonstrating the dynamic decision-making capabilities of health-aware digital twins.

Contents
1 Introduction 2

2 Predictive digital twins using physics-based models and machine learning 3

2.1 Probabilistic graphical model for predictive digital twins . . . . . . . . . . . . . . . 4
2.2 Numerical models for simulation-based damage identification . . . . . . . . . . . . 7
2.3 Data assimilation via artificial neural networks . . . . . . . . . . . . . . . . . . . . 8

3 Numerical experiments 10
3.1 L-shaped cantilever beam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.1 Dataset assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.2 Digital twin framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.3 Results: two available actions . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.4 Results: four available actions . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Railway bridge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
∗ [email protected]
† [email protected]
‡ [email protected]
§ [email protected]
¶ [email protected]

1
 3.2.1 Dataset assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.2 Digital twin framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

4 Conclusions 20

A Implementation details 26

1 Introduction
The optimal management of deteriorating structural systems is an important challenge in modern
engineering. In particular, the failure or non-optimized maintenance planning of civil structures
may entail high safety, economic, and social costs. Within this context, enabling a digital twin
(DT) perspective for structural systems that are critical for either safety or operative reasons, is
crucial to allow for condition-based or predictive maintenance practices, in place of customarily
employed time-based ones. Indeed, having an up-to-date digital replica of the physical asset of
interest can yield several benefits spanning its entire lifecycle, including performance and health
monitoring, as well as maintenance, inspection, and management planning [1].
The DT concept [2, 3, 4, 5, 6] has been recently applied to several fields for operational moni-
toring, control, and decision support, including structural health monitoring (SHM) and predictive
maintenance [7, 8], additive manufacturing [9], smart cities [10], urban sustainability [11], and
railway systems management [12]. It allows for a personalized characterization of a physical asset,
in the form of computational models and parameters of interest, that evolves over time and is
kept synchronized with its physical counterpart by means of data-collecting devices. Within a civil
SHM framework, such a twinning perspective can be enabled by the assimilation of data through
data-driven structural health diagnostics (from physical to digital), possibly accommodating the
quantification and propagation of relevant uncertainties related to, e.g., measurement noise, mod-
eling assumptions, environmental and operational variabilities [13, 14, 15, 16, 17]. The resulting
updated digital state should then enable prediction of the physical system evolution, as well as
inform optimal planning of maintenance and management actions (from digital to physical).
In this work, we propose a DT framework for civil engineering structures. The overall computa-
tional strategy is based upon a probabilistic graphical model (PGM) inspired by the foundational
model proposed in [18], which provides a general framework to carry out data assimilation, state
estimation, prediction, planning, and learning. Formally, such a PGM is a dynamic Bayesian
network with the addition of decision nodes, i.e., a dynamic decision network [19, 20]. This is
employed to encode the end-to-end information flow, from physical to digital through assimilation
and inference, and back to the physical asset in the form of informed control actions. A graphical
abstraction of the proposed DT strategy is depicted in Fig. 1. The figure shows a physical-to-digital
information flow and a digital-to-physical information flow. These bi-directional information flows
repeat indefinitely over time. In particular, we have:

• From physical to digital. Structural response data are gathered from the physical system
and assimilated with deep learning (DL) models, see e.g., [21, 22], to estimate the current
structural health in terms of presence, location, and severity of structural damage. To solve
this inverse problem, we refer to vibration-based SHM techniques, see e.g., [23, 24, 25, 26],
which exploit the aforementioned collected data, such as displacement or acceleration time
histories. This first estimate of the digital state is then employed to estimate an updated
digital state, according to control-dependent transition dynamics models describing how the
structural health is expected to evolve.
• From digital to physical. The updated digital state is exploited to predict the future evolution
of the physical system and the associated uncertainty, thereby enabling predictive decision-
making about maintenance and management actions feeding back to the physical system.
• Offline learning phase. The DT setup considered in this work takes advantage of a prelimi-
nary offline learning phase. This phase involves the training of the DL models underlying the
structural health identification, and learning the control policy to be applied at each time
step of the online phase. The DL models are trained in a supervised fashion, with labeled

2
 data pertaining to specific damage conditions generated by exploiting physics-based numeri-
cal models. To efficiently assemble a training dataset representative of potential damage and
operational conditions the structure might undergo during its lifetime, we exploit a reduced-
order modeling strategy for parametrized systems relying on the reduced basis method [27].
The health-dependent control policy is also computed offline, by maximizing the expected
future rewards for the planning problem induced by the PGM.

Digital to physical Physical to digital

Optimal policy Sensing

enactment

Probabilistic graphical model for predictive digital twins

Sensed structural response

Digital state tracking

Structural health
identification

Digital state inference and evolution prediction

Figure 1: Predictive digital twin framework for civil engineering structures: graphical abstraction
of the end-to-end information flow enabled by the probabilistic graphical model.

The elements of novelty that characterize this work are the following: (i) the adaptation of
the PGM digital twinning framework to the health monitoring, maintenance, and management
planning of civil engineering structures; (ii) the assimilation of vibration response data is car-
ried out by exploiting DL models, which allow automated selection and extraction of optimized
damage-sensitive features and real-time assessment of the structural state. This work shows how to
incorporate in the DT framework high-dimensional multivariate time series describing the sensor
measurements, while tracking the associated uncertainties. The proposed computational frame-
work is made available in the public repository digital-twin-SHM [28]. The code implements
the PGM framework as a dynamic decision network. This enables us to easily specify the graph
topology from a few time slices, and then unroll it for any number of time steps in the future.
The remainder of this paper is organized as follows. In Sec. 2, we describe the proposed DT
framework. In Sec. 3, the computational procedure is assessed on two test cases, respectively
related to a cantilever beam and a railway bridge. Conclusions and future developments are drawn
in Sec. 4.

2 Predictive digital twins using physics-based models and

machine learning
In this section, we describe the methodology characterizing our DT framework in terms of the
PGM encoding the asset-twin coupled dynamical system in Sec. 2.1; the population of training

3
datasets exploiting physics-based numerical models in Sec. 2.2; and the DL models underlying the
structural health identification in Sec. 2.3.

2.1 Probabilistic graphical model for predictive digital twins

The digital twin assimilates vibration recordings shaped as multivariate time series U(µ) =
[u1 (µ), . . . , uNu (µ)] ∈ RL×Nu , consisting of Nu time series made of L sensor measurements equally
spaced in time, for instance in terms of accelerations or displacements. The vector µ ∈ RNpar com-
prises the parameters representing the operational, damage, and (possibly) environmental condi-
tions. Each recording refers to a time interval (0, T ), within which measurements are recorded with
a sampling rate fs . For the problem settings we consider, the interval (0, T ) is short enough for
the operational, environmental, and damage conditions to be considered time-invariant, yet long
enough not to compromise the identification of the structural behavior.
The PGM that defines the elements comprising the asset-twin coupled dynamical system, and
mathematically describes the relevant interactions via observed data and control inputs, is the
dynamic decision network sketched in Fig. 2. Circle nodes in the graph denote random variables
at discrete times, square nodes denote actions, and diamond nodes denote the objective function.
Bold outlines denote observed quantities, while thin outlines denote estimated quantities. The
directed acyclic structure of the PGM encodes the assumed conditional dependencies. Edges in
the graph represent dependencies between random variables. Solid edges represent the variables’
dependencies encoded via conditional probability distributions, while dashed edges represent the
variables’ dependencies encoded via deterministic functions.
We consider a non-dimensional time discretization, and denote discrete time steps by t. The
physical time duration between successive time steps may vary depending on the application, and
is governed by the update frequency of the DT via data assimilation, so that the DT is updated
once per time step. Thanks to the modeled conditional dependencies between random variables,
the graph topology is specified from the first two time steps, and can then be unrolled for any
number of time steps.

S0 S1 S2

O0 O1

A
U−1 D0NN U0 U0A D1NN U1 U1A

D0 D1

Q0 R0 Q1 R1

t=0 t=1 t=2

| | |

Figure 2: Dynamic decision network encoding the asset-twin coupled dynamical system. Circle
nodes denote random variables, square nodes denote actions, and diamond nodes denote the ob-
jective function. Bold outlines denote observed quantities, while thin outlines denote estimated
quantities. Directed solid edges represent the variables’ dependencies encoded via conditional prob-
ability distributions, while directed dashed edges represent the variables’ dependencies encoded via
deterministic functions.

The physical state St ∼ p(st ), with st denoting the realization of the random variable St at
time t, encapsulates the variability in the health state of the asset, which is usually only partially
observable. The probability distribution encoding the relative likelihood that St = st , for any
possible st , is denoted with p(st ). The digital state Dt ∼ p(dt ) is characterized by those parameters
employed to capture the variability of the physical asset by means of the computational models

4
comprising the DT. In our framework, the digital state is given as a vector of length two, describing
the presence/location and magnitude of damage in the asset. The physical-to-digital information
flow is governed by the observed data Ot = ot , which are assimilated by the DT to update the
digital state. The assimilation is carried out using the DL models described in Sec. 2.3, providing
a first estimate of the digital state DtNN ∼ p(dNN t ). This estimation is then used in a Bayesian
inference formulation, together with the prior belief Dt−1 from the previous time step, to estimate
an updated digital state Dt according to a control-dependent transition dynamics model describing
how the digital state is expected to evolve. The updated digital state can thus be exploited to
compute quantities of interest Qt ∼ p(qt ), such as modal quantities or other response features,
through the computational models comprising the DT. For instance, quantities of interest can
be useful to perform posterior predictive checks on the tracking capabilities of the DT to assess
how it matches the reality, by comparing sensor measurements with the corresponding posterior
estimates. However, we point out that this capability is not exploited in the present work, and
that the Qt node is kept in the graph in agreement with the foundational model proposed in [18].
Nevertheless, the updated digital state Dt is eventually exploited to inform the digital-to-physical
information flow, in the form of control inputs; in Fig. 2, Ut ∼ p(ut ) and UtA = uA t denote the
belief about what action to take and the control input effectively enacted on the asset, respectively.
At each time step, Ut is estimated according to a health-dependent control policy, that maps the
belief over the digital state onto the control actions feeding back to the physical asset. Finally, the
reward Rt ∼ p(rt ) quantifies the performance of the asset for the time step and can be equivalently
perceived as a negative cost to be maximized.
The key assumptions behind our PGM are that the physical state is only observable indirectly
via the sensed structural response, and the physical and digital states evolve according to a Marko-
vian process. This implies that the conditional probabilities associated with the random variables
at one time step depend only on the random variables at the previous time step, and are indepen-
dent of all past states. The resulting graph topology encodes a conditional independence structure
that allows us to conveniently cast the asset tracking within a sequential Bayesian inference frame-
work. Indeed, by exploiting conditional independence and Bayes rule, the joint distributions over
variables can be factorized up to the current time step tc , as follows:
p(D0NN , . . . , DtNN
c
, D0 , . . . , Dtc , Q0 , . . . , Qtc , R0 , . . . , Rtc , U0 , . . . , Utc |o0 , . . . , otc , uA A
0 , . . . , utc )
tc h
Y i
∝ ϕdata
t ϕhistory
t ϕNN QoI control reward
t ϕt ϕ t ϕt , (1)
t=0

with factors:
ϕdata
t = p(Ot = ot |DtNN ), (2)
ϕhistory
t
A
= p(Dt |Dt−1 , Ut−1 = uA
t−1 ), (3)
ϕNN
t = p(Dt |DtNN ), (4)
ϕQoI
t = p(Qt |Dt ), (5)
ϕcontrol
t = p(Ut |Dt ), (6)
ϕreward
t = p(Rt |Dt , UtA = uA
t ). (7)

