Received: 16 July 2019 Revised: 29 January 2020 Accepted: 19 March 2020

DOI: 10.1002/stc.2556

RESEARCH ARTICLE

Modified Hamiltonian Monte Carlo-based Bayesian finite

element model updating of steel truss bridge

Shubham Baisthakur Arunasis Chakraborty

Department of Civil Engineering, Indian

Institute Of Technology, Guwahati, India Summary
The aim of this study is to develop a Hamiltonian Monte Carlo-based algorithm
Correspondence
Arunasis Chakraborty, Department of for finite element model updating in the Bayesian framework. The proposed
Civil Engineering, Indian Institute Of algorithm uses adaptive prior-based approach, which helps to generate the inter-
Technology, Guwahati, India.
Email: [email protected]
mediate pdf s. Numerical analysis is carried out with different coefficient of
variations of the prior for model updating. Guidelines for their proper selec-
tion procedure are also prescribed in this work. The efficiency of the proposed
method is demonstrated using synthetic experiments and actual test data for
updating the finite element model of a steel truss bridge. Finally, performance
of this algorithm is compared with the standard Markov chain Monte Carlo
algorithm to demonstrate its advantages.

K E Y WO R D S
Bayesian inference, maximum likelihood, Metropolis–Hasting algorithm, model updating, modified
Hamiltonian Monte Carlo simulation

1 I N T RO DU CT ION

Due to the recent advancements in computer simulations and finite element analysis, large complex structures can be
easily modeled for detailed analysis and design. These models are often developed to get more insights into the structural
performance and its behavior under various anticipated loading conditions. Once the model is updated, it can be used for
design life assessment and reliability evaluations of the existing structures1 using advanced simulation techniques. For
these analyses to be reliable, computer models developed for such studies must be sufficiently accurate to replicate the
realistic physical behavior of the actual structure. However, it is often observed that the model-predicted response differs
from the actual response as it is an approximate representation of the in situ structure due to various modeling assump-
tions and uncertainties.2 These uncertainties can be attributed to unmodeled structural behavior (e.g., chaotic systems)
or simplification of complex boundary conditions (e.g., assumption of ideal fixed joint), which are classified under mod-
eling uncertainties. They can also arise due to inherent variability in the parameters involved, whose exact values cannot
be fixed beforehand and are classified as parametric uncertainties. All these uncertainties often lead to erroneous predic-
tion of the structural response. In this context, the main objective of model updating is to identify the uncertain model
parameters and figure out their actual values to obtain the model-predicted response that closely matches with the actual
structural behavior. The structural models with readjusted values of the uncertain parameters are called the updated mod-
els. These updated models have found wide range of applications in structural health monitoring,3-6 structural control,7
uncertainty quantification8,9 and reliability/risk assessment.10,11
Model updating problem can be solved in two different ways, that is, the deterministic approach and probabilistic
approach. In deterministic approach, model updating is usually cast as an optimization problem where the parame-
ters are obtained by minimizing a goodness-of-fit function, which quantifies the error between the recorded response
and the model-predicted response (e.g., natural frequencies, mode shapes, etc.). Over a period of time, various deter-

Struct Control Health Monit. 2020;27:e2556. wileyonlinelibrary.com/journal/stc © 2020 John Wiley & Sons, Ltd. 1 of 22
https://doi.org/10.1002/stc.2556
2 of 22 BAISTHAKUR AND CHAKRABORTY

ministic model updating formulations using genetic algorithms,12 sensitivity-based optimization,13,14 single-objective
optimization,15 and multiobjective optimization16 are proposed. Jaishi and Ren16 have demonstrated the efficiency of
multiobjective optimization over single-objective optimization by comparing the simulated response of a simply sup-
ported beam. In this work, single-objective optimization is carried out using gradient-based optimization scheme while
pareto optimality is adopted for multiobjective version. The robustness of multiobjective optimization is subsequently
established for field application using a 6-span continuous bridge. However, deterministic algorithms, in general, do
not utilize any information related to the initial modeling assumptions and uncertainties associated with the measured
structural responses. Thus, due to the nature of this problem, it may lead to an overfitted model such that it can match
with the measured structural response very well, but fails to accurately predict the response for any other loading condi-
tions. Since the focus of the deterministic algorithm is to minimize the data fitting error using optimization techniques,
model parameters obtained from this procedure give the single best output, with no control over its accuracy in the
future. Jaishi and Ren16 have studied the sensitivity of such algorithm when the recorded data are contaminated with
measurement noise. Since the process of model updating is a reverse formulation, there exists an inherent ill-posedness
where more than one potential solution may exist,17 which is difficult to identify within the deterministic model updat-
ing framework. The probabilistic approach, on the other hand, involves complete information from the recorded data
as well as initial modeling assumptions. It also incorporates likelihood function, which, unlike the goodness-of-fit func-
tion, is further used for statistical inference.18,19 Also, probabilistic model updating does not offer the single best output;
however, it forms a probability distribution over the parameter space, specifying the relative plausibility of each uncer-
tain model parameter achieving a particular value, which is conditioned by the initial modeling assumptions and the
recorded structural behavior. Recording structural response during on-site experiment followed by parameter identifica-
tion, which are subsequently used for model updating, in itself, is a challenging task and involves lot of complexities.
Due to the limitations of the number of sensors that can be actually mounted and monitored during on-site experi-
ments, limited responses are available in most of the time. Moreover, changes in ambient conditions (e.g., temperature
and wind) induce changes in the structural behavior20,21 very often, which can introduce uncertainties in the response.
Also, there is a possibility of distortion of the recorded signal due to the measurement noise. Owing to these complex-
ities and uncertainties involved in identifying actual parameters, model updating is best tackled in the probabilistic
domain.
In this context, as large number of random experiments are difficult to be carried out for model updating in the
probabilistic approach, frequentist definition of probability cannot be applied to quantify the plausibility of uncertain
parameters achieving a particular value. To address this issue, model updating in the probabilistic domain is dealt within
the Bayesian framework. The groundbreaking work for this purpose is done by Beck and Katafygiotis18 and Katafygio-
tis et al.19 where the authors have extended the central idea of Bayesian inference for updating the continuous-valued
uncertain parameters 𝜃 using the measurement of continuous-valued structural response D.18 Through their continual
research, the authors have developed a general Bayesian statistical formulation for updating the model parameters and
quantitative assessment of their relative accuracy using Bayesian inference.17,18 In this approach, the uncertain parame-
ter 𝜃 attaining a specific value is considered as a hypothesis where the initial degree of belief towards this hypothesis is
updated using measurement D in terms of posterior probability. It is conditioned by the structural behavior and a class of
mathematical model C such that

p(D|𝜃, C)p(𝜃|C)
p(𝜃|D, C) = = K0 p(D|𝜃, C)p(𝜃|C). (1)
p(D|C)

The term p(𝜃|D, C) is the updated or posterior pdf of the uncertain model parameter 𝜃 based on the findings D (i.e.,
measurement), which provides a quantitative judgment of the plausibility of a set of uncertain parameter values. The
term p(D|𝜃, C) is called the likelihood function, which quantifies the likelihood of obtaining the structural response D for
a specified set of model parameter values 𝜃. The term p(𝜃|C) is called the prior pdf, as it gives the initial plausibility of
1
parameters 𝜃 for a mathematical model in its actual state when the structural response is not available. K0 = p(D|C) is a
normalizing constant, such that integrating the right-hand side of the above equation over the parameter space 𝛩 gives
unity, that is,
p(𝜃|D, C) d𝜃 = K0 p(D|𝜃, C)p(𝜃|C) d𝜃 = 1. (2)
∫Θ ∫Θ
In the above formulation, exact evaluation of the posterior pdf involves a multidimensional integral, which cannot
be performed using analytical procedure. In general, the exact nature of this Bayesian predictive pdf is not known a
BAISTHAKUR AND CHAKRABORTY 3 of 22

