The criterion of relevance, used in this work with
nonlinear classifiers, definitely proves that only four of the
six initial features were relevant for the classification of
class defects studied. In this case when the inputs were made
up of four features, the nonlinear classifiers produced high-
performance results for the training data. An input made,
using only the combination of three relevant features,
maintained the indices of success of the classifier, really
proving that the importance is the ‘quality’ of the feature
employed and not the ‘quantity’.
The use of the principal nonlinear discrimination
components, to demonstrate the classes of defects in two
dimensions, gives an idea of the problem of separation
between classes, although for classification three or four
dimensions are still used. When the components are used as
input vectors of the classifiers, good results are obtained.
These methodologies can perfectly well be extended to
larger size problems where the calculations are relatively
much more complex.
The calculations of probability of success of the
classification, using nonlinear classifiers by implemented
neural networks, were high, signifying that even with the
lack of test data to check their generalization, the probability
of success of the classifiers is high for new samples.
Certainly, as can be shown in various publications cited here,
the lack of a high number of samples to increase
the reliability of the results is a frequent problem and
remains as a motive for further investigation by the authors.
It is important to point out that the defect classes such as
cracks and lack of fusion, important in terms of weld beads,
have not yet been evaluated by this technique due to a lack
of the quantity of reliable observations available, but
certainly this will be done in a future work with the
acquisition of new radiographic patterns.
The criterion of relevance, used in this work with
nonlinear classifiers, definitely proves that only four of the
six initial features were relevant for the classification of
class defects studied. In this case when the inputs were made
up of four features, the nonlinear classifiers produced high-
performance results for the training data. An input made,
using only the combination of three relevant features,
maintained the indices of success of the classifier, really
proving that the importance is the ‘quality’ of the feature
employed and not the ‘quantity’.
The use of the principal nonlinear discrimination
components, to demonstrate the classes of defects in two
dimensions, gives an idea of the problem of separation
between classes, although for classification three or four
dimensions are still used. When the components are used as
input vectors of the classifiers, good results are obtained.
These methodologies can perfectly well be extended to
larger size problems where the calculations are relatively
much more complex.
The calculations of probability of success of the
classification, using nonlinear classifiers by implemented
neural networks, were high, signifying that even with the
lack of test data to check their generalization, the probability
of success of the classifiers is high for new samples.
Certainly, as can be shown in various publications cited here,
the lack of a high number of samples to increase
the reliability of the results is a frequent problem and
remains as a motive for further investigation by the authors.
It is important to point out that the defect classes such as
cracks and lack of fusion, important in terms of weld beads,
have not yet been evaluated by this technique due to a lack
of the quantity of reliable observations available, but
certainly this will be done in a future work with the
acquisition of new radiographic patterns.
การแปล กรุณารอสักครู่..