Herein, ϕdata
t encodes the assimilation of observed data through the DL models underlying the
identification of the structural health. ϕhistory
t and ϕNN
t factorize the belief about the digital state
Dt , conditioned on the digital state at the previous time step Dt−1 , the last enacted control input
A
Ut−1 = uA NN
t−1 , and the data assimilation outcome Dt . In our PGM, the spaces of the digital states
and control inputs are discrete, thus the relevant causal relationships are modeled by means of
conditional probability tables (CPTs). In particular, ϕhistory
t plays the role of a predictor forward
in time, parametrized by means of a control-dependent CPT describing how the digital state is
expected to evolve. Such a CPT should embody any a priori knowledge that the DT designer has
with respect to the asset and the relevant operational conditions. ϕhistory
t can be estimated offline
from historical data, see e.g., [29, 30], or learned online from experience. On the other hand, ϕNN t
updates the digital state estimate to account for data assimilation. This is encoded by means of a
CPT mapping the estimate DtNN provided by the DL models, onto a belief about Dt . Such a CPT
is a confusion matrix measuring the offline (expected) performance of the DL models in correctly
identifying the digital state among all the possible outcomes of Dt . ϕQoI t and ϕreward
t respectively

5
encapsulate the evaluation of the computational models comprising the DT to estimate quantities
of interest, and the computation of the reward function quantifying the performance of the asset.
Finally, the control factor ϕcontrol
t is encoded by means of a health-dependent control policy π(Dt )
computed as described in the following. Since the spaces of the unobserved variables are discrete,
we can propagate and update the relative belief exactly with a single pass of the sum-product
message-passing algorithm [19].
The control policy π(Dt ) is computed offline under the simplifying assumption of sufficient
sensing capability to provide an accurate estimate of the structural health, allowing us to decouple
the sensing and control problems. This involves solving the planning problem induced by the
expected evolution of the structural health, maximizing the expected reward over the planning
horizon. Considering an infinite planning horizon, this can be stated as the optimization problem:
+∞
X
π(Dt ) = arg max γ t E[Rt ], (8)
π
t=0

where γ ∈ [0, 1] is the discount factor. Here, this is solved using the dynamic-programming value
iteration algorithm [31]. The reward function to be optimized is chosen as:

Rt (Ut , Dt ) = Rtcontrol (Ut ) + αRthealth (Dt ). (9)

Herein, Rtcontrol (Ut ) and Rthealth (Dt ) quantify the rewards relative to control inputs and health
state, respectively, and α ∈ R is a weighting factor, useful to tune the trade-off between risk-averse
and risk-seeking behavior. After learning π(Dt ), UtA is selected as the best point estimate of Ut .
Starting from the updated digital state Dtc at the current time step tc , future prediction is
achieved by unrolling until a prediction time tp the portion of PGM relative to Dt , Qt , Rt , and Ut
(see Fig. 3). All other nodes are removed from the prediction graph, as neither data assimilation
nor actions are performed on the asset while forecasting its evolution. The factorization in Eq. (1)
can be extended over the prediction horizon as:

p(D0NN , . . . , DtNN
c
, D0 , . . . , Dtp , Q0 , . . . , Qtp , R0 , . . . , Rtp , U0 , . . . , Utp |o0 , . . . , otc , uA A
0 , . . . , u tc )
tp h tc h
iY i
Y
∝ ϕhistory
t ϕQoI
t ϕt
control reward
ϕt ϕdata
t ϕ NN
t . (10)
t=0 t=0

U0 U1 U2 U3

D0 D1 D2 D3

Q0 R0 Q1 R1 Q2 R2 Q3 R3

t = tc t = tc + 1 t = tc + 2 t = tc + 3
| | | |

Figure 3: Dynamic decision network employed to predict the future evolution of the digital state
and the associated uncertainty. Circle nodes denote random variables, and diamond nodes denote
the objective function. Directed solid edges represent the variables’ dependencies encoded via con-
ditional probability distributions, while directed dashed edges represent the dependencies encoded
via deterministic functions.

The algorithmic description of the online phase of the proposed digital twinning framework is
reported in Algorithm 1. The operations repeat each time new observational data are provided.
Note that the considered PGM digital twinning framework is general, and can easily be adapted
to deal with physical assets other than civil engineering structures by reorganizing the topology of
the graph, if necessary.

6
Algorithm 1 Online phase – algorithmic description
Input: observational data Ot = ot
1: assimilate ot with the DL models to provide DtNN = dNN t . ▷ (Ot ) → (DtNN )
A NN history
2: infer Dt and Ut by updating dt−1 , given ut−1 , dt , and the CPTs encoding ϕt , ϕNN
t and
control NN A
ϕt . ▷ (Dt−1 , Dt , Ut−1 , ) → (Dt , Ut )
history
3: infer the future evolution of Dt and Ut , given the updated dt , and the CPTs encoding ϕt
control
and ϕt . ▷ (Dtc ) → (Dtp , Utp )
4: select UtA = uA t as the best point estimate of Ut = ut . ▷ (Ut ) → (UtA )
A
5: return control input to be enacted ut , expected evolution of Dt and Ut .

2.2 Numerical models for simulation-based damage identification

As anticipated in the previous section, the assimilation of structural response data to identify the
structural state is carried out through DL models. A simulation-based strategy is exploited to
train DL models on the basis of vibration responses. The training data are numerically generated
by simulating physics-based models so that the effect of damage on the structural response can
be systematically reproduced [32]. In particular, the structure to be monitored is modeled as a
linear-elastic continuum, discretized in space through finite elements. Its dynamic response to
the applied loadings, under the assumption of linearized kinematics, is described by the following
semi-discretized form of the elasto-dynamic problem:

 Mẍ(t) + C(µ)ẋ(t) + K(µ)x(t) = f (t, µ), t ∈ (0, T )
x(0) = x0 , (11)
ẋ(0) = ẋ0 ,


which is referred to as the full-order model (FOM). Here t ∈ (0, T ) denotes time; x(t), ẋ(t), ẍ(t) ∈
RNFE are the vectors of nodal displacements, velocities and accelerations, respectively; NFE is
the number of degrees of freedom (dofs); M ∈ RNFE ×NFE is the mass matrix; C(µ) ∈ RNFE ×NFE
is the damping matrix, assembled according to the Rayleigh’s model; K(µ) ∈ RNFE ×NFE is the
stiffness matrix; f (t, µ) ∈ RNFE is the vector of nodal forces induced by the external loadings;
and x0 and ẋ0 are the initial conditions (at t = 0), in terms of nodal displacements and velocities,
respectively. The mass matrix M is not a function of µ because the mass properties of the structure
are unaffected by the employed damage description or by the operational conditions. The solution
of Problem (11) is advanced in time using the Newmark integration scheme (constant average
acceleration method) [33], to provide xl , ẋl and ẍl , for l = 1, . . . , L, with xl being the vector of
nodal displacements at time l.
With reference to civil structures, we focus on the early detection of damage patterns character-
ized by a small evolution rate, whose prompt identification can reduce lifecycle costs and increase
the safety and availability of the structure. In this context, a localized reduction of the material
stiffness stands as the simplest damage mechanism resulting from a time scale separation between
damage growth and damage assessment, see e.g., [34, 35, 36]. Here, local stiffness reduction is
obtained by parametrizing the stiffness matrix via two variables y ∈ N and δ ∈ R, included in
the parameter vector µ, respectively describing the location and magnitude of the applied stiffness
reduction, similarly to [37, 38, 39]. In particular, y ∈ {0, . . . , Ny } labels the specific damage region,
among a set of predefined Ny damage locations, where y = 0 identifies the damage-free baseline.
The parameter δ ∈ R describes the magnitude of the stiffness reduction taking place within the
predesignated region associated with y.
As NFE increases, the computational cost associated with the solution of the FOM for any
sampled µ also grows, and the generation of synthetic datasets becomes prohibitive. To address
this challenge, a projection-based reduced-order model (ROM) is exploited in place of the FOM
to speed up the offline dataset population phase, similarly to [38, 39]. The ROM is obtained by
a proper orthogonal decomposition (POD)-Galerkin reduced basis method [27, 40, 41, 42]. This
reduced-order modeling strategy is chosen because POD has been investigated and validated in
the context of structural dynamics [43, 44] and structural analysis [45, 46], its appealing offline-
online decoupling, and the availability of efficient criteria for the selection of POD basis functions.
It is worth noting that alternative reduced-order modeling approaches can also be employed to
alleviate the computational burden during the offline dataset generation. For instance, one could
use spectral POD [47, 48, 49], or Grassmannian diffusion maps [50], as viable alternatives to the

7
reduced basis method.
The ROM approximation to the solution of Problem (11) is obtained by linearly combining
NRB ≪ NFE POD basis functions wk ∈ RNFE , k = 1, . . . , NRB , as x(t, µ) ≈ Wb x(t, µ), where
W = [w1 , . . . , wNRB ] ∈ RNFE ×NRB is the basis matrix collecting the POD basis functions and
b(t, µ) ∈ RNRB is the vector of unknown POD coefficients. By enforcing the orthogonality between
x
the residual and the subspace spanned by the first NRB POD modes through a Galerkin projection,
the following NRB -dimensional semi-discretized form is obtained:
¨ (t) + Cr (µ)x
b˙ (t) + Kr (µ)b

 Mr x b x(t) = fr (t, µ), t ∈ (0, T )
b(0) = W⊤ x0 ,
x (12)
 ˙ ⊤
x
b(0) = W ẋ0 .