priori, and hence, it is approximated by a Gaussian distribution centered at the optimal points, which globally mini-
mizes the objective function.18 An algorithm for efficient searching of parameter space to find these optimum points
has also been proposed in the literature.17 However, the asymptotic Gaussian approximation is feasible only for the
identifiable cases where a nonconvex global optimization exists, while for locally identifiable cases, finding these opti-
mal points involves a series of local optimization problems. Therefore, the computational efforts required to solve this
problem grow with the dimension of the parameter space, which makes it infeasible for high-dimensional cases.22 These
factors of model identifiability and asymptotic approximation of the posterior pdf are studied by Katafygiotis et al,17
where the authors have also addressed practical issues like noise and incomplete measurements due to limited number of
sensors.
Earlier work in model updating is carried out using ambient time histories for updating the modal parameters. Sub-
sequently, the formulation is extended to incorporate load-dependent Ritz vectors6 and eigenvalues5 as the parameter
to detect damage. Here, it may be noted that the Metropolis–Hastings (MH) algorithm23 is often used to simulate the
samples of the random variable distributed according to an arbitrary continuous pdf. Hence, this method finds its way
into Bayesian model updating where samples for the successive iterations are simulated by this method. It has been
successfully implemented in model updating, bypassing other computationally exhaustive optimization algorithms.24,25
However, since the high-probability region of the target pdf p(𝜃|D, C) is concentrated in narrow bands within the proba-
bility space, direct application of the MH technique to simulate samples distributed as per target pdf may lead to higher
rejection rate and form a chain with large number of repeated samples, which may lead the chain into nonergodic state.
These problems are dealt with various adaptive MH algorithms developed in the recent past. The adaption is done either
on the proposal pdf used to generate the candidate state3,26,27 or the target posterior distribution22,28,29 to obtain the infor-
mation about the manifold in an iterative manner. The later case involves a sequence of intermediate pdfs such that they
gradually converge to the posterior pdf while maintaining the proper acceptance rate. In this context, use of intermediate
pdfs faces the problem of high rejection rate in Markov chains due to the narrow regions of high probability concentra-
tion, which cannot be anticipated beforehand. Beck and Au22 have proposed an adaptive MH (AMH) algorithm that uses
predefined sequence of intermediate pdfs and requires kernel density estimation to approximate them, which makes it
inefficient for high-dimensional problems. Ching and Chen28 have proposed a new adaptive approach called transitional
Markov chain Monte Carlo (TMCMC) where it uses resampling to bypass the kernel density estimation. Unlike AMH,
TMCMC can automatically generate intermediate pdfs. In a recent work, Wan and Ren30 have used Gaussian process as
a surrogate to bypass the evaluation of time-consuming finite element models. The authors have also presented a novel
variance-based global sensitivity analysis scheme to select uncertain parameters, which, unlike other local sensitivity
analysis, can also study the joint effects of input on the model output. Both these schemes, along with the Delayed Rejec-
tion Adaptive Metropolis (DRAM) algorithm, are then applied for model updating of a 120 m long pedestrian foot bridge,
which significantly reduce the computational cost. Thus, the recent trends for developing efficient simulation techniques
include the algorithms inspired by Darwinian's theory in the genetic algorithm,31 Gibbs sampling,32 enhanced MCMC
technique,33 and online model updating.34
All the above-mentioned Markov chain-based simulation techniques exhibit random walk while generating the can-
didate states. It needs longer time to discover the areas of significant probability content. Duane et al.35 have presented
a new technique for the numerical simulation of the lattice field theory called the hybrid Monte Carlo (also known as
Hamiltonian Monte Carlo [HMC]) technique. It uses a deterministic mechanism inspired by the principles of Hamilto-
nian dynamics for simulating the samples following a target distribution. Thus, it alleviates random walk with a consistent
exploration of the probability space. Its application in Bayesian model updating is first studied by Cheung and Beck,36
where the authors have used this method for finite element model updating of a 10-story building using simulated data. In
general, HMC offers faster convergence as compared with conventional random walk-based MCMC methods.36 Boulka-
ibet et al.37 have proposed an improved version of the HMC technique called the shadow hybrid Monte Carlo technique
where the sampling is done from a modified Hamiltonian function instead of the conventional Hamiltonian function. In
both these versions of the HMC algorithm, the authors have used the leapfrog algorithm to generate the candidate state
where the gradient of the logarithm of posterior pdf acts as a guide to explore the high-probability region of the target pdf.
However, numerical evaluations of the gradient over the uncertain parameter space are computationally exhaustive and
even infeasible for multidimensional problems involving large-scale complex finite element models. In a recent work,
Wang et al.38 have proposed a modified version of HMC algorithms where the authors have derived a closed-form solution
for the simulation of standard Gaussian random variables. These authors have also studied its application in reliability
analysis using subset simulation where this method has shown improvements over the standard MCMC algorithm.
4 of 22 BAISTHAKUR AND CHAKRABORTY

1.1 Problem formulation

The literature review presented above summarizes the development of Bayesian model updating over the last few decades.
It also highlights the recent trends aiming at improving the efficiency of model updating techniques. However, in all the
adaptive simulations developed so far, adaption is done either on the proposal pdf or the target posterior pdf, while the
effect of using adaptive prior pdf on model updating is not yet explored. The recently developed HMC algorithm38 has
shown great potential in the subset simulation. Therefore, application of this algorithm in model updating problem is
expected to improve the efficiency of searching the probability space, which is yet to be studied. Also, most of the studies
involving Bayesian algorithms are focused on its performance using synthetic experiments21,31,32,39-42 and/or small-scale
laboratory applications.37,43,44 Only a handful of studies are available for large-scale field structures.3,4,8,33,45 In particular,
the application of HMC algorithm involving full-scale structure has never been studied before. With this in view, the
present study aims to address the following issues:
• develop an adaptive HMC algorithm in the Bayesian model updating framework and also validate its efficiency and
accuracy using controlled synthetic experiments prior to its application on actual structure;
• propose a new adaptive scheme for the generation of intermediate pdfs based on the magnitude of error function and
prior probabilities;
• apply the proposed HMC algorithm to update the Finite Element (FE) model of a full-scale truss bridge for experimental
validation where modal parameters are identified using time-frequency analysis of the recorded ambient vibration
response; and
• demonstrate the importance of the selection of prior pdf in model updating of actual structures to address special issue
related to field problems where modeling and assumptions play vital role and discusses its effect on the health of the
structure.

2 MOTIVATION BEHIND HMC S IMULATION

HMC is a nonrandom walk-based MCMC method, which employs a deterministic mechanism derived from the dynamics
to simulate the samples according to a target distribution. It uses the concept of canonical ensemble from the statistical
mechanics in conjunction with the Hamiltonian principles to give a physical interpretation of the problem with useful
intuitions. A canonical ensemble is a probability distribution that assigns a finite probability to a particle attaining a
particular state of the system based on the energy content. For a system, where x represents the state and E(x) is the energy
function, the canonical ensemble is defined as
( )
1 E(x)
𝑓 (x) = exp − , (3)
Z T

where T is the temperature of the system and Z is a normalizing constant. If we consider a Hamiltonian system such that
it is entirely determined by its position (𝛽) and momentum (𝛾), then the total energy of the system can be defined in terms
of the Hamiltonian as
H(𝛽, 𝛾) = U(𝛽) + K(𝛾), (4)
where U(𝛽) represents the potential energy of the system, which depends only on the position, and K(𝛾) is kinetic energy,
which depends on the momentum. For this system, the canonical ensemble becomes
( )
1 H(𝛽, 𝛾)
𝑓 (𝛽, 𝛾) = exp − . (5)
Z T

If the potential energy and kinetic energy of a system are defined such that the canonical ensemble represents a target
probability distribution, then the solution of the governing Hamiltonian equations describes the time evolution of the
system. A more detailed description of the properties of this evolution can be found in other studies.35,38,46

2.1 Overview of the HMC method

Developing a deterministic Monte Carlo method using the concept of Hamiltonian dynamics requires constructing a
Hamiltonian system, which is mathematically equivalent to the probability space of interest. To establish such a relation,
BAISTHAKUR AND CHAKRABORTY 5 of 22

the variable (x) is viewed as position 𝛽 of the Hamiltonian system, and an auxiliary momentum variable 𝛾 is introduced
to complete the position–momentum phase space. The potential energy, V(𝛽), of the system is expressed in terms of the
target pdf as a canonical ensemble in the following form:

V(𝛽) ≡ −log(𝜋(𝛽)). (6)

Also, the kinetic energy, K(𝛾), of the system is expressed in terms of the momentum as follows:

𝛾 T M −1 𝛾
K(𝛾) ≡ , (7)
2

where M is a positive-definite symmetric matrix. For simplicity, this matrix is always chosen as a scalar multiple of the
identity matrix. Thus, the Hamiltonian equivalent of the mathematical system becomes

H(𝛽, 𝛾) = V(𝛽) + K(𝛾)

𝛾 T M −1 𝛾 . (8)
= −log(𝜋(𝛽)) +
2

Setting the temperature T = 1, the canonical ensemble for this system in terms of the position vector 𝛽 and the
momentum vector 𝛾 is given by

1 −H(𝛽,𝛾) 1 −V(𝛽) −K(𝛾) 1 𝛾 T M −1 𝛾

𝑓 (x) = 𝜋(𝛽, 𝛾) = e = e e = 𝜋(𝛽) e− 2 , (9)
Z Z Z

where Z is a normalizing constant such that it makes the area under the pdf equal to unity. The above equation shows
that the position vector 𝛽 and the momentum vector 𝛾 are statistically independent of each other. Also, it is evident that
the position vector 𝛽 follows the target probability distribution 𝜋(𝛽) and the momentum vector 𝛾 follows a Gaussian
distribution with zero mean and covariance M. Therefore, if the samples are simulated as per the pdf 𝜋(𝛽, 𝛾), then the
momentum component 𝛾 can be easily projected out to get the samples distributed as per the target probability distribution
𝜋(𝛽).

