Artigo 2 – Progress in Nuclear Energy
140
even when using high-speed photography, is that the picture is often confusing and
difficult to interpret, especially when dealing with high velocity flows. In addition, there
are systems that are opaque where flow visualization is impossible; then,
such analysis
is also not possible by this method (Wu et al., 2001, Jin et al., 2003). Therefore, a non-
invasive system that can provide material volume fraction (MVF) predictions regardless
of a priori knowledge of the flow regime, without subjective evaluation, is a great
contribution.
Together with the detection system artificial neural networks (ANNs) (Haykin,
1994) has been used in order to interpret the pulse height distributions
(PHDs) obtained
by gamma-ray radiation detectors to identify the flow regime (Mi et al. 1997, 1998, Wu
et al., 2001; Jin et al., 2003) and predict the MVFs (Salgado et al., 2007, 2009, Bishop
and James, 1992; open et al., 1999). ANNs are mathematical models inspired in the
human brain, which has the ability of learning by examples. ANNs are able to discover
behaviors and patterns from a finite set of data (called the “training set” or “training
patterns”). If an adequate training set is provided, the ANN is able to generalize the
knowledge acquired during (learning) process, responding adequately to new situations
(not comprised in the training set).
The training and test patterns (different volume fractions for the three flow
regimes) were obtained by means of static and ideal mathematical models for annular,
stratified and homogeneous regimes.
These models were developed by mathematical simulation using the Monte
Carlo N-Particle eXtended
(MCNP-X) computer code (Pelowitz, 2005) based on the
method of Monte Carlo (MC) (Open et al., 1998, 1999). The MC technique is a widely
used simulation tool for radiation transport, mainly in situations where physical
measurements are inconvenient or impracticable. In this work the MCNP-X code, which
is specific for simulating electron and photon transport through materials with various
geometries, has been used. The model developed in the MCNP-X code considers the
main effects of radiation with the matter involved and the PHDs from the NaI(Tl)
detectors. The energy resolution, dimensions and characteristics of a real detector are
also considered; in general, the model presented tends to approach the realistic case.
In this work, the whole gamma-ray PHDs obtained by detectors are directly used
to feed the ANNs without any parameterization of the signal, which allowed the use of
simplified detection geometry consists of two NaI(Tl) detectors, the first one positioned
at 180° diametrically opposed to sources of
241
Am and
137
Cs and the second one at 45°.
In addition, the system considers the transmitted (I
T
) and scattered (I
S
) beam
measurements in order to increase the visualization of the cross-section, making the
response less dependent on the flow regime and also to obtain sufficient information to
determine precisely the volume fractions regardless, a priori, of the flow regime.
The developed ANN system comprises of four ANNs. The first one is trained to
identify the dominant flow regime and other ones are trained for volume fraction
predictions of each specifically regime. An evaluation of the quality training of ANN
was made from 25 patterns not used during the training phase, also generated by
mathematical code.
In this study, the training patterns (combination of the volume fractions of each
material) were distributed uniformly throughout the search space; moreover, the choice
of each data set was performed manually. Another important enhancement over the
mathematical model used in previous work (Salgado et al, 2009) is the use of a more
realistic model of the NaI(Tl) detector, considering the real dimensions and materials
compositions, as well as its energy resolution. These improvements allowed the