The solution of this reduced-order system is advanced in time using the same strategy employed
for the FOM model, and then projected onto the original FOM space as x(t, µ) ≈ Wb x(t, µ).
Here, reduced matrices Mr , Cr , and Kr , and the reduced vector fr play the same role as their
high-fidelity counterparts, yet with dimension NRB × NRB instead of NFE × NFE , according to the
following relationships:

Mr ≡ W⊤ MW, Cr (µ) ≡ W⊤ C(µ)W,

(13)
Kr (µ) ≡ W⊤ K(µ)W, fr (t, µ) ≡ W⊤ f (t, µ).

The basis matrix W is obtained by POD, exploiting the so-called method of snapshots as
follows. First, a snapshot matrix S = [x1 , . . . , xNS ] ∈ RNFE ×NS is assembled from NS solution
snapshots, computed by integrating in time the FOM solution for different values of parameters
µ. The computation of an optimal reduced basis is then carried out by factorizing S through a
singular value decomposition. We use a standard energy-based criterion to set the order NRB of
the approximation. For further details see, e.g., [27, 42, 23, 21].
To populate the training dataset D, the parametric space of vector µ is taken as uniformly
distributed, and then sampled via the Latin hypercube rule. The number of samples is equal to
the number I of instances collected in D as:

D = {(Ui , yi , δi )}Ii=1 , (14)

where the vibration recordings Ui associated with the i-th sampling of µ, with i = 1, . . . , I, are
labeled by the corresponding values of yi and δi , and are obtained as follows. With reference to
displacement recordings, nodal values in (0, T ) are first collected as Vi = [Wb xL ]i ∈
x1 , . . . , Wb
RNFE ×L by solving Problem (12). The relevant vibration recordings Ui are then obtained as:

Ui = (TVi )⊤ , (15)

where T ∈ RNu ×NFE is a Boolean matrix whose (n, m)–th entry is equal to 1 only if the n–
th sensor output coincides with the m–th dof. In order to mimic the measurement noise, each
vibration recording in D is corrupted by adding an independent, identically distributed Gaussian
noise, whose statistical properties depend on the target accuracy of the sensors. In the following,
the index i will be dropped for ease of notation, unless necessary.

2.3 Data assimilation via artificial neural networks

The ϕdata
t factor in our PGM encodes the assimilation of observed data through the DL models
underlying the identification of the structural health. In this section, we describe the adopted DL
models, the aspects related to their training, and how they are used to assimilate observational
data to detect, locate, and quantify the presence of structural damage.
Every time new observational data U are acquired, they are first processed with a classification
model NNCL to address damage detection/localization. Classification involves the prediction of
an output class to categorize a given input. Here, the classes are those described through the y
parameter. Whenever a damage is identified in the j-th region, j = 1, . . . , Ny , the observational
data U are further processed with regression models NNjRG , one for each damageable region, to
quantify the associated amount of damage δ.
The aforementioned classification and regression tasks are addressed by means of DL models.
The use of DL models for SHM purposes has the advantage of automating the feature engineering

8
stage characterizing the pattern recognition paradigm for SHM [35, 51]. Indeed, a DL model
is trained to select and extract optimized damage-sensitive features from raw sensor recordings
through an end-to-end learning process. Moreover, since the DL model is learned offline, the
structural state can be next assessed in real-time regardless of considering continuous or discrete
variables, which would be difficult to achieve with other optimization techniques, such as nonlinear
programming, stochastic optimization, and metaheuristic methods.
The model NNCL addresses the multi-class classification task underlying the damage detec-
tion/localization problem, namely NNCL : U → b ∈ RNy +1 . The target label b categorizes one
of the Ny + 1 predefined damage scenarios described through parameter y. In particular, b is a
one-hot encoding b = [b0 , . . . , bNy ]⊤ , with entries bm equal to 1 if the target class y is m and 0
otherwise, with m = 0, . . . , Ny . This is needed because DL models cannot operate on nominal data
directly. They require all input variables and output variables to be numeric. The one-hot encoding
converts the nominal feature described by the y parameter into a multidimensional binary vector.
The number of dimensions corresponds to the number of categories, and each category gets its
dimension. Each category is encoded by mapping it to a vector in which the entry corresponding
to the category’s dimension is 1, and the rest are 0.
The estimated counterpart of b is obtained as b b = NNCL (U). By employing a Softmax ac-
tivation function for the output layer of NNCL , the entries of b b = (bb0 , . . . , bbNy )⊤ ∈ RNy +1 are
m
interpreted as the confidence levels bb by which U is assigned to the m-th damage class, with
m = 0, . . . , Ny . In particular, the Softmax activation function converts the real-valued vector
a = (a0 , . . . , aNy )⊤ ∈ RNy +1 , provided by the output layer of NNCL , into a discrete probability
distribution as:
m
exp(a )
b = Softmax(a),
b with bbm (a) = PNy , m = 0, . . . , Ny . (16)
k
k=0 exp(a )

When NNCL is exploited for prediction, the most likely class is selected as the one that best
categorizes the processed measurements U.
The model NNjRG addresses the regression task underlying the damage quantification problem,
namely NNjRG : U → δ ∈ R, with j = 1, . . . , Ny . The estimated counterpart of δ is obtained as
δb = NNjRG (U). Hence, the regression models, one for each damageable region, map the vibration
recordings U associated with the j-th damage region, onto the estimated magnitude of the stiffness
reduction taking place within the relative damage region. Since all NNjRG models are learned
following the same procedure, the index j will be dropped in the following for ease of notation.
Since the space of digital states in the PGM is discrete, the outcomes of NNCL and NNRG
are accommodated within the PGM by discretizing the range in which the damage level δ can
take values in Nδ uniform intervals, thus resulting in Nd = 1 + Nδ Ny possible states. The same
reasoning is followed to compute the confusion matrix encoding the ϕNN
t factor. In particular, ϕNN
t
measures the offline performance of NNCL and NNRG in assimilating noisy FOM data to classify
the digital state, among the Nd possible outcomes of Dt .
The models NNCL and NNRG are trained separately. The datasets dedicated to the training
of NNCL and NNRG are derived from dataset D in Eq. (14) as follows. The dataset used to learn
NNCL is obtained from Eq. (14), as

DCL = {(Ui , bi )}Ii=1 . (17)

The dataset used to learn NNRG is derived from Eq. (14), as

DRG = {(UiRG , δiRG )}IiRG

RG
=1 , (18)

where IRG is the number of training instances in DRG , all characterized by a structural damage
within the same predefined region.
The set of weights and biases parametrizing NNCL is denoted as ΘCL . This is optimized
minimizing the probabilistic categorical cross-entropy [52, 37] LCL between the predicted and
target class labels over DCL :
I Ny
1XX m
LCL (ΘCL , DCL ) = − b log(bbm
i ), (19)
I i=1 m=0 i

9
which provides a measure of the distance between the discrete probability distribution describing
b, and its estimated counterpart b
b = NNCL (U).
The set of weights and biases ΘRG parametrizing NNRG is learned through the minimization
of the following mean squared error loss function:
IRG
1 X
LRG (ΘRG , DRG ) = (δiRG − δbiRG )2 , (20)
IRG iRG =1

which provides a measure of the distance between the target magnitude of the stiffness reduction
δ, and its approximated counterpart δb = NNRG (U).
The algorithmic description of the procedures and computations characterizing the prelimi-
nary offline phase of the proposed digital twinning framework is reported in Algorithm 2. The
implementation details of the deep learning models are reported in Appendix A.

Algorithm 2 Preliminary offline phase – algorithmic description

Input: parametrization of the operational and damage conditions
PGM implementing the prediction graph
1: set up the physics-based numerical model of the structure to be monitored.
2: assemble the snapshot matrix of the structural response via FOM analyses.
3: compute the POD basis functions via singular value decomposition of the snapshots matrix.
4: use the ROM to populate the training dataset D with vibration recordings at sensor location.
5: use the recordings and labels in D to derive DCL and DRG .
6: train the classification model NNCL on DCL and the regression models NNRG on DRG .
7: test the generalization capabilities of NNCL and NNRG on noisy FOM data.
8: compute the confusion matrix encoding the ϕNN t factor.
9: compute the control policy π(Dt ) by solving the planning problem induced by the PGM.
10: return trained DL models, ϕNN t factor, control policy π(Dt ).

3 Numerical experiments
This section demonstrates the proposed methodology for two test cases: an L-shaped cantilever
beam and a railway bridge.
The FOM and ROM in Problem (11) and Problem (12) are implemented in the Matlab en-
vironment, using the redbKIT library [53]. The PGM framework for predictive digital twins is
implemented in Python, using the pgmpy library [54]. All computations have been carried out on a
PC featuring an AMD RyzenTM 9 5950X CPU @ 3.4 GHz and 128 GB RAM. The NN architectures
are implemented through the Tensorflow-based Keras API [55], and trained on a single Nvidia
GeForce RTXTM 3080 GPU card.

3.1 L-shaped cantilever beam

The first test case deals with the L-shaped cantilever beam depicted in Fig. 4. The structure is
made of two arms, each one having a length of 4 m, a width of 0.3 m, and a height of 0.4 m. The
assumed mechanical properties are those of concrete: Young’s modulus E = 30 GPa, Poisson’s ratio
3
ν = 0.2, density ρ = 2500 kg/m . The structure is excited by a distributed vertical load q(t), acting
2
on an area of (0.3×0.3) m close to its tip. The load varies in time according to q(t) = Q sin (2πf t),
with Q ∈ [40, 80] kPa and f ∈ [10, 60] Hz, respectively being the load amplitude and frequency.
Following the setup described in Sec. 2, these parameters have a uniform distribution within their
respective ranges.