2.2 Time evolution of the Hamiltonian system

To simulate a random variable distributed as per the target distribution 𝜋(𝛽), the equivalent Hamiltonian system can be
assumed as a conservative system for which the governing Hamiltonian equations are

d𝛽 𝜕H
= , (10a)
dt 𝜕𝛾

d𝛾 𝜕H
=− . (10b)
dt 𝜕𝛽
These equations are solved numerically by assuming a seed value (𝛽(0), 𝛾(0)) for some finite time t using time-stepping
algorithm to get the position and momentum after time t as (𝛽(t), 𝛾(t)). Generally, the leapfrog algorithm is used to solve
the above-mentioned differential equations,35,36 where the position and momentum of particle after some finite time 𝛿t
is given as
𝛿t
𝛽(t + 𝛿t) = 𝛽(t) + 𝛿tM −1 {𝛾(t) − ΔV[𝛽(t)]}, (11)
2
𝛿t
𝛾(t + 𝛿t) = 𝛾(t) − {ΔV[(𝛽(t)] + ΔV[(𝛽(t + 𝛿t)]}. (12)
2
Sample generation by this numerical technique requires evaluation of FE model in each step of the leapfrog algorithm,
which makes it computationally infeasible for higher dimensions. However, when the target pdf follows Standard Gaus-
sian distribution and M becomes an identity matrix (i.e. K(𝛾) = 𝛾T𝛾/2), Hamiltonian for such a system reduces to
6 of 22 BAISTHAKUR AND CHAKRABORTY

H(z, 𝛾) = V(z) + K(𝛾)

𝛾T𝛾
= −log(𝜑(z)) + (13)
2
zT z𝛾 𝛾 T
= + .
2 2
The system in this form has an analytical solution, which is given by38

z(t) = 𝛾in sin t + zin cos t, (14)

𝛾(t) = 𝛾in cos t − zin sin t, (15)

where zin and 𝛾 in denote the initial position and momentum, respectively. This closed-form solution describes the time
evolution of the system and bypasses computationally exhaustive solution of the leapfrog algorithm. For generating the
candidate state, the time parameter t is assigned with an arbitrary value at the beginning of the chain, which is then
modified based on the acceptance ratio of the candidates and the limits of acceptance38 such that if the rate of acceptance
(a) is less than the minimum acceptance rate (amin ), then t is decreased to

t = sin−1 {sin t exp[(a − amin )∕2]}, (16)

whereas if the acceptance ratio exceeds the maximum rate of acceptance (amax ), then t is increased by

t = sin−1 {sin t exp[(a − amax )∕2]}. (17)

However, considering the circular behavior of Equations (16) and (17), it is reasonable to assign an initial value to t such
that t ∈ [−𝜋∕2, 𝜋∕2]. Using the above formulation and the acceptance–rejection criteria, the HMC method can be applied
for Bayesian model updating with standard Gaussian prior, which is explained in the following subsections.

2.3 HMC for model updating

The HMC algorithm is applied for model updating by Cheung and Beck36 where the leapfrog strategy is used to generate
the candidate state. In this approach, generation of a new candidate state requires the evaluation of gradient of the poten-
tial energy function with respect to each uncertain parameter. However, gradient evaluation using numerical techniques
or approximate allied methods is a cumbersome task and the computational cost grows with the dimension as well as
the complexity of the potential energy function. Cheung and Beck36 have used the simultaneous perturbation stochastic
approximation (SPSA) optimization algorithm to estimate the starting point 𝜃 0 closer to the high-probability region of
the posterior pdf, which itself requires large number of evaluations of the posterior pdf, further consuming the compu-
tational resources. Also, the problems of low acceptance rate and multimodal posterior density are often encountered in
model updating due to its inverse nature and the narrow regions with high probability concentration. These problems are
addressed in MCMC-based simulation methods with the help of intermediate pdfs and adaptive proposal distribution,
respectively. However, in HMC-based techniques, the time parameter (t) can be tuned to maintain a consistent explo-
ration of the probability space with a required rate of acceptance. However, as the candidate state is proposed through a
close form solution or a numerical time-stepping method, the liberty of using information from the simulated posterior
is not available. Therefore, this method fails to identify well-separated high-probability regions, leading to a multimodal
posterior density function. These shortcomings limit the application of this method to less complicated mathematical
models having unimodal posterior distribution. To address this issue, a new HMC-based model updating algorithm is
presented in the following section which provides a way to overcome these drawbacks using the closed-form solution as
given in Equations (14) and (15) for simulating the random variables following an adaptive prior-based formulation. It
can be used to update more complicated finite element model having multimodal posterior pdfs.

2.4 Proposed adaptive prior-based HMC algorithm

Generally, the prior pdf in model updating problem is formulated based on the crude knowledge about the nature and
bounds of the uncertain parameters, which are independent of the measurement D. This proposal is updated using both
BAISTHAKUR AND CHAKRABORTY 7 of 22

measurements and initial assumptions through the updating procedure. The posterior values obtained after running the
model updating algorithm give a better insight into the nature of uncertain parameters as compared with the initial
formulation at the beginning of the chain. A Gaussian distribution fitted to these posterior values acts as a better candidate
of prior pdf for updating in the subsequent chains. The proposed model updating procedure using this formulation is
shown in Algorithm 1, and the detailed description along with the formulation of each term is explained in subsections
below.

2.4.1 Formulation of Prior

The prior pdf p(𝜃|C) denotes the probability of the uncertain parameters attaining a particular value based on the previous
knowledge of the problem or user's judgment, independent of the measurement D. While carrying out experimental
studies to understand the actual behavior of the structure, there is a possibility that the recorded data contain certain
amount of error. The prior pdf addresses this error of the recorded data. It induces sufficient inertia into the system such
that the parameter values far from their respective mean are accepted only if it reduces the error term significantly. In
general, two types of prior pdf are proposed in the literature: improper prior and conjugate prior.2 For a particular problem
where the primary emphasis is on the measured data and no initial values of the parameter are to be favored based on
judgments, the prior pdf is considered to be constant. This constant improper prior gets absorbed into the constant K0 in
Equation (1) and the likelihood function alone governs the posterior density function. Integrating this prior pdf over the
8 of 22 BAISTHAKUR AND CHAKRABORTY

parameter space does not give unity, and the solution obtained using this term yields the maximum likelihood estimate
of the uncertain parameters. However, selecting a prior pdf conjugate to the likelihood function results in the posterior
pdf following the same family of distribution as that of the prior pdf. It controls the contribution of initial information
while formulating the posterior. A prior pdf selected for any problem should always be slowly varying so that it does not
introduce significant bias in the acceptance probability of the specific parameter values. The selection of prior pdf for a
particular problem is done by maximizing the information entropy based on the constraints over the parameter space and
the available data. This approach gives a suitable class of probability distribution for the problem. However, it does not
specify the values of the distribution parameter. The importance factor used in the HMC algorithm governs the acceptance
of a candidate set and is calculated as the discrepancy in the value of Hamiltonian between the proposed candidate and
initial state, which can be split in terms of likelihood and prior as

