Professional Documents
Culture Documents
Inference
Mayank Mishra
Graduate Student, Department of Geotechnical and Structural Engineering
Università degli Studi della Basilicata, Potenza, Italy
Chavon Grande
Structural Engineering Contractor & Project Engineer
CH2M HILL, Portland, Oregan, USA
ABSTRACT: The main objective of this paper is to combine test data gathered from non-destructive testing
(NDT) and fuse it using Bayesian inference. The Bayesian updating technique is applied to data obtained from
Ferroscan testing which is used to estimate the depth of concrete cover (cc) left. The tests were done while
performing the seismic verification of one of the university buildings at university of Padua, Italy. In monuments
and heritage sites its not always feasible to take out samples from the core for laboratory testing as the art work,
graffiti, paintings etc. needs to be preserved, so NDT is always a better option for these class of buildings.
The NDT test data has many uncertainties present in it and to deduce meaningful information out of it is a
challenging task which is addressed in this paper. Bayesian approach is more rational as it gives more realistic
value as compared with other approaches which result in a more conservative estimate. The research will help
engineering community working in inspection field to choose parameters based on a data from limited number of
NDT data tests. Vague information about the structure can be supplemented with the help of experimental data to
arrive to a posterior distribution of concrete cover and hence to estimate 95% confidence interval for parameter
of interest (clear concrete cover). A Matlab toolbox was also developed (Ramos et al. 2015) to combine NDT
test data without having obligation for practitioners to understand the complex mathematics involved behind
the Bayesian black box . A posterior probability density function is derived in this case for concrete cover (cc)
taking into account the NDT test data from Ferroscan testing. This paper only uses data from Ferroscan test but
the same methodology can be adapted to different data sets.
Yes No
Level 3 is addressed in this paper. After collecting Figure 4: Flowchart in the updating process for calculating con-
data from Ferroscan test, it was analyzed and carried crete cover
forward for the data fusion process. The first thing
suggested is that the type of data should be same i.e Choice of prior can vary from very accurate prior
if the data from source is telling some other parame- with less standard deviation to even no knowledge
ter then it must be converted into equivalent data by which is non-informative Jeffrey’s prior. It is reason-
developing some regression relationship or by using able to assume prior as a normal distribution (for
some empirical formulas/conversion charts (level 2 mathematical considerations) for modelling of many
fusion). Ang & Tang (1984) suggested a three step mechanical parameters LNEC Report (1983). When
procedure. The initial prior is obtained by combining the posterior and the prior have the same paramet-
the indirect test data with prior pdf to obtain the pos- ric form is called conjugacy. For example, the Gaus-
terior PDF. The posterior PDF is considered as a prior sian family is conjugate to itself (or self-conjugate)
for the second set of data and another updated poste- with respect to a Gaussian likelihood function. If the
rior is obtained. This posterior got from two indirect likelihood function is Gaussian, choosing a Gaussian
test data is considered as prior when we combine this prior over the mean will ensure that the posterior dis-
with the direct test data to obtain the final posterior tribution is also Gaussian. A conjugate prior is an
preference of the values available from literature is
known so 50% weightage was given to both of them
in the mathematical algorithm used to arrive to a pos-
terior. The data fusion center shown in figure 5 sum-
marises all the results with their respective values in
case of conjugate prior. Table 2 can be referred for de-
tailed results on all parameters after Bayesian updat-
Figure 5: Data Fusion system using Bayesian approach for up- ing using Jeffreys and conjugate prior. The clear cover
dating clear cover using data obtained from Ferroscan instrument obtained from conjugate prior (40.59 mm) was found
using conjugate prior to be more conservative value than obtained from Jef-
freys prior (29.35 mm). This was expected as data
algebraic convenience, giving a closed-form expres- from Jeffreys prior reports only the readings at the
sion for the posterior: otherwise a difficult numeri- present state of the building without taking into ac-
cal integration may be necessary. All members of the count the clear cover at the time of construction which
exponential family have conjugate priors. For details is a higher value. Figure 6 presents the posterior prob-
of the bayesian algorithm used to combine NDT test ability distributions of the population of clear cover
data and fusion operator (Miranda et al. 2009, Mishra considering the mean values of the mean and standard
2013, Ramos et al. 2015) can be referred. deviation. The main aspect to focus on when compar-
ing the posterior results for the population of cc using
Table 1: Posterior distributions considering Jeffrey’s and Conju- the Jeffreys and conjugate priors is that the variance is
gate prior distributions
higher for the latter. In fact, as it is clearly illustrated
Posterior Normal Distribution in figure 6, the uncertainty is higher for the conjugate
Jeffrey’s µ, σ 2 − N (31.15, 7.622 /20) distributions. This fact is due to the consideration of
(n − 1)s2 /σ 2 − 1102.6/σ 2 − > χ2n−1 the prior information uncertainty which does not exist
Conjugate µ, σ 2 |X − N (40.5, σ 2 /40) when using Jeffreys prior.
1/σ 2 − γ(40.5, 1/3905.1)
Table 2: Posterior estimates of clear cover in mm using Jeffreys
and Conjugate prior
0.04 Parameter Jeffreys Conjugate
Normal Posterior−Jeffreys
0.035
Normal Posterior−Conjugate µ 30.26 40.58
σ(µ) 3.89 1.41
0.03
σ 7.89 8.84
0.025 σ(σ) 1.31 1.01
Probability Density
5 RESULTS