3.1.1 Dataset assembly

Synthetic displacement time histories U are obtained in relation to Nu = 8 dofs along the bottom
surface of the structure, to mimic the monitoring system depicted in Fig. 4. Each recording is

10
Figure 4: L-shaped cantilever beam: details of synthetic recordings related to displacements
u1 (t), . . . , u8 (t), loading condition, and predefined damage regions Ω1 , . . . , Ω7 .

provided for a time interval (0, T = 1 s) with an acquisition frequency fs = 200 Hz. Recordings
are corrupted with an additive Gaussian noise yielding a signal-to-noise ratio of 100.
In addition to the damage-free baseline condition, damage is simulated by considering Ny = 7
possible damage classes, each referring to a reduction of the material stiffness within a subdomain
Ωj , with j = 1, . . . , Ny , as depicted in Fig. 4. The stiffness reduction can occur with a magnitude
δ ∈ [30%, 80%], and is held constant within the considered time interval.
The FOM is obtained with a finite element discretization using linear tetrahedral elements and
resulting in NFE = 4659 dofs. The basis matrix W is obtained from a snapshot matrix S, assembled
through 400 evaluations of the FOM, at varying values of the input parameters µ = (Q, f, y, δ)⊤
sampled via Latin hypercube rule. By prescribing a tolerance ϵ = 10−3 on the fraction of energy
content to be disregarded in the approximation, the order of the ROM approximation turns out
to be NRB = 56.
The dataset D is built with I = 10, 000 instances collected using the ROM. This is then
employed to train NNCL and NNRG , as described in the previous section. In the absence of
experimental data, the testing phase of NNCL and of NNRG is carried out through noise-corrupted
FOM solutions. In particular, the asset is monitored by processing batches of Nobs = 10 noisy
observations, relative to the same damage location y and damage magnitude δ, yet featuring
varying operational conditions set by Q and f . As the health of the asset evolves over time, the
DT assimilates a batch of noisy observations {Uk }N k=1 at each time step, to dynamically estimate
obs

the variation in the structural health parameters underlying the digital state.

3.1.2 Digital twin framework

The two structural health parameters within the digital state are d = (y, δ)⊤ . In order to accom-
modate the outcome of the DL models within the PGM and to compute the CPT encoding the
ϕNN
t factor, the range in which the damage level δ can take values is discretized in Nδ = 6 intervals
{[30%, 35%], [35%, 45%], [45%, 55%], [55%, 65%], [65%, 75%], [75%, 80%]}, thus resulting in Nd = 43
possible digital states. The number of δ intervals and the width of each interval are chosen arbi-
trarily, and there are no restrictions in this respect. The resulting digital states are then sorted to
follow the lexicographic order.
The confusion matrix reported in Fig. 5 measures the offline performance of NNCL and NNRG
in assimilating noisy FOM data to classify the digital state, among the Nd possible outcomes
of Dt . The (unknown) ground truth digital state is detected by the DL models with an overall
classification accuracy of 93.61%. Moreover, it can be argued from the confusion matrix that most
of the misclassifications are due to the damage scenarios related to a stiffness reduction within Ω6
or within Ω7 . This is a quite expected outcome since measurements closer to the clamped side are
only marginally affected by the presence of damage close to the free end of the beam, thus yielding
a smaller sensitivity of sensor recordings to damage. This confusion matrix then serves as the CPT

11
 100%
1

2 75%

Classification accuracy
3

Target state
4 50%

5
25%
6

7
0%
1 2 3 4 5 6 7
Predicted state
Figure 5: L-shaped cantilever beam - Confusion matrix measuring the offline performance of the
DL models in correctly categorizing the digital state. Results are reported in terms of classification
accuracy, measuring how observational data are classified with respect to the ground truth digital
state. Digital states are ordered first for damage location and then for damage level.

encoding the ϕNNt factor.

For the present case, we consider four possible control inputs, each provided with a CPT
modeling the transition probability p(Dt+1 |Dt , UtA = uAt ) from Dt to Dt+1 after taking the action
A history
ut , and collectively encoding the ϕt factor. These internal models of how structural health is
expected to evolve do not reflect the prescribed ground truth evolution, which is unknown to the
DT. The considered control inputs are the following:
• Do nothing (DN) action. There is no maintenance action planned in this case and the physical
state will evolve according to a stochastic deterioration process.
• Minor imperfect maintenance (MI) action. A maintenance action is performed and the asset
may be restored from its current condition to a healthier damage state. This can be traced
back to, e.g., patching and sealing cracked surfaces, rectifying and replacing expansion joints,
or tightening/replacing loose/missing knot bolts for steel members.
• Major imperfect maintenance (MA) action. A maintenance action is performed and the
asset may be restored from its current condition to a healthier damage state, with a higher
probability of improvements than in the previous case. This can be traced back to, e.g.,
repairing heavily damaged slabs, piers, and steel members, and retrofitting compromised
structural elements.
• Perfect maintenance (PM) action. A maintenance action is performed and the asset is re-
stored from its current condition to the damage-free state. This can be traced back to the
replacement of excessively compromised structural elements.

3.1.3 Results: two available actions

We first illustrate the DT capabilities to assimilate observational data and track the structural
health evolution, by restricting the available actions to DN and PM. The (unknown) ground truth
evolution of structural health varies depending on the most recently applied control input, which
can be either DN or PM. In the absence of maintenance, the physical state evolves following a
deterioration process. We prescribe a (simulated) stochastic degradation process that monotoni-
cally deteriorates the structural health. The degradation process features a probability of damage
inception (y ̸= 0) equal to 0.5. Damage may develop in any of the predefined regions with δ = 30%,
and then propagate with δ increments sampled from a Gaussian probability density function (pdf)
centered at 1.5% and featuring a standard deviation equal to 1% (negative increments are rounded
to zero). The effect of a PM action is simulated by restoring the physical state to its undamaged

12
configuration. At each time step during the operation, new observational data are simulated ac-
cording to the (unknown) ground truth structural health and the most recently enacted control
input. The DT assimilates the data and estimates the digital state, eventually suggesting the next
control input to enact. Note that the prescribed trajectory of the structural health parameters is
arbitrarily chosen to fully display the capabilities of the DT. Nevertheless, the DT would be equally
capable of tracking the structural health evolution also considering either more or less aggressive
degradation processes.
The state transition model encoding ϕhistory
t is conditioned on the most recently issued control
input. The transition probability p(Dt+1 |Dt , UtA = uA t ) from Dt to Dt+1 associated with the DN
action assumes that damage may start in any subdomain Ωj , with j = 1, . . . , Ny , with probability
0.05, and then grow to the next δ interval with the same probability. The transition model assumed
for the PM action instead maps the Dt belief to a belief Dt+1 associated with a damage-free
condition, independently of the current condition. The corresponding CPTs are transition matrices,
where the diagonal entries represent the probability of staying in the same state. The lower-left
and upper-right triangles are associated with the probabilities of the system of deteriorating and
improving its condition, respectively. Therefore, the DN transition matrix is a lower-left triangular
matrix, with the highest probability assigned to remaining in the same state, consistent with what
is expected for the deterioration of civil structures. The transition to the next δ interval is the
second most likely transition, while improvements have a zero probability. Once the structure has
reached the last δ interval, it remains in this condition with a probability equal to 1. In contrast,
the PM transition matrix is an upper-right triangular matrix with probabilities equal to 1 in the
first row.
At each time step, the DT selects a control input uA t ∈ {DN, PM} to be enacted on the asset.
Taking a DN action yields a positive reward, but also gives the chance of worsening the asset’s
structural health. On the other hand, the PM action responds to the degrading structural health,
yet yields a negative reward. The computation of the costs associated with the health state and
control inputs encapsulates the evaluation of the ϕreward
t factor quantifying the performance of the
asset. In particular, the two reward functions in Eq. (9) are defined as:

 A  +0.1, if y = 0,
+12, if ut = DN,
Rtcontrol (uA ) = R health
(d t ) = −exp(6δ/5), if y ∈ {1, 2, 3, 4}, (21)
t −20, if uAt = PM,
t
−exp(δ), if y ∈ {5, 6, 7},


where Rtcontrol targets the cost assigned to each control input and Rthealth measures the cost asso-
ciated with the structural health state. These non-dimensional rewards represent indicative values
the decision-maker is charged due to the condition of the structure. Although these values are not
based on real data, actual values are not usually hard to find. State agencies and companies pro-
vide lists with services and costs [56]. The three cases in Rthealth distinguish between the absence
of damage, the presence of damage within the harm closed to the clamped side, and the presence
of damage far from the clamp, respectively. Note how these penalize the progressive deterioration
of the structural health as a function of δ. Rthealth can resemble a variety of aspects, like reduction
in the level of service due to deterioration, working accidents, structural reliability, and structural
failure probability [56].
During the offline phase, we solve the planning problem induced by the PGM to compute the
control policy π(Dt ), which maps the digital state belief to actions and encodes the control factor
ϕcontrol
t . The optimization of Rt (Ut , Dt ) is carried out as described in Sec. 2.1, assuming a discount
factor γ = 0.95 and a weighting factor α = 2. The computed control policy π(Dt ) recommends
that the asset operates until when δ ∈ [65%, 75%] and δ ∈ [75%, 80%], respectively if y ∈ {1, 2, 3, 4}
and if y ∈ {5, 6, 7}, at which point it should be repaired.
Fig. 6 depicts a simulated online phase of the DT up to time step tc = 50. Results are reported
in terms of the (unknown) ground truth digital state, and the corresponding DT estimate after
assimilating the observational data. The graphs report the evolution of the digital state only for the
damaged regions, nevertheless, damage can potentially affect all Ω1 , . . . , Ω7 predefined damaged
regions. The DT proves capable of accurately tracking the digital state evolution with relatively
low uncertainty. The corresponding estimation of the control inputs is reported in the bottom part
of the figure, demonstrating that the DT is able to promptly suggest the PM action within one
time step of when the (unknown) ground truth structural health demands it.
Fig. 7 depicts the predicted evolution of the digital state and of the corresponding informed

13
 Damage probability in region 4
1.0
80%
Digital twin
60%

Damage
30%
0% No Damage
Damage probability in region 6
80%
0.5
60% Ground truth
Damage

30%
0% No Damage
Action probability
PM Optimal control
Actions

DN
0 10 20 30 40 50 0.0
Time step

Figure 6: L-shaped cantilever beam - Online phase of the digital twin framework with two possible
actions: DN (do nothing), and PM (perfect maintenance). Probabilistic and best point estimates
of: (top) digital state evolution against the ground truth digital state; (bottom) control inputs
informed by the digital twin, against the optimal control input under ground truth. In the top
A
panels the background color corresponds to p(Dt |Dt−1 , DtNN , Ut−1 = uAt−1 ). In the bottom panel
it corresponds to p(Ut |Dt ).

control inputs, starting from tc = 50. The prediction horizon is extended over 20 time steps in
the future so that tp = tc + 20. The DT prediction engine informs about the expected future
degradation of the structural health, allowing to plan future interventions.