[ ]
I𝑓 = min 1, exp(−ΔH)
[ ]
= min 1, exp(H(qin , pin ) − H(qcand , pcand ))
[ ]
= min 1, exp(V(qin ) + K(pin ) − V(qcand ) − K(pcand ))
[ ]
= min 1, e(V(qin )−V(qcand )) e(K(pin )−K(pcand ))
[ ]
= min 1, e−log(𝜋(qin )+log(𝜋(qcand )) k
[ ] (18)
𝜋(qcand )
= min 1, k
𝜋(qin )
[ ( )]
p(D|𝜃in , C) p(𝜃in |C)
= min 1, × k
p(D|𝜃cand , C) p(𝜃cand |C)
[ ( )]
= min 1, rl × rp × k ,

where the dimensionless terms, rl and rp , represent the ratio of the likelihood and the prior probability of the candidate
state and the initial state, respectively, while the term k is the measure of the discrepancy in the kinetic energy. By the
nature of this problem, the term k in the above equation always remains close to 1, and the acceptance or rejection of a
candidate state is governed by rl and rp . However, other model updating algorithms developed so far directly use the ratio
of posterior probabilities while remaining silent on the individual contribution of likelihood and prior. For an efficient
model updating algorithm, these two dimensionless terms should have a similar scale of variation so that both initial
assumptions and error term have equal contribution while accepting or rejecting the parameter values. For a selected
class of prior based on any independent analysis, the ratio rp is governed by the distribution parameters. The mean value
of the prior pdf can be fixed to the initial values of the parameter. However, assigning a proper value to the coefficient
of variation is a difficult task. Selection of an arbitrary coefficient of variation without any detailed analysis may lead
to biased model updating, favoring either the prior assumptions or the likelihood function. For a selected class of prior
pdf with a fixed mean, the ratio rp depends on the coefficient of variation (COV) such that a large value of it smoothens
the variation of probabilities and the ratio rp . A small value of the COV represents a low uncertainty level and leads to
a narrow pdf with sharp drop of the probabilities away from the mean, which introduces bias in accepting particular
parameter values. This problem of low COV is encountered in the adaptive prior method proposed above, as the COV
decreases with the increase of the number of chains. The smooth variation of the ratio rp even with a low COV is ensured
by using the property of Gaussian distribution in which irrespective of the COV, the parameter values closer to the mean
always have the higher probability than the values away from mean. Also, it is evident from Equation (18) that the ratio
of probabilities (i.e., rp ) is more important than their individual values. Therefore, mean and standard deviation for prior
pdf are obtained by fitting a Gaussian distribution to the sampled parameter values, and then, the COV is adjusted such
that it ensures smooth variation of the ratio rp . This process of adaptive prior is explained in Algorithm 2.

2.4.2 Formulation of likelihood function

The likelihood function, p(D|𝜃, C), ensures the closeness of the model response to the measurement D when the uncertain
parameters attain a set of value 𝜃. It is formulated based on the goodness-of-fit function (also known as objective function),
which acts as the measure of error between the structural response and the model response for a given set of parameter
values 𝜃. Likelihood function governs the contribution of the measurements in the formulation of posterior pdf. When
BAISTHAKUR AND CHAKRABORTY 9 of 22

the model is updated using a set of identified natural frequencies, the objective function is formulated as follows

[ 2 ]
2 2
1 ∑ 𝜔nact − 𝜔nmod (𝜃)
N
Jg (𝜃) = , (19)
N n=1 𝜔2nact

where 𝜔nmod (𝜃) and 𝜔nact are the nth model-predicted natural frequency and the nth identified system natural frequency,
respectively. Similarly, for the problems where both natural frequencies and mode shapes are considered for model
updating, the above-mentioned objective function is modified in the following form:

N [ ]2
1 ∑ 𝜔nact − 𝜔nmod (𝜃) ( )
Jg (𝜃) = + 1 − |𝜙Tnact 𝜙nmod (𝜃)|2 , (20)
N n=1 𝜔nact

where 𝜙nmod (𝜃) and 𝜙nact are the respective mode shapes obtained from the model and the actual structure, respectively.
Assuming Gaussian model for the error distribution, the likelihood function is then constructed as
[ ]
1
p(D|𝜃, C) = exp − Jg (𝜃) . (21)
2

2.4.3 Formulation of posterior pdf

The posterior pdf p(𝜃|D, C) is obtained by updating the parameters 𝜃 using measurement D for a selected class of mathe-
matical model C, while quantifying the relative plausibility of 𝜃 over a set of its values. It is obtained by multiplying the
likelihood function and the prior pdf as follows:
[ ]
1
p(𝜃|D, C) = K0 exp − Jg (𝜃) p0 (𝜃|C), (22)
2

where Jg is the goodness-of-fit function and p0 (𝜃|C) is the prior pdf. To compensate for the effect of the low rate of accep-
tance due to narrow region of high probability concentration in the posterior pdf, a set of intermediate pdfs is proposed
by Beck and Au:22
[ ]
Jg (𝜃)
pi = ci exp − 2 , (23)
2𝜖i
10 of 22 BAISTHAKUR AND CHAKRABORTY

where 𝜖i2 ≈ 2m−i 𝜖 2 and m is the number of intermediate prior pdfs considered in their study. However, as HMC algorithm
can efficiently identify high-probability regions using deterministic candidate states, the adaption scheme in Beck and
Au22 slows down the convergence of the model updating process. To address this issue, an adaption scheme that correlates
the magnitude of the prior probabilities and the likelihood function is proposed in this study where the parameter 𝜖 2 is
evaluated as follows:
−Jg (𝜇i ) k
𝜖i2 = × (24a)
2lnr i
and
p(𝜃|D, C)
r= . (24b)
p(𝜃|C)
In Equation (24a), Jg (𝜇i ) represents the objective function evaluated at mean values of 𝜃 and r represents the domains of
the intermediate pdfs in different chains where i corresponds to the chain number. Since the term r cannot be predicted
accurately in advance, the term k should be tuned such that it maintains the trade-off between the rate of convergence of
the algorithm and the rate of acceptance of the simulated candidates in the chain.

3 NUMERICAL ST UDY

The efficiency of the proposed adaptive prior-based HMC algorithm is demonstrated with the help of three examples.
The first and second examples are validation exercises where synthetic data are used to check the performance of the
proposed algorithm. For this purpose, portal frames whose parameters have well-separated high-probability regions are
considered. Finally, a steel truss bridge is tested for experimental verification. These examples are discussed in detail in
the following subsections.

3.1 Example 01: 2D portal frame for globally identifiable parameters

An analytical 3DOF shear building model of a moment-resisting frame is considered in this example for the identification
of interstory stiffness parameter using simulated noisy frequency data set. The model is assumed to have the bay width of
3.2 m and the storey height of 4 m each. The column cross section has a dimension of 0.2 × 0.25 m with the concrete grade
of M25 while the lumped floor mass is assumed to be 42.5 kN for the ground and the first floor and 40 kN for the second
floor. The natural frequencies of this structure are 1.824, 5.084, and 7.294 Hz. To study the efficiency of the proposed
HMC algorithm for parametric identification, noisy versions of the natural frequencies are generated by adding zero
mean Gaussian white noise with 5% and 10% variance of the original frequencies and the interstory stiffness parameters
are updated using both HMC and standard MCMC method such that the model frequencies match with the simulated
frequencies. The parameterized stiffness matrix for this case in terms of interstory stiffness modifier 𝜃 is shown in Figure 1.
In this analysis, k0 = 27.832 × 105 N/m is the nominal interstory stiffness value obtained by assuming columns with fixed
boundary condition at both ends. For this problem, the prior pdfs of the parameters 𝜃 1 , 𝜃 2 , and 𝜃 3 are assumed to be
independent Gaussian centered at their nominal values, that is, 𝜃 1 = 𝜃 2 = 𝜃 3 = 1. The COV is assumed to be 30% at the
beginning. For HMC algorithm, the time parameter t is assigned with initial value of 𝜋∕3, which is further updated based
on the rate of acceptance of the simulated Markov chain. The goodness-of-fit function, as mentioned in Equation (19), is
used to formulate the likelihood function. The posterior pdf in this case is given by
( )
Jg
p(𝜃|D, C) = c exp − 2 p(𝜃|C). (25)
2𝜖

Here, it may be noted that the success of the model updating algorithm depends on the choice of the adaptive scheme.
For this problem, the effects of different values of 𝜖 2 on the algorithm are studied, and it is observed that the
√most efficient
adaptive scheme is when the sequence of probability√concentration of the intermediate pdfs decrease by 1∕ 2 of its earlier
value, that is, pi+1 ∼ p2i , for which r is equal to 1∕ 2. Also, the parameter k = 3 is found to maintain an average 60%
rate of acceptance of the candidates. The average values of the ratio of likelihood and prior probability of the proposed
candidate state to the initial candidate for the first four chains are presented in Table 1. This analysis shows that in
most of the cases, the average values of rl and rp for the proposed algorithm are more than 1, which proves that the
proposed deterministic candidate generation scheme has more candidates in higher probability region than the local
random walk-based candidate generation scheme used in standard MCMC algorithm. Results of the parameter estimation
BAISTHAKUR AND CHAKRABORTY 11 of 22

FIGURE 1 Details of portal frame used in

Example 01

Chain 5% error 10% error TABLE 1 Average ratio of the likelihood and the
rl rp rl rp prior probabilities of a candidate state to initial state
HMC MH HMC MH HMC MH HMC MH
1 0.9923 0.9691 1.1688 0.9359 0.9999 0.9698 1.2356 0.8973
2 1.0385 0.9854 1.1316 1.0108 1.0040 0.9823 1.1034 1.0393
3 1.0135 0.9755 1.0793 0.9735 1.0085 0.9684 1.1306 0.9564
4 1.0167 0.9770 1.0166 0.9721 0.9920 0.9865 1.0751 0.9774
Abbreviations: HMC, Hamiltonian Monte Carlo; MH, Metropolis–Hastings.

Parameter Initial value Noise level Updated value SD COV (%) TABLE 2 Parameter identification
𝜃HMC
∗
𝜃MH
∗
𝜎 HMC 𝜎 MH HMC MH results for globally identifiable case
𝜃1 1 0.9103 0.9098 0.0093 0.0432 0.93 4.61 (i.e., Example 01)
𝜃2 1 5% 0.9909 0.9838 0.0124 0.0452 1.25 4.67
𝜃3 1 0.9560 0.9658 0.0114 0.0445 1.14 4.45
𝜃1 1 0.6954 0.7095 0.0146 0.0121 2.17 1.69
𝜃2 1 10% 1.1839 1.1588 0.0279 0.0299 2.26 2.69
𝜃3 1 0.9979 1.0140 0.0320 0.0277 3.24 2.70
Abbreviations: HMC, Hamiltonian Monte Carlo; MH, Metropolis–Hastings.

using the proposed HMC algorithm and the standard MCMC algorithm for different noise levels are summarized in
Table 2. Using these updated parameters, natural frequencies are identified, which are reported in Table 3. A close match
between the original frequencies and the updated frequencies is observed from this table. In this context, subscripts in, sim
and upd correspond to the frequencies observed from the initial, simulated, and updated models, respectively. Figures 2
and 3 show the parameters simulated by the adaptive HMC and the standard MCMC algorithms, where the proposed
algorithm has converged in less number of chains (i.e., 4 and 6 for 5% and 10% noise level, respectively) as compared with
the standard MCMC algorithm (i.e., 10 for two different noise levels). These figures show the efficiency of the proposed
algorithm in rapidly identifying the region of high probability concentration. The marginal posterior densities obtained
from two different methods are presented in Figures 4 and 5. These results also show that the proposed HMC-based
algorithm outperforms the standard MCMC algorithm in identifying the narrow region of interest of the posterior pdf.
Further, to compare the performance of the proposed approach in terms of computational cost, the numerical model
presented above is updated using the proposed approach, the standard MCMC method and the conventional HMCMC
method. The candidate states in conventional HMCMC method are generated using the leapfrog algorithm presented in
Equations (11)and 12. The time required for convergence by various methods are presented in Table 4, and their posterior
estimates are compared in Figure 6. These results conclusively demonstrate that the proposed approach is more efficient
in terms of computational efficiency and in identifying the narrow region of importance of posterior distribution as com-
pared with both standard MCMC and HMCMC method using leapfrog algorithm. In this context, the process of finding
12 of 22 BAISTHAKUR AND CHAKRABORTY

TABLE 3 Updated natural frequencies for globally Mode 𝜔in Noise level 𝜔n
exp
𝜔n
upd
n
identifiable case (i.e., Example 01) HMC MH
1 1.8246 1.7699 (3.00) 1.7703 (0.02) 1.7688 (0.06)
2 5.0844 5% 4.9429 (2.78) 4.9413 (0.03) 4.9521 (0.18)
3 7.2939 7.1849 (1.49) 7.1833 (0.02) 7.1824 (0.03)
1 1.8246 1.6697 (3.60) 1.6690 (0.03) 1.6774 (0.46)
2 5.0844 10% 4.8744 (4.13) 4.8770 (0.05) 4.9023 (0.57)
3 7.2939 7.5566 (8.49) 7.5344 (0.30) 7.5130 (0.57)
Note. Percentage error is mentioned in parenthesis. Abbreviations: HMC, Hamiltonian
Monte Carlo; MH, Metropolis–Hastings.

FIGURE 2 Samples generated

using Markov chain Monte Carlo
and HMC method for 5% noise
case. HMC, Hamiltonian Monte
Carlo; MH, Metropolis–Hastings

FIGURE 3 Samples
generated using Markov chain
Monte Carlo and HMC method
for 10% noise case. HMC,
Hamiltonian Monte Carlo; MH,
Metropolis–Hastings

out the exact posterior distribution for a model updating problem is a cumbersome task and is not very widely attempted
in literature because of the complexity of multidimensional integral involved. Moreover, direct application of brute force
Monte Carlo simulation is not an efficient technique as also mentioned by Beck and Au.22 Thus, the computation of exact
distribution is not attempted in this work. However, the results show that the posterior distribution obtained by the pro-
posed approach can be accepted as the close representation of the exact posterior as it closely represents the simulated
data and respective model properties. Further, a comparative study is carried out between the conventional HMCMC
method using leapfrog algorithm and the proposed approach using closed-form solution to demonstrate the efficiency
of the proposed approach in handling high-dimensional parameter space. For this purpose, the DOF is increased from 2
to 10 where the third story is repeatedly added in every successive case. Here, the time required by different methods to
update the numerical model of a portal frame with increasing number of DOF is plotted in Figure 7a. From this figure, it
BAISTHAKUR AND CHAKRABORTY 13 of 22

FIGURE 4 Marginal posterior

distribution of 𝜃 1 , 𝜃 2 , and 𝜃 3 using the
proposed Hamiltonian Monte Carlo
algorithm for 5% noise case

FIGURE 5 Marginal posterior

distribution of 𝜃 1 , 𝜃 2 , and 𝜃 3 using the
standard Markov chain Monte Carlo
algorithm for 5% noise case

Method Computational time (s) TABLE 4 Computational time required by various methods in Example
Proposed HMC Algorithm 08.08 01
Standard MCMC 13.62
HMCMC (with leapfrog algo.) 18.65
Abbreviations: HMC, Hamiltonian Monte Carlo; HMCMC, Hamilto-
nian Markov chain Monte Carlo; MH, Metropolis–Hastings.

is evident that the computational cost required in conventional HMCMC method rapidly increases with the increase in
dimension of problem. This increase in computational time for conventional HMCMC is due to the requirement of large
number of model evaluations with the increase in dimension (i.e., parameter space). The candidate generation scheme
presented in Equations (11) and (12) involving numerical differentiation requires twice the number of model evalua-
tions for generating each set of candidate state, whereas the proposed closed-form candidate generation technique (i.e.,
Equations 14 and 15) is independent of the dimension of the parameter space. Therefore, the proposed approach can han-
dle higher dimensional parameter space more efficiently. The percentage of time spent on generating candidate states
with increasing dimension for conventional HMCMC approach is presented in Figure 7b.