Damage probability in region 6

1.0
80%

60%
Damage

30% 0.5
0% No Damage
Action probability
PM
Actions

DN
0 5 10 15 20 0.0
Time step
Figure 7: L-shaped cantilever beam - Digital twin future predictions with two possible actions:
DN (do nothing), and PM (perfect maintenance). The starting time is tc = 50. In the top panel
the probability p(Dt |Dt−1 , Ut−1 ) relates to the amount of damage in Ω6 . In the bottom panel it
corresponds to p(Ut |Dt ).

14
3.1.4 Results: four available actions
We now consider all four possible control inputs. We prescribe a stochastic degradation process
with a probability of damage inception (y ̸= 0) equal to 0.5. Damage may develop in any of the
predefined regions with damage level sampled from a uniform distribution δ ∈ [30%, 70%], and
then propagate as in the previous case. This more aggressive degradation process is used to spot
in a few time steps the effectiveness of the decision-making capabilities of the DT. The effect of the
MI and MA actions on the asset is simulated according with stochastic repair processes, for which
the structural health is forced to monotonically improve. The effect of a MI action is simulated
with δ decrements sampled from a Gaussian pdf centered at −12.5% and featuring a standard
deviation equal to 1%, while the effect of a MA action is modeled with δ decrements sampled
from a Gaussian pdf centered at −17.5% and featuring a standard deviation equal to 1%. In both
cases, the damage-free condition is assumed to be recovered if the resulting structural state features
δ < 30%.
The transition model p(Dt+1 |Dt , UtA = uA t ) associated with the MI action assumes no improve-
ment in the structural health with probability 0.1, improvement of one δ interval with probability
0.75, and improvement of two δ intervals with probability 0.15. The resulting CPT is an upper-
right triangular transition matrix, as deterioration from any state upon a repair action is assumed
to have zero probability. The highest probability is assigned to improvements of one δ interval,
followed by improvements of two δ intervals. There is also a lower probability of remaining in the
same deteriorated state, which reflects a failed maintenance. Similarly, the MA action assumes no
improvement with probability 0.05, improvement of one δ interval with probability 0.3, improve-
ment of two δ intervals with probability 0.4, and improvement of three δ intervals with probability
0.25. In this case, the highest probability is assigned to improvements of two δ intervals, followed
by improvements of one δ intervals, three δ intervals, and finally, the lowest probability is associated
with the possibility of a failed maintenance.
The two reward functions in Eq. (9) are chosen as:

+12, if uA t = DN,

 +0.1, if y = 0,

 A
−20, if u = PM,

Rtcontrol (uA ) = t
R health
(dt ) = −exp(5δ), if y ∈ {1, 2, 3, 4}, (22)
t −8, if uA t = MI,
t
−exp(4δ), if y ∈ {5, 6, 7}.

 
−15, if uA t = MA,


We assume a discount factor γ = 0.95, and a weighting factor α = 2.5. The resulting control
policy π(Dt ) recommends that the asset should operate until when δ ∈ [30%, 35%], after which: if
y ∈ {1, 2, 3, 4}, a MI action should be performed when δ ∈ [35%, 45%], and a PM action should
be performed when δ > 45%; while, if y ∈ {5, 6, 7}, the MI and MA actions should be performed,
respectively when δ ∈ [35%, 55%] and when δ ∈ [55%, 75%], and a PM action should be performed
when δ > 75%.
Fig. 8 depicts a simulated online phase of the DT up to tc = 50. The DT accurately tracks
the digital state evolution and timely suggests the appropriate control inputs most of the time. In
particular, the DT proposes the optimal control input, except for the time steps t = 43 and t = 50
featuring a sub-optimal action. In both cases, a MI action is proposed in place of a DN, because
the DT estimates a δ ∈ [35%, 45%] instead of a δ ∈ [30%, 35%] related to a stiffness reduction
within Ω7 . This is in line with what was observed in the confusion matrix of Fig. 5, due to the
limited sensitivity of recordings to damage scenarios affecting the terminal region of the beam. This
peculiar type of misclassification turns out to be the most pathological in the confusion matrix and
is therefore capable of potentially spoiling the assimilation of observational data. Nevertheless, the
DT reverts to correctly tracking the structural health of the asset within one time step.
Fig. 9 depicts the predicted evolution of the digital state and control inputs, from tc = 21 and
over 20 time steps in the future. The DT prediction correctly suggests taking with high probability
a MA action, followed by two MI actions, and accordingly predicts the corresponding evolution of
the structural health. Comparing the DT prediction with what is effectively experienced during
the online phase (see Fig. 8), note how the DT prediction closely resembles the actual evolution
of the digital state and control inputs. This is a remarkable result in terms of DT prediction
capabilities, since the DT is not aware of the future values of the structural health parameters,
and the relative transition models do not match their real (stochastic) evolution.

15
 Damage probability in different regions
80% 1.0
1

Damage
60% Digital twin
30%
0%
80%
2
Damage

60% Ground truth

30%
0%
80%
3
Damage

60%

30%
0%
0.5
80%
5
Damage

60%

30%
0%
80%
7
Damage

60%

30%
0%
Action probability
MA
Actions

MI
PM
DN 0.0
0 10 20 30 40 50
Time step
Figure 8: L-shaped cantilever beam - Online phase of the digital twin framework with four possible
actions: DN (do nothing), PM (perfect maintenance), MI (minor imperfect maintenance), and
MA (major imperfect maintenance). Probabilistic and best point estimates of: (top) digital state
evolution against the ground truth digital state; (bottom) control inputs informed by the digital
twin, against the optimal control input under ground truth. In the top panels the background color
corresponds to p(Dt |Dt−1 , DtNN , Ut−1
A
= uA
t−1 ). In the bottom panel it corresponds to p(Ut |Dt ).

3.2 Railway bridge

The second case study concerns the railway bridge depicted in Fig. 10. It is an integral concrete
portal frame bridge located along the Bothnia line in the Swedish suburbs of Hörnefors. It features
a span of 15.7 m, a free height of 4.7 m, and a width of 5.9 m (edge beams excluded). The
thickness of the structural elements is 0.5 m for the deck, 0.7 m for the frame walls, and 0.8 m
for the wing walls. The bridge is founded on two plates connected by stay beams and supported
by pile groups. The concrete is of class C35/45, whose mechanical properties are: E = 34 GPa,
3
ν = 0.2, ρ = 2500 kg/m . The superstructure consists of a single track with sleepers spaced 0.65 m
3
apart, resting on a ballast layer 0.6 m deep, 4.3 m wide and featuring a density ρB = 1800 kg/m .
The geometrical and mechanical modeling data have been adapted from former research activities
on the relevant soil-structure interaction, see [57, 58].
The bridge is subjected to the transit of Gröna Tåget trains type, at a speed υ ∈ [160, 215] km/h.
Only trains composed of two wagons are considered, thus characterized by 8 axles, each one carrying
a mass ψ ∈ [16, 22] ton. The corresponding load model is described in [38], and consists of 25
equivalent distributed forces transmitted by the sleepers to the deck through the ballast layer with
a slope 4 : 1, according with Eurocode 1 [59].

16
 Damage probability in region 5
1.0
80%

60%

Damage
30% 0.5
0% No Damage
Action probability
MA
Actions

MI
PM
DN 0.0
0 5 10 15 20
Time step
Figure 9: L-shaped cantilever beam - Digital twin future predictions with four possible actions:
DN (do nothing), PM (perfect maintenance), MI (minor imperfect maintenance), and MA (ma-
jor imperfect maintenance). The starting time is tc = 21. In the top panel the probability
p(Dt |Dt−1 , Ut−1 ) relates to the amount of damage in Ω5 . In the bottom panel it corresponds to
p(Ut |Dt ).

Figure 10: Hörnefors railway bridge.

3.2.1 Dataset assembly

Synthetic displacement time histories U are obtained from Nu = 10 sensors deployed as depicted in
Fig. 11. Each recording is provided for a time interval (0, T = 1.5 s) with an acquisition frequency
fs = 400 Hz. This setting allows to record train passages at the lowest speed of 160 km/h, and
properly catches the structural response at the maximum speed of 215 km/h. Recordings are
corrupted with an additive Gaussian noise yielding a signal-to-noise ratio of 120.
In addition to the undamaged condition, the presence of damage in the structure is accounted
for using a localized stiffness reduction that can take place within Ny = 6 predefined subdomains
Ωj , with j = 1, . . . , Ny , as depicted in Fig. 11. The stiffness reduction can occur with a magnitude
δ ∈ [30%, 80%], and is kept fixed while a train travels across the bridge.
The FOM features NFE = 17, 292 dofs, resulting from a finite element discretization with an
element size of 0.80 m and a reduced size of 0.15 m for the deck, to enable a smooth propagation of
the traveling load. The presence of the ballast layer is accounted for through an increased density for
the deck and for the edge beams. The embankments are accounted for through distributed springs,
3
modeled as a Robin mixed boundary condition (with elastic coefficient krobin = 108 N/m ) applied
on the surfaces facing the ground. The structural dissipation is modeled by means of a Rayleigh’s
damping matrix, assembled to account for a 5% damping ratio on the first two structural modes.
The ROM is obtained from a snapshot matrix S, assembled through 400 evaluations of the FOM
for different values of parameters µ = (υ, ψ, y, δ)⊤ . By setting the error tolerance to ϵ = 10−3 ,
NRB = 133 POD modes are to be considered.
The training dataset D is built with I = 10, 000 instances collected using the ROM. Also in

17
Figure 11: Railway bridge: details of synthetic recordings related to displacements u1 (t), . . . , u10 (t),
and predefined damage regions Ω1 , . . . , Ω6 .

this case, the testing phase of NNCL and of NNRG is carried out considering noisy FOM solutions.
The monitoring of the asset is then simulated by assimilating Nobs = 1 noisy observations at each
time step. As the structural health of the bridge evolves over time, the DT estimates the variation
in the structural health parameters every time a train travels across the bridge.