3.2 Example 02: 2D portal frame for locally identifiable parameters

In reality, inverse problem of model updating often encounters the possibility of multiple solutions, which is referred as
locally identifiable models. In this validation exercise, one such problem is considered to study the accuracy of the pro-
posed method. A 2DOF portal frame2 with locally identifiable solutions is considered here to demonstrate the capability
of the proposed algorithm to identify well-separated regions of probability concentration. It has bay width of 5 m and
story height of 2.5 m. The beams and columns have cross-sectional area of 0.01 and 0.02 m with the moment of inertia of
6.0 × 10−4 m4 and 1.5 × 10−4 m4 , respectively. The frame is made up of a material having Young's modulus of 200 Gpa and
14 of 22 BAISTHAKUR AND CHAKRABORTY

FIGURE 6 Posterior distribution

comparison for 𝜃 1 , 𝜃 2 , and 𝜃 3 using various
methods

FIGURE 7 Time required by

conventional HMCMC with
leapfrog algorithm and proposed
HMC with closed-form solution: (a)
Model updating and (b) candidate
generation. HMCMC, Hamiltonian
Markov chain Monte Carlo

mass density of 7, 800kg∕m3 . Furthermore, the floor mass of 11.17×103 kg is assumed to be lumped at each level. Stiffness
matrix in terms of parameter 𝜃 is formulated in similar way as described in the previous problem where k0 = 46.08 MN/m.
Again, a noisy set of frequencies are generated as 𝜔1 = 5.5 and 𝜔2 = 14.9 Hz for system identification. There exist two
maxima in the posterior density at the parameter values (0.73, 0.84) and (1.6, 0.38). The standard MCMC algorithm and
proposed HMC algorithm are applied to identify these two distinct maxima. Here, independent Gaussian distributions
with mean 0.9 and 1.2 and unit standard deviation are considered as the prior pdfs. In this analysis, 𝜖i2 = 1∕2i−1 is used
√
for the standard MCMC algorithm, whereas an adaption scheme with r = 1∕ 3 and k = 2 is used for the proposed
algorithm. The results obtained from these two methods are presented in Figures 8 and 9. Figure 8 shows the perfor-
mance of the standard MCMC algorithm where 10 iterations are required for convergence as opposed to 4 iterations by
the proposed HMC algorithm using closed-form solution as shown in Figure 9. This clarifies the advantage of the pro-
posed model updating scheme, which clearly shows 60% reduction of computational cost. Besides number of iterations,
the volume of sample generations in each iterations is also less as shown in the previous example. However, these results
are not produced here to avoid repetition. The estimated parameters in Table 5 show higher uncertainties in the posterior
upd
density, which is due to the presence of multiple solutions. However, the updated natural frequencies (𝜔n ) and simu-
lated frequencies (𝜔sim
n ) are in close agreement with each other as shown in Table 6. Since the maximum at the location
(0.73, 0.84) has higher posterior probability than the maximum at the location (1.6,0.38), the first maximum is selected
as the updated value of uncertain parameters. Overall, these results show that the proposed method can distinctly iden-
tify the well-separated regions of probability concentration in less number of chains as compared with standard MCMC
algorithm and can solve the locally identifiable problems more efficiently.

3.3 Example 03: Model updating of truss bridge

Finally, a 55 m long steel truss bridge over the river Singi-Chu near Pasakha in Bhutan as shown in Figure 10 is con-
sidered in this study. The structural details of this truss bridge is provided in the Appendix. This bridge is designed and
constructed as per Indian standard guidelines in the recent past and is yet to be inaugurated due to severe underper-
formance during initial testing immediately after the erection. This underperformance is envisaged due to a flash flood
during the construction of the superstructure, when a sudden surge of water from the uphill caused misalignment of the
joints, leading to inaccurate fabrication of the members. Figure 11a shows the positions of these dislocated joints. It may
BAISTHAKUR AND CHAKRABORTY 15 of 22

FIGURE 8 Scatter plot and

marginal posterior distribution
of samples generated using
standard Hamiltonian Markov
chain Monte Carlo with
Metropolis–Hastings algorithm
for locally identifiable case

FIGURE 9 Scatter plot and

marginal posterior distribution
of samples generated using the
proposed Hamiltonian Monte
Carlo method for locally
identifiable case

Parameter Initial value Updated value SD COV TABLE 5 Identification parameters for locally
𝜃HMC
∗
𝜃MH
∗
𝜎 HMC 𝜎 MH HMC MH identifiable case (i.e., Example 02)
𝜃1 1 0.7304 0.7369 0.4405 0.4606 44.05 39.33
𝜃2 1 0.8507 0.8368 0.2408 0.2538 24.08 41.37
Abbreviations: HMC, Hamiltonian Monte Carlo; MH, Metropolis–Hastings.

be noted that the right half of the bridge is affected as the scaffolding beneath this section over the gorge has suffered the
thrust of the incoming water. The site survey reveals that maximum distortion of the joints at the deck level is 0.45% and
0.89% the roof level of the span as shown in Figure 11a.
The proposed adaptive HMC-based Bayesian updating strategy is used to study the in situ condition of the members
and its impact on the overall performance of the structure. To assess the health of the bridge in its as-built condition, the
first static load test is carried out to check the deflection due to standard loading as per the IRC-SP-51 guidelines. The
results show higher deflection than it is expected as per structural drawings. This anomaly advocated for dynamic tests for
detailed assessment of the health of this structure. Therefore, an ambient vibration test is conducted using a 22T vehicular
load traveling at a speed of 20 km/h. The test load is fixed well below the design load as per the Indian Road Congress
(IRC) guidelines. In this experiment, three sets of accelerometers are used on either sides of the bridge deck as marked
in Figure 11 (i.e., along the grid lines 4-4, 6-6, and 8-8). Here, it may be noted that sensors in 6-6 grid lines correspond
to the maximum deformation in the first mode while those in 4-4 and 8-8 correspond to the same in the second mode.
Although these locations can be optimized for more efficient measurements, that exercise itself demands the updated
finite element model of the bridge. These are wireless G-Link sensors from Lord Corporation that are connected to the
WSDA Base Station from the same manufacturer as shown in Figure 10. The test vehicle is driven at a constant speed, and
the acceleration responses along horizontal and vertical directions are measured using these sensors. The free vibrations
16 of 22 BAISTHAKUR AND CHAKRABORTY

TABLE 6 Updated natural frequencies for locally identifiable Mode 𝜔in 𝜔sim 𝜔n
upd
n n
case (i.e., Example 02) HMC MH
1 06.3177 05.5000 (12.94) 05.5062 (0.11) 05.5136 (0.11)
2 16.5401 14.9000 (09.92) 14.9600 (0.40) 14.8824 (0.24)
Note. Percentage error is mentioned in parentheses. Abbreviations: HMC, Hamil-
tonian Monte Carlo; MH, Metropolis–Hastings.

FIGURE 10 Field test details: (a)

Pasakha bridge, (b) test truck, (c)
G-Link wireless accelerometer, and (d)
base station

FIGURE 11 Distorted joint locations

and ambient vibration response: (a) joint
locations at top and bottom, (b) vertical
response, and (c) horizontal response

records in these two directions are used further to identify the in situ natural frequencies of this bridge. Figure 11 shows
the locations of accelerometers and a sample response time history at the middle of the bridge. These data are analyzed
using wavelet-based time-frequency analysis to identify the in situ natural frequencies of this bridge. Figure 12a,b shows
the scalograms obtained from signal processing using Morlet wavelet with central frequency of 5 rad/s. The details of
this modal identification following the procedure outlined by Mahato and Chakraborty47 are not discussed here as that
is not the theme of this study. Reader may refer to Mahato and Chakraborty47 for the detailed description of this signal
processing for modal identification. The identified first three natural frequencies are shown in Table 7.
Here, it may be noted that the scalogram in Figure 12b clearly shows the existence of dominant frequencies as marked
in them. Also, these two scalograms indicate that the frequency close to 4.1 Hz (i.e., second mode of torsional vibration)
is active. This is also observed during the experiment as the bridge vibrates significantly in the side-wise (rolling) direc-
tion. Since the bridge deck is supported by two 2D plane truss on either sides, it is expected to predominantly vibrate in
the vertical plane as the truss members are allowed to take only axial loads. However, significant vibration in the trans-
verse direction indicates the deviation of the bridge behavior from its design. In this context, the finite element model of
the bridge is developed as per the structural drawing using SAP2000 where the members are modeled as truss element,
longitudinal and cross beams at the deck level are modeled using beam element, and the deck is modeled using plate ele-
ment. The geometric and material properties are used as mentioned in the structural drawings. Additional thickness of
the road surface layer and the secondary loading from footpath and railings are also included in the baseline finite ele-
ment model. The frequencies reported in Table 7 clearly indicate that the actual bridge (i.e., as constructed) has higher
natural frequencies as compared with its design values, implying that the fabricated bridge is stiffer than it is expected as
per design. This increase in stiffness of the bridge is most likely caused by the misalignment of the truss joints. Due to this
misalignment, truss members (which are designed to carry axial forces only) are subjected to end moments, and hence,
the joints experience rigidity. This, in turn, implies that the truss members experience partial fixity at its ends. To quantify
BAISTHAKUR AND CHAKRABORTY 17 of 22

FIGURE 12 Frequency identification

results and dominant mode shapes: (a)
scalogram of vertical response, (b)
scalogram of horizontal response, (c)
first fundamental mode, and (d) second
fundamental mode

Mode number 𝜔nd (Hz) 𝜔ni (Hz) 𝜔nupd (Hz) TABLE 7 Identified and updated natural frequencies of Pasakha bridge
1 2.396 2.733 (14.06) 2.612 (04.43)
2 4.349 4.122 (05.22) 4.293 (04.15)
3 4.940 5.061 (02.45) 5.033 (00.55)

Note. Subscripts d and i stand for ‘designed’ and ‘identified’. Percentage

error is mentioned in parentheses.

this partial effectiveness of the truss joints, the initial bridge model is modified by restraining the member ends. These
joints restraints introduce bending stiffness (i.e., effectively rotational springs at the ends) by the beam elements instead of
the truss elements in actual drawing. However, fully restrained joints make the bridge super stiff whose natural frequen-
cies are far from their designed values. Here, it may be noted that the stiffness offered by a beam element is a function of
three parameters, namely, boundary/joint conditions, material properties, and the sectional properties of the constituent
members. Thus, to replicate the actual scenario (i.e., partial fixity at joints), strength parameters are adjusted so that the
frequencies from the model match closely with the observed values. Hence, the joint stiffness is updated by the proposed
algorithm using the identified natural frequencies. For this purpose, a sensitivity analysis is first carried out in the light
of joint distortion to study the effects of misalignment on the performance of the bridge with major emphasis on the first
two modal frequencies. From this analysis, four joints in the right half of the bridge (i.e., 7-7, 8-8, 9-9, and 10-10) as shown
in Figure 11a are used for updating the bridge model. These joints are the most effective compared with central joint as
they are affected most during fabrication. Also, the rolling vibration (i.e., second mode as shown in Figure 12d) observed
during field experiment clearly supports this claim as the joints marked above are prone to maximum displacement dur-
ing torsional mode of vibration. The joint stiffness at these locations is then updated by proportionately modifying the
Young's modulus of all the members connected to a particular joint. In this context, depending on the end condition (i.e.,
fixed or pinned), the stiffness contributed by member connected to a joint is expressed in terms of EI L3
or AE
L
, respectively.
Therefore, the fixity of truss joint is quantified in terms of a dimensionless parameter 𝜃, which modifies Young's modu-
lus of all the members connected to that joint such that the value of 𝜃 less than 1 indicates partial fixity of the joint. In
this way, updating the Young's modulus will directly update the stiffness contribution of the member as this is the only
remaining parameter that controls joint strength apart from end fixity. Thus, independent Gaussian density functions
centered at the initial value of 𝜃 i = 1; i = 1, 2, 3, 4 are considered as the prior pdfs. Here, it may be noted that as 𝜃 repre-
sents the partial effectiveness of the truss joint, its value is always less than 1. Therefore, the lower and upper bounds for
these parameters are assumed to be 0 and 1, respectively. The value of COV is uniformly increased from 0.1 to 1.0, and
its effect on the ratio rl and rp are shown in Figure 13. From this figure, it is clear that for small values of COV, rl and rp
can vary over a wider range. Thus, selection of COV in the lower range may lead to biased model updating, favoring the
18 of 22 BAISTHAKUR AND CHAKRABORTY

FIGURE 13 Variation of
scale rl and rp with COV

FIGURE 14 Contribution of likelihood and

prior in formulation of posterior density for
different values of COV

TABLE 8 Identified parameters of Pasakha bridge Parameter Initial value Updated value SD COV(%)
𝜃1 1 0.5385 0.0690 10.70
𝜃2 1 0.5400 0.0706 12.35
𝜃3 1 0.6333 0.0682 09.39
𝜃4 1 0.7394 0.0528 07.95

prior assumptions over the likelihood function. Figure 14 shows the contribution of the likelihood function and prior pdf
in formulating the posterior pdf for different values of COV. From this analysis, the value of COV is fixed at 1.0 so that
there is significant contribution from both likelihood function and prior pdf in the formulation of the posterior density
function. As the maximum deviation of the natural frequency is found in the first mode in this problem, the likelihood
function is formulated using weighted objective function where 60% weight is assigned to the first frequency and 30% and
10% weight
√ is used for second and third natural frequency, respectively. The adaption scheme is constructed by assigning
r = 1∕ 3 and k = 1. The results of this parameter updating process are summarized in Tables 7 and 8. Table 7 shows
that the updated model offers frequencies that closely match with the observed/identified frequencies with error below
5%. Figure 15 shows the variation of marginal posterior pdfs in different iterations. These results demonstrate that the
proposed HMC algorithm rapidly identifies the important region of probability concentration, establishing the degree of
compliance between the FE model and the actual structure. However, the marginal of uncertain parameters for actual
structure has higher COV as compared with the previously addressed numerical models, which can be attributed to the
higher uncertainties involved in the modeling and measurements in this example.
These updated values imply that the selected joints provide 74.41%, 64.30%, 53.64%, and 60.04% of the respective mem-
ber stiffness with fixity at both ends as opposed to the pin joint assumption in the design of fully functional truss members.
BAISTHAKUR AND CHAKRABORTY 19 of 22

FIGURE 15 Marginal Posterior

distribution of four uncertain joint
parameters 𝜃 of Pasakha bridge

As the members connected to these joints are subjected to end moments due to improper connections, it is reflected in
the stiffer superstructure. Since the gusset plates are not designed to carry end moments, these joints possess a potential
threat of joint failure at these locations due to repeated loading during the operational life span of the bridge. Using this
updated model, the load carrying capacity of the bridge is verified and it is envisaged to carry one third of its design load.
As the bridge rating is beyond the scope of this study, these details are omitted in this paper.

4 CO N C LU S I O N

The new adaptive prior-based HMC algorithm for Bayesian model updating is proposed in this paper, which shows
considerable improvements over the standard MCMC algorithm in exploring the narrow regions of high probability con-
centration of the posterior pdf. The proposed algorithm is successfully implemented for updating simulated cases of both
globally and locally identifiable numerical models and the actual finite element model of a truss bridge. Together, these
results clearly show the efficiency of this method to handle modeling errors while exploring well-separated regions of
interest. The deterministic candidate generation scheme implemented in this paper is found to outperform the conven-
tional random walk-based method to simulate the candidates in higher probability region. Also, the adaptive scheme
proposed for generating intermediate pdfs considers the magnitude of prior probabilities and the likelihood function,
which acts as a better candidate than the earlier proposals. The bifurcation of the acceptance ratio into likelihood and
prior offers better insight into the independent contribution of these components and can avoid potential bias towards
one of them. Also, the guidelines presented in this work to select proper distribution parameters based on the ratio of like-
lihood and prior probabilities help to standardize the selection of proper values for the prior distribution, which is often
left to the user's judgment. The field problem considered in this study presents a particularly interesting case where the
natural frequencies of the structure are found to be higher than its design values. The joint stiffness updating methodol-
ogy applied to tackle this issue acts as a guideline to address the commonly encountered problems of joint rigidity due to
on-site fabrication difficulties during the construction of steel bridges. In general, it can be concluded that the proposed
algorithm acts as a better solution for model updating than the conventional approach using sensitivity analysis.

ORCID
Arunasis Chakraborty https://orcid.org/0000-0002-1678-3068

REFERENCES
1. Papadimitriou C, Beck JL, Katafygiotis LS. Updating robust reliability using structural test data. Probabilistic Eng Mech. 200104;16:103-113.
2. Yuen K-V. Bayesian methods for structural dynamics and civil engineering. Singapore: John Wiley & Sons; 2010.
3. Behmanesh I, Moaveni B. Probabilistic identification of simulated damage on the Dowling Hall Footbridge through Bayesian finite element
model updating. Struct Control Health Monit. 2015;22(3):463-483.
4. Mustafa S, Debnath N, Dutta A. Bayesian probabilistic approach for model updating and damage detection for a large truss bridge. Int J
Steel Struct. 2015;15(2):473-485.
20 of 22 BAISTHAKUR AND CHAKRABORTY