3.2.2 Digital twin framework

As in the previous case, the two structural health parameters within the digital state are d =
(y, δ)⊤ . The range in which the damage level δ can take values is discretized in Nδ = 6 intervals.
The resulting Nd = 37 possible digital states are sorted first for damage location and then for
damage level.
The confusion matrix measuring the offline performance of NNCL and of NNRG in correctly
categorizing the digital state is reported in Fig. 12. The ground truth digital state is detected
with an overall classification accuracy of 91.39%. In this case, the majority of misclassifications
are due to confusing adjacent digital states relative to the same damage location, thus yielding a
tridiagonal band matrix.
For the present case, we consider the following three possible control inputs:
• Do nothing (DN) action. There is no maintenance action planned in this case and the physical
state will evolve according to a stochastic deterioration process.
• Perfect maintenance (PM) action. A maintenance action is performed and the asset is re-
stored from its current condition to the damage-free state.
• Restrict operational conditions (RE) action. The operational conditions of the bridge are
restricted by allowing only lightweight trains, carrying less than 18 ton per axle, to travel
across the bridge. Such a restriction results in a lower deterioration rate, but also yields a
lower revenue generated by the infrastructure.
In the cases where the most recently issued control input is either DN or RE, the physical state
undergoes a degradation process that monotonically deteriorates the structural health. When
operational conditions are not restricted, we prescribe a stochastic degradation process featuring a
probability of damage inception (y ̸= 0) equal to 0.5. Damage may develop in any of the predefined
regions with damage level sampled from a uniform distribution δ ∈ [30%, 35%], and then propagate
with δ increments sampled from a Gaussian pdf centered at 1.5% and featuring a standard deviation
equal to 1% (negative increments are rounded to zero). When the operations are restricted and
only lightweight trains are allowed to travel across the bridge, we instead assume a probability
of damage inception equal to 0.25. In this eventuality, damage may develop with damage level

18
 100%
1

75%

Classification accuracy
2

Target state
3
50%
4

5 25%

6
0%
1 2 3 4 5 6
Predicted state
Figure 12: Railway bridge - Confusion matrix measuring the offline performance of the DL models
in correctly categorizing the digital state. Results are reported in terms of classification accuracy,
measuring how observational data are classified with respect to the ground truth digital state.
Digital states are ordered first for damage location and then for damage level.

sampled from a uniform distribution δ ∈ [30%, 35%], and then propagate with δ increments sampled
from a Gaussian pdf centered at 0.95% and featuring a standard deviation equal to 0.5%. The
resulting trajectory of the structural health parameters is intended to represent periods of gradual
degradation in the structural health, as well as sudden changes due to discrete damage events. Also
in this case, the effect of a PM action is simulated by restoring the physical state to its undamaged
configuration.
The transition model p(Dt+1 |Dt , UtA = uA t ) associated with the DN action assumes that damage
may start in any subdomain Ωj , with j = 1, . . . , Ny , with probability 0.1, and then grow to the
next δ interval with the same probability. For the transition model associated with the RE action,
this probability is assumed to decrease to 0.03. The CPTs associated with the DN and RE
actions are therefore lower-left triangular transition matrices. The highest probability assigned to
remaining in the same state, followed by the transition to the next δ interval, with zero probability
of improvements. The transition model assumed for the PM action instead maps the Dt belief to a
belief Dt+1 associated with a damage-free condition, independently of the current condition. The
CPT associated with the PM action is therefore an upper-right triangular transition matrix with
probabilities equal to 1 in the first row.
In this case, the two reward functions in Eq. (9) are chosen as:
 
 +30, if uA
t = DN,  +0, if y = 0,
Rtcontrol (uA
t )= −250, if uA t = PM, Rthealth (dt ) = −exp(5δ) + 4, if y ̸= 0, (23)
+27, if uA
t = RE, −250, if δ ≥ 79%,
 

where the last contribution in Rthealth penalizes excessively compromised structural states with a
significantly negative reward.

3.2.3 Results
During the offline phase, we solve the planning problem in Eq. (8) by assuming a discount factor
γ = 0.90, and a weighting factor α = 1. The resulting control policy π(Dt ) recommends that the
asset operates in ordinary conditions until when δ ∈ [30%, 35%], after which point it should fall
back to the more conservative RE regime in order to minimize further degradation. Once reached
δ ≥ 65%, the bridge should be finally repaired.
Fig. 13 reports a sample simulation of the DT online phase up to time step tc = 60. The DT
correctly tracks the digital state with relatively low uncertainty. Damage initially develops within
Ω5 , and the DT follows its evolution with a limited delay of at most two time steps, with respect
to the ground truth, due to the need of updating the relative prior belief from the previous time

19
steps. The RE action is suggested as soon as the DT estimates a δ ∈ [35%, 65%], after which point
the DT keeps on tracking the structural health parameters evolving with a lower deterioration rate.
A PM action is finally suggested due to an excessively compromised structural state. A similar
behavior can be observed for the following damage scenario affecting Ω6 .

Damage probability in region 5

1.0
80%
Digital twin
60%
Damage

30%
0% No Damage
Damage probability in region 6
80%
0.5
60% Ground truth
Damage

30%
0% No Damage
Action probability
RE
Actions

PM Optimal control
DN 0.0
0 10 20 30 40 50 60
Time step

Figure 13: Railway bridge - Online phase of the digital twin framework with three possi-
ble actions: DN (do nothing), PM (perfect maintenance), and RE (restrict operational con-
ditions). Probabilistic and best point estimates of: (top) digital state evolution against the
ground truth digital state; (bottom) control inputs informed by the digital twin, against the op-
timal control input under ground truth. In the top panels the background color corresponds to
A
p(Dt |Dt−1 , DtNN , Ut−1 = uA
t−1 ). In the bottom panel it corresponds to p(Ut |Dt ).

Fig. 14 reports the predicted evolution of the digital state and control inputs, from tc = 5
and over 20 time steps in the future. The DT predicts the expected degradation of the structural
health according to the transition model associated with the DN action, before predicting to take
a RE action with relatively high probability after a few time steps. The DT prediction is close to
what is effectively experienced online (see Fig. 13). However, besides having the estimated digital
state two time steps behind the ground truth value, the prediction is also too optimistic in terms
of deterioration rate, which suggests the use of a more refined transition model.

4 Conclusions
In this work we have proposed a predictive digital twin approach to the health monitoring, main-
tenance, and management planning of civil structures, to advance condition-based and predictive
maintenance practices. The presented strategy relies upon a probabilistic graphical model inspired
by [18]. This framework is used to encode the asset-twin coupled dynamical system, the relevant
end-to-end information flow via observational data (physical to digital) and control inputs (digital
to physical), and its evolution over time, all with quantified uncertainty. The assimilation of obser-
vational data is carried out with deep learning models, leveraging the capabilities of convolutional
layers to automatically select and extract damage-sensitive features from raw vibration recordings.
The structural health parameters comprising the digital state are used to capture the variability
of the physical asset. They are continually updated in a sequential Bayesian inference fashion,
according to control-dependent transition dynamics models describing how the structural health is

20
 Damage probability in region 5
1.0
80%

60%

Damage
30% 0.5
0% No Damage
Action probability
RE
Actions

PM
DN
0 5 10 15 20 0.0
Time step
Figure 14: Railway bridge - Digital twin future predictions with three possible actions: DN (do
nothing), PM (perfect maintenance), and RE (restrict operational conditions). The starting time
is tc = 5. In the top panel the probability p(Dt |Dt−1 , Ut−1 ) relates to the amount of damage in
Ω5 . In the bottom panel it corresponds to p(Ut |Dt ).

expected to evolve. The updated digital state is eventually exploited to predict the future evolu-
tion of the physical system and the associated uncertainty. This enables predictive decision-making
about maintenance and management actions.
The computational procedure takes advantage of a preliminary offline phase which involves:
(i) using physics-based numerical models and reduced order modeling, to overcome the lack of
experimental data for civil applications under varying damage and operational conditions while
populating the datasets for training the deep learning models; (ii) learning the health-dependent
control policy to be applied at each time step of the online phase, to map the belief over the digital
state onto actions feeding back to the physical asset.
The proposed strategy has been assessed against the simulated monitoring of an L-shaped
cantilever beam and a railway bridge. In the absence of experimental data, the tests have been
carried out considering high-fidelity simulation data, corrupted with an additive Gaussian noise.
The obtained results have proved the digital twin capabilities of accurately tracking the digital state
evolution under varying operational conditions, with relatively low uncertainty. The framework
is also able to promptly suggest the appropriate control input, within at most two time steps of
when the (unknown) ground truth structural health demands it.
Although the capabilities of health-aware digital twins are showcased in the specific context of
monitoring the structural integrity of civil structures to advance predictive maintenance practices,
the applicability of the presented framework is general. Indeed, the proposed framework can be
adapted for various types of structures and engineering systems by adjusting the components
within the dynamic Bayesian network to align with the specific characteristics of the problem at
hand. The solution to the inverse problem (if any) can be estimated by assimilating available
observational data using methods other than deep neural networks, for instance through Markov
chain Monte Carlo sampling algorithms. Similarly, the state transition models are closely tied to
the employed parametrization of the digital state and the availability of historical data. The same
applies to the available control inputs, which are likely to vary for different structures, such as those
in mechanical or aerospace systems, and the method chosen for solving the associated planning
problem. Additionally, the graph topology can be easily reorganized to adapt to situations where
observational data are not acquired after issuing a control input, or when control inputs are issued
with a different frequency than that governing the digital twin update.
Future research lines will investigate the ability of the digital twin to update the transition
dynamics models by learning from previous data. As suggested by the railway bridge case study,
this will allow for a more accurate prediction of the expected evolution of the digital state, thus
enabling predictive decision-making better tailored to the monitored asset. Another aspect of
interest concerns solving the planning problem induced by the probabilistic model using reinforce-
ment learning algorithms, capable of taking into account a finite planning horizon representing the

21
design lifetime of the asset.

Data availability
The observational data used to run the experiments presented in Sec. 3 are available in the public
repository digital-twin-SHM [28]. The Matlab library for finite element simulation and reduced-
order modeling of partial differential equations employed to generate these data is available in the
public repository Redbkit [53].

Code availability
The implementation code used for the experiments presented in Sec. 3 is available in the public
repository digital-twin-SHM [28]. The code implements the proposed digital twin framework
and can be used to generate the graphs of digital state estimation and prediction reported in this
paper. The DL models trained according to the implementation details reported in Appendix A
are also made available in the same repository.

Acknowledgements
Matteo Torzoni gratefully acknowledges the financial support from the Politecnico di Milano
through the interdisciplinary Ph.D. Grant “Physics-informed deep learning for structural health
monitoring”. Marco Tezzele and Karen E. Willcox acknowledge support from the NASA University
Leadership Initiative under Cooperative Agreement 80NSSC21M0071. Andrea Manzoni acknowl-
edges the project “Dipartimento di Eccellenza” 2023-2027, funded by MUR, and the project FAIR
(Future Artificial Intelligence Research), funded by the NextGenerationEU program within the
PNRR-PE-AI scheme (M4C2, Investment 1.3, Line on Artificial Intelligence).

References
[1] A. Thelen, X. Zhang, O. Fink, Y. Lu, S. Ghosh, B. D. Youn, M. D. Todd, S. Mahadevan,
C. Hu, and Z. Hu, “A comprehensive review of digital twin – part 1: modeling and twinning
enabling technologies,” Structural and Multidisciplinary Optimization, vol. 65, p. 354, 2022.
[2] E. J. Tuegel, A. R. Ingraffea, T. G. Eason, and S. M. Spottswood, “Reengineering Aircraft
Structural Life Prediction Using a Digital Twin,” International Journal of Aerospace Engi-
neering, vol. 2011, p. 154798, 2011.

[3] E. Tuegel, “The Airframe Digital Twin: Some Challenges to Realization,” in 53rd
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference,
2012.
[4] M. Grieves and J. Vickers, “Digital twin: Mitigating unpredictable, undesirable emergent be-
havior in complex systems,” Transdisciplinary Perspectives on Complex Systems: New findings
and approaches, pp. 85–113, 2017.
[5] A. Rasheed, O. San, and T. Kvamsdal, “Digital twin: Values, challenges and enablers from a
modeling perspective,” IEEE Access, vol. 8, pp. 21980–22012, 2020.
[6] S. A. Niederer, M. S. Sacks, M. Girolami, and K. Willcox, “Scaling digital twins from the
artisanal to the industrial,” Nature Computational Science, vol. 1, no. 5, pp. 313–320, 2021.
[7] Y. Ye, Q. Yang, F. Yang, Y. Huo, and S. Meng, “Digital twin for the structural health
management of reusable spacecraft: a case study,” Engineering Fracture Mechanics, vol. 234,
p. 107076, 2020.
[8] M. Podskarbi and D. J. Knezevic, “Digital Twin for Operations - Present Applications and
Future Digital Thread,” in OTC Offshore Technology Conference, (Houston, Texas), May
2020.

22
 [9] A. Phua, C. Davies, and G. Delaney, “A digital twin hierarchy for metal additive manufac-
turing,” Computers in Industry, vol. 140, p. 103667, 2022.
[10] M. Jans-Singh, K. Leeming, R. Choudhary, and M. Girolami, “Digital twin of an urban-
integrated hydroponic farm,” Data-Centric Engineering, vol. 1, p. 20, 2020.
[11] A. Tzachor, S. Sabri, C. E. Richards, A. Rajabifard, and M. Acuto, “Potential and limitations
of digital twins to achieve the sustainable development goals,” Nature Sustainability, vol. 5,
no. 10, pp. 822–829, 2022.

[12] G. Arcieri, C. Hoelzl, O. Schwery, D. Straub, K. G. Papakonstantinou, and E. Chatzi, “Bridg-

ing pomdps and bayesian decision making for robust maintenance planning under model
uncertainty: An application to railway systems,” Reliability Engineering & System Safety,
p. 109496, 2023.
[13] M. G. Kapteyn, D. J. Knezevic, D. Huynh, M. Tran, and K. E. Willcox, “Data-driven physics-
based digital twins via a library of component-based reduced-order models,” International
Journal for Numerical Methods in Engineering, vol. 123, no. 13, pp. 2986–3003, 2020.
[14] B. Moya, A. Badı́as, I. Alfaro, F. Chinesta, and E. Cueto, “Digital twins that learn and correct
themselves,” International Journal for Numerical Methods in Engineering, vol. 123, no. 13,
pp. 3034–3044, 2022.

[15] W. Sisson, P. Karve, and S. Mahadevan, “Digital Twin Approach for Component Health-
Informed Rotorcraft Flight Parameter Optimization,” AIAA Journal, vol. 60, no. 3, pp. 1923–
1936, 2022.
[16] B. Moya, I. Alfaro, D. Gonzalez, F. Chinesta, and E. Cueto, “Physically sound, self-learning
digital twins for sloshing fluids,” PLoS One, vol. 15, no. 6, p. e0234569, 2020.

[17] K. Cherifi, V. Mehrmann, and P. Schulze, “Simulations in a Digital Twin of an Electrical

Machine.” arXiv preprint arXiv:2207.02171, 2022.
[18] M. G. Kapteyn, J. V. Pretorius, and K. E. Willcox, “A probabilistic graphical model foun-
dation for enabling predictive digital twins at scale,” Nature Computational Science, vol. 1,
no. 5, pp. 337–347, 2021.
[19] D. Koller and N. Friedman, Probabilistic Graphical Models: Principles and Techniques. Cam-
bridge, MA: MIT Press, 2009.
[20] S. J. Russell, Artificial Intelligence: A Modern Approach. London, UK: Pearson Education,
2020.

[21] M. Torzoni, A. Manzoni, and S. Mariani, “Structural health monitoring of civil structures: a
diagnostic framework powered by deep metric learning,” Computers & Structures, vol. 271,
p. 106858, 2022.
[22] V. Giglioni, I. Venanzi, V. Poggioni, A. Milani, and F. Ubertini, “Autoencoders for unsuper-
vised real-time bridge health assessment,” Computer-Aided Civil and Infrastructure Engineer-
ing, vol. 38, pp. 959–974, 2023.
[23] M. Torzoni, A. Manzoni, and S. Mariani, “A multi-fidelity surrogate model for structural
health monitoring exploiting model order reduction and artificial neural networks,” Mechanical
Systems and Signal Processing, vol. 197, p. 110376, 2023.

[24] M. Torzoni, A. Manzoni, and S. Mariani, “A Deep Neural Network, Multi-fidelity Surrogate
Model Approach for Bayesian Model Updating in SHM,” in European Workshop on Structural
Health Monitoring, pp. 1076–1086, Springer International Publishing, 2023.
[25] L. Rosafalco, M. Torzoni, A. Manzoni, S. Mariani, and A. Corigliano, “A Self-adaptive Hybrid
Model/data-Driven Approach to SHM Based on Model Order Reduction and Deep Learning,”
in Structural Health Monitoring Based on Data Science Techniques, pp. 165–184, Springer
International Publishing, 2022.

23
[26] S. Eftekhar Azam and S. Mariani, “Online damage detection in structural systems via dynamic
inverse analysis: A recursive Bayesian approach,” Engineering Structures, vol. 159, pp. 28–45,
2018.
[27] A. Quarteroni, A. Manzoni, and F. Negri, Reduced Basis Methods for Partial Differential
Equations: An Introduction, vol. 92. Springer, 2015.
[28] M. Torzoni and M. Tezzele, “digital-twin-shm,” 2023. https://github.com/mtezzele/
digital-twin-SHM.
[29] G. Morcous, “Performance prediction of bridge deck systems using markov chains,” Journal
of Performance of Constructed Facilities, vol. 20, no. 2, pp. 146–155, 2006.
[30] P. D. Thompson and M. B. Johnson, “Markovian bridge deterioration: developing models
from historical data,” Structure and Infrastructure Engineering, vol. 1, no. 1, pp. 85–91, 2005.
[31] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction. Cambridge, MA:
MIT Press, 2018.
[32] T. Taddei, J. D. Penn, M. Yano, and A. T. Patera, “Simulation-Based Classification; a Model-
Order-Reduction Approach for Structural Health Monitoring,” Archives of Computational
Methods in Engineering, vol. 25, pp. 23–45, 2018.
[33] N. M. Newmark, “A method of computation for structural dynamics,” Journal of Engineering
Mechanics, vol. 85, no. 3, pp. 67–94, 1959.
[34] A. Teughels, J. Maeck, and G. De Roeck, “Damage assessment by FE model updating using
damage functions,” Computers & Structures, vol. 80, no. 25, pp. 1869–1879, 2002.
[35] C. Farrar and K. Worden, Structural Health Monitoring: A Machine Learning Perspective.
John Wiley & Sons, 2013.
[36] A. Pandey and M. Biswas, “Damage Detection in Structures Using Changes in Flexibility,”
Journal of Sound and Vibration, vol. 169, no. 1, pp. 3–17, 1994.
[37] L. Rosafalco, A. Manzoni, S. Mariani, and A. Corigliano, “Fully Convolutional Networks
for Structural Health Monitoring through Multivariate Time Series Classification,” Advanced
Modeling and Simulation in Engineering Sciences, vol. 7, p. 38, 2020.
[38] L. Rosafalco, M. Torzoni, A. Manzoni, S. Mariani, and A. Corigliano, “Online structural
health monitoring by model order reduction and deep learning algorithms,” Computers &
Structures, vol. 255, p. 106604, 2021.
[39] M. Torzoni, L. Rosafalco, A. Manzoni, S. Mariani, and A. Corigliano, “SHM under varying
environmental conditions: an approach based on model order reduction and deep learning,”
Computers & Structures, vol. 266, p. 106790, 2022.
[40] F. Chinesta, A. Huerta, G. Rozza, and K. Willcox, “Model Reduction Methods,” in Ency-
clopedia of Computational Mechanics, Second Edition (E. Stein, R. de Borst, and T. J. R.
Hughes, eds.), pp. 1–36, John Wiley & Sons, 2017.
[41] P. Benner, S. Grivet-Talocia, A. Quarteroni, G. Rozza, W. H. A. Schilders, and L. M. Silveira,
eds., Volume 1: System- and Data-Driven Methods and Algorithms. Berlin, Boston: De
Gruyter, 2021.
[42] G. Rozza, G. Stabile, and F. Ballarin, Advanced Reduced Order Methods and Applications
in Computational Fluid Dynamics. Philadelphia, PA: Society for Industrial and Applied
Mathematics, 2022.
[43] G. Kerschen and J. C. Golinval, “Physical interpretation of the proper orthogonal modes
using the singular value decomposition,” Journal of Sound and Vibration, vol. 249, no. 5,
pp. 849–865, 2002.
[44] G. Kerschen, J. C. Golinval, A. Vakakis, and L. A. Bergman, “The Method of Proper Or-
thogonal Decomposition for Dynamical Characterization and Order Reduction of Mechanical
Systems: An Overview,” Nonlinear Dynamics, vol. 41, no. 1, pp. 147–169, 2005.