5. Vanik MW, Beck JL, Au S. Bayesian probabilistic approach to structural health monitoring. J Eng Mech. 2000;126(7):738-745.
6. Sohn H, Law KH. Application of load-dependent Ritz vectors to Bayesian probabilistic damage detection. Probab Eng Mech.
2000;15(2):139-153.
7. Hemez FM, Doebling SW. Review and assessment of model updating for non-linear, transient dynamics. Mech Syst Signal Process.
2001;15(1):45-74.
8. Simoen E, Moaveni B, Conte JP, Lombaert G. Uncertainty quantification in the assessment of progressive damage in a 7-story full-scale
building slice. J Eng Mech. 2013;139(12):1818-1830.
9. Mallapur S, Platz R. Quantification and evaluation of uncertainty in the mathematical modelling of a suspension strut using Bayesian
model validation approach. Model validation and uncertainty quantification, Vol. 3. Cham, Switzerland: Springer; 2017:113-124.
10. Zhang R, Mahadevan S. Model uncertainty and Bayesian updating in reliability-based inspection. Struct Safety. 2000;22(2):145-160.
11. Kalantarnia M, Khan F, Hawboldt K. Dynamic risk assessment using failure assessment and Bayesian theory. J Loss Prevention Process
Indust. 2009;22(5):600-606.
12. Zimmerman DC, Yap K, Hasselman T. Evolutionary approach for model refinement. Mech Syst Sig Process. 1999;13(4):609-625.
13. Teughels A, De Roeck G. Structural damage identification of the highway bridge z24 by fe model updating. J Sound Vibr.
2004;278(3):589-610.
14. Bakir PG, Reynders E, De Roeck G. Sensitivity-based finite element model updating using constrained optimization with a trust region
algorithm. J Sound Vibr. 2007;305(1-2):211-225.
15. Ren W-X, Chen H-B. Finite element model updating in structural dynamics by using the response surface method. Eng Struct.
2010;32(8):2455-2465.
16. Jaishi B, Ren W-X. Finite element model updating based on eigenvalue and strain energy residuals using multiobjective optimisation
technique. Mech Syst Sig Process. 2007;21(5):2295-2317.
17. Katafygiotis LS, Papadimitriou C, Lam H-F. A probabilistic approach to structural model updating. Soil Dyn Earthq Eng.
1998;17(7-8):495-507.
18. Beck JL, Katafygiotis LS. Updating models and their uncertainties. I: Bayesian statistical framework. J Eng Mech. 1998;124(4):455-461.
19. Katafygiotis LS, Mickelborough N, Yuen K-V. Statistical identification of structural modal parameters using ambient response data. In:
Proceedings 3rd International Conference on Computational Stochastic Mechanics. Santorini (Thera), Greece; 1998:93-102.
20. Sohn H, Dzwonczyk M, Straser EG, Law KH, Meng TH-Y, Kiremidjian AS. Adaptive modeling of environmental effects in modal param-
eters for damage detection in civil structures. Smart structures and materials 1998: Smart systems for bridges, structures, and highways,
Vol. 3325: International Society for Optics and Photonics; 1998:127-139.
21. Behmanesh I, Moaveni B, Lombaert G, Papadimitriou C. Hierarchical Bayesian model updating for structural identification. Mech Syst
Sig Process. 2015;64:360-376.
22. Beck JL, Au S-K. Bayesian updating of structural models and reliability using Markov chain Monte Carlo simulation. J Eng Mech.
2002;128(4):380-391.
23. Chib S, Greenberg E. Understanding the Metropolis–Hastings algorithm. Am Stat. 1995;49(4):327-335.
24. Kuczera G, Parent E. Monte Carlo assessment of parameter uncertainty in conceptual catchment models: The Metropolis algorithm. J
Hydrology. 1998;211(1-4):69-85.
25. Bates BC, Campbell EP. A Markov chain Monte Carlo scheme for parameter estimation and inference in conceptual rainfall-runoff
modeling. Water Resources Res. 2001;37(4):937-947.
26. Haario H, Saksman E, Tamminen J, et al. An adaptive metropolis algorithm. Bernoulli. 2001;7(2):223-242.
27. Andrieu C, Thoms J. A tutorial on adaptive MCMC. Stat Comput. 2008;18(4):343-373.
28. Ching J, Chen Y-C. Transitional Markov chain Monte Carlo method for Bayesian model updating, model class selection, and model
averaging. J Eng Mech. 2007;133(7):816-832.
29. Yuen K-V, Beck JL, Au SK. Structural damage detection and assessment by adaptive markov chain monte carlo simulation. Struct Control
Health Monit. 2004;11(4):327-347.
30. Wan H-P, Ren W-X. Stochastic model updating utilizing Bayesian approach and Gaussian process model. Mech Syst Sig Process.
2016;70:245-268.
31. Qian F, Zheng W. An evolutionary nested sampling algorithm for Bayesian model updating and model selection using modal measure-
ment. Eng Struct. 2017;140:298-307.
32. Cheung SH, Bansal S. A new Gibbs sampling based algorithm for Bayesian model updating with incomplete complex modal data. Mech
Syst Signal Process. 2017;92:156-172.
33. Lam H-F, Yang J, Au S-K. Bayesian model updating of a coupled-slab system using field test data utilizing an enhanced markov chain
monte carlo simulation algorithm. Eng Struct. 2015;102:144-155.
34. Rocchetta R, Broggi M, Huchet Q, Patelli E. On-line Bayesian model updating for structural health monitoring. Mech Syst Signal Process.
2018;103:174-195.
35. Duane S, Kennedy AD, Pendleton BJ, Roweth D. Hybrid Monte Carlo. Phys Lett B. 1987;195(2):216-222.
36. Cheung SH, Beck JL. Bayesian model updating using hybrid Monte Carlo simulation with application to structural dynamic models with
many uncertain parameters. J Eng Mech. 2009;135(4):243-255.
37. Boulkaibet I, Mthembu L, Marwala T, Friswell M, Adhikari S. Finite element model updating using the shadow hybrid Monte Carlo
technique. Mech Syst Sig Process. 2015;52:115-132.
BAISTHAKUR AND CHAKRABORTY 21 of 22

38. Wang Z, Broccardo M, Song J. Hamiltonian Monte Carlo methods for subset simulation in reliability analysis. Struct Safety. 2019;76:51-67.
39. Tasaka S, Shinozuka M, Chaudhuri S, Na U. Bayesian inference for prediction of carbonation depth of concrete using MCMC. Mem Akashi
Tech Coll. 2009;52:45-50.
40. Beck JL. Bayesian system identification based on probability logic. Struct Control Health Monit. 2010;17(7):825-847.
41. Prajapat K, Ray-Chaudhuri S. Posterior resolution and structural modification for parameter determination in Bayesian model updating.
Int J Cybern Info. 2016;5(1):193-202.
42. Straub D. Stochastic modeling of deterioration processes through dynamic Bayesian networks. J Eng Mech. 2009;135(10):1089-1099.
43. Ma Y, Zhang J, Wang L, Liu Y. Probabilistic prediction with Bayesian updating for strength degradation of RC bridge beams. Struct Safety.
2013;44:102-109.
44. Prajapat K, Ray-Chaudhuri S. Detection of multiple damages employing best achievable eigenvectors under Bayesian inference. J Sound
Vibr. 2018;422:237-263.
45. Lam H, Peng H, Au S. Development of a practical algorithm for Bayesian model updating of a coupled slab system utilizing field test data.
Eng Struct. 2014;79:182-194.
46. Neal RM. MCMC using hamiltonian dynamics. Handbook of Markov chain Monte Carlo. Boca Raton, London, New York: CRC Press;
2011;2:2.
47. Mahato S, Chakraborty A. Sequential clustering of synchrosqueezed wavelet transform coefficients for efficient modal identification. J
Civil Struct Health Monit. 2019;9(2):271-291.

How to cite this article: Baisthakur S, Chakraborty A. Modified Hamiltonian Monte-Carlo Based
Bayesian Finite Element Model Updating of Steel Truss Bridge. Struct Control Health Monit. 2020;27:e2556.
https://doi.org/10.1002/stc.2556

APPENDIX A

The members properties used to model the truss bridge structure addressed in this study in FE package SAP2000 are
provided in Table A1, whereas the member locations are shown in Figure A1

Section name Area(m2 ) Section name Area(m2 ) TABLE A1 Cross section details of Bridge members
B1 0.0068 B2 0.0091
BXB1 0.005 DB 0.0188
S1 0.047 S2 0.037
S3 0.0292 S4 0.0314
S5 0.0394 S6 0.0437
S7 0.041 S8 0.021
S9 0.0198 S10 0.012
S11 0.0068 S12 0.009
S13 0.0154 S14 0.0125
S15 0.0045 S16 0.0045
S17 0.0038 S18 0.0086
22 of 22 BAISTHAKUR AND CHAKRABORTY

FIGURE A1 Member locations for the bridge

addressed in the study: (a) elevation view @ Y =
0/12.45 m,(b) plan view @ Z = 0 m, (c) plan
view @ Z = 6.4 m