24
[45] M. Guo and J. S. Hesthaven, “Reduced order modeling for nonlinear structural analysis us-
ing gaussian process regression,” Computer Methods in Applied Mechanics and Engineering,
vol. 341, pp. 807–826, 2018.

[46] M. Tezzele, L. Fabris, M. Sidari, M. Sicchiero, and G. Rozza, “A multi-fidelity approach

coupling parameter space reduction and non-intrusive POD with application to structural
optimization of passenger ship hulls,” International Journal for Numerical Methods in Engi-
neering, vol. 124, no. 5, pp. 1193–1210, 2023.
[47] M. Sieber, C. O. Paschereit, and K. Oberleithner, “Spectral proper orthogonal decomposition,”
Journal of Fluid Mechanics, vol. 792, pp. 798–828, 2016.
[48] A. Towne, O. T. Schmidt, and T. Colonius, “Spectral proper orthogonal decomposition and
its relationship to dynamic mode decomposition and resolvent analysis,” Journal of Fluid
Mechanics, vol. 847, pp. 821–867, 2018.

[49] A. Lario, R. Maulik, O. T. Schmidt, G. Rozza, and G. Mengaldo, “Neural-network learning

of SPOD latent dynamics,” Journal of Computational Physics, vol. 468, p. 111475, 2022.
[50] K. R. dos Santos, D. G. Giovanis, K. Kontolati, D. Loukrezis, and M. D. Shields, “Grass-
mannian diffusion maps based surrogate modeling via geometric harmonics,” International
Journal for Numerical Methods in Engineering, vol. 123, no. 15, pp. 3507–3529, 2022.

[51] C. M. Bishop, Pattern Recognition and Machine Learning. Information Science and Statistics,
New York, NY: Springer, 2006.
[52] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. Cambridge, MA: MIT Press, 2016.
[53] F. Negri, “redbkit, version 2.2,” 2016. http://redbkit.github.io/redbKIT.

[54] A. Ankan and A. Panda, “pgmpy: Probabilistic graphical models using Python,” in Proceed-
ings of the 14th Python in Science Conference (SCIPY 2015), Citeseer, 2015.
[55] F. Chollet et al., “Keras,” 2015. https://keras.io.
[56] K. Papakonstantinou and M. Shinozuka, “Planning structural inspection and maintenance
policies via dynamic programming and Markov processes. Part II: POMDP implementation,”
Reliability Engineering & System Safety, vol. 130, pp. 214–224, 2014.
[57] M. Ülker-Kaustell, Some aspects of the dynamic soil-structure interaction of a portal frame
railway bridge. PhD thesis, KTH Royal Institute of Technology, 2009.
[58] T. Arvidsson and J. Li, Dynamic analysis of a portal frame railway bridge using frequency
dependent soil structure interaction. Master thesis, KTH Royal Institute of Technology, 2011.
[59] European Committee for Standardization, “Part 2: Traffic loads on bridges,” in EN 1991-2
Eurocode 1: Actions on structures, pp. 66–74, 2003.
[60] P. Battaglia, J. B. C. Hamrick, V. Bapst, A. Sanchez, V. Zambaldi, M. Malinowski, A. Tac-
chetti, D. Raposo, A. Santoro, R. Faulkner, C. Gulcehre, F. Song, A. Ballard, J. Gilmer,
G. E. Dahl, A. Vaswani, K. Allen, C. Nash, V. J. Langston, C. Dyer, N. Heess, D. Wierstra,
P. Kohli, M. Botvinick, O. Vinyals, Y. Li, and R. Pascanu, “Relational inductive biases, deep
learning, and graph networks.” arXiv preprint arXiv:1806.01261, 2018.
[61] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep feedforward neural
networks,” Journal of Machine Learning Research, vol. 9, pp. 249–256, 2010.
[62] D. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” in 3rd International
Conference on Learning Representations, vol. 3, pp. 1–13, 2015.

25
A Implementation details
In this Appendix, we discuss the implementation details of the DL models described in Sec. 2.3.
The architectures, as well as the relevant hyperparameters and training options, have been chosen
through a preliminary study, aimed at minimizing LCL and LRG , while retaining the generalization
capabilities of NNCL and of NNjRG , with j = 1, . . . , Ny . Since all NNjRG models share the same
architecture, the index j will be dropped in the following for ease of notation.
In the present work, NNCL and NNRG are set as 12-layers DL models, whose architecture is
outlined in Tab. 1a and in Tab. 2a, respectively. NNCL and NNRG feature a damage-sensitive
feature extractor required to be insensitive to transformations in the input not related to dam-
age. This is implemented through the composition of three one-dimensional (1D) convolutional
units. Convolutional layers naturally embed good relational inductive biases such as locality and
translation equivariance [60, 52], which prove highly effective to detect time correlations within
and across time series. The resulting sparse connectivity and parameter sharing also make them
computationally efficient. Each convolutional unit consists of a convolutional layer, followed by a
Tanh activation function, max pooling, and dropout. The extracted features are expected to be
sensitive to the presence of damage, but insensitive to measurement noise and operational vari-
ability. The extracted features are then reshaped through a flatten layer and run through a stack
of three fully-connected layers: the first two are Tanh-activated, while the output layer of NNCL is
Softmax-activated, and the output layer of NNRG has no activation function.

Table 1: NNCL - (a) employed architecture, and (b) selected hyperparameters and training options.

(a) (b)

Layer Output shape Activation Input Convolution kernel size: 25, 13, 7
Dropout rate: 5%
0 - Input (BCL , L, Nu ) None –
Weight initializer: Xavier
1 - Conv1D (BCL , L, 32) Tanh 0
2 - MaxPooling1D (BCL , L/2, 32) None 1 L2 regularization rate: λCL = 10−3
3 - Dropout (BCL , L/2, 32) None 2 Optimizer: Adam
4 - Conv1D (BCL , L/2, 64) Tanh 3 Batch size: BCL = 32
5 - MaxPooling1D (BCL , L/4, 64) None 4 Initial learning rate: ηCL = {10−3 , 10−4 }
6 - Dropout (BCL , L/4, 64) None 5 Allowed epochs: 250
4
7 - Conv1D (BCL , L/4, 32) Tanh 6 Learning schedule: 5 cosine decay
8 - MaxPooling1D (BCL , L/8, 32) None 7 Weight decay: 0.05
9 - Dropout (BCL , L/8, 32) None 8 Early stop patience: 15 epochs
10 - Flatten (BCL , 4L) None 9 Train-val split: 80 : 20
11 - Dense (BCL , 64) Tanh 10
12 - Dense (BCL , 16) Tanh 11
13 - Dense (BCL , Ny + 1) Softmax 12

Table 2: NNRG - (a) employed architecture, and (b) selected hyperparameters and training options.

(a) (b)

Layer Output shape Activation Input Convolution kernel size: 25, 13, 7
Dropout rate: 10%
0 - Input (BRG , L, Nu ) None –
Weight initializer: Xavier
1 - Conv1D (BRG , L, 32) Tanh 0
2 - MaxPooling1D (BRG , L/2, 32) None 1 L2 regularization rate: λRG = 10−3
3 - Dropout (BRG , L/2, 32) None 2 Optimizer: Adam
4 - Conv1D (BRG , L/2, 64) Tanh 3 Batch size: BRG = 32
5 - MaxPooling1D (BRG , L/4, 64) None 4 Initial learning rate: ηRG = {10−3 , 10−4 }
6 - Dropout (BRG , L/4, 64) None 5 Allowed epochs: 250
4
7 - Conv1D (BRG , L/4, 32) Tanh 6 Learning schedule: 5 cosine decay
8 - MaxPooling1D (BRG , L/8, 32) None 7 Weight decay: 0.05
9 - Dropout (BRG , L/8, 32) None 8 Early stop patience: 15 epochs
10 - Flatten (BRG , 4L) None 9 Train-val split: 80 : 20
11 - Dense (BRG , 64) Tanh 10
12 - Dense (BRG , 16) Tanh 11
13 - Dense (BRG , 1) None 12

Using the Xavier’s weight initialization [61], NNCL and NNRG are trained by minimizing the

26
following loss functions, respectively:

LR 2
CL (ΘCL , DCL ) = LCL (ΘCL , DCL ) + λCL ∥ΘCL ∥2 , (24)
LR
RG (ΘRG , DRG ) = LRG (ΘRG , DRG ) + λRG ∥ΘRG ∥22 , (25)

where λCL and λRG denote the L2 regularization rate over the relative model parameters ΘCL
and ΘRG . The loss functions LR R
CL and LRG are minimized using the first-order stochastic gradient
descent optimizer Adam [62], for a maximum of 250 allowed epochs. The corresponding learning
rates ηCL and ηRG are initially set to {10−3 , 10−4 }, and decreased for 4/5 of the allowed training
steps using a cosine decay schedule with weight decay equal to 0.05. The optimization is carried
out considering an 80:20 splitting ratio of the dataset for training and validation purposes, with
20% of the data randomly taken and set aside to monitor the learning process. We use an early
stopping strategy to interrupt learning, whenever the loss function value attained on the validation
set does not decrease for a prescribed number of patience epochs in a row. The hyperparameters
and training options for NNCL and for NNRG are reported in Tab. 1b and in Tab. 2b, respectively.

27