Application of the Hilbert-Huang Transform in Ultrasound Inspection Signals for Defect Classification using Neural Networks
Ultrasound testing, TOFD, Non-Destructive Testing, Hilbert-Huang Trans-
form, Non-linear models, Empirical Mode Decomposition, Neural Networks.
The use of digital signal processing as an aid for diagnosing failures in manufac-
turing processes has been very promising, as it allows to increase the efficiency of such
processes and ensures product quality and the safety of installations. This work proposes
the use of Hilbert-Huang transform (HHT) as a tool for feature extraction of ultrasound
signals obtained in an experimental procedure on welded joints using the TOFD techni-
que. HHT is a time-frequency decomposition that uses base functions for transformation
that are estimated from the signals of interest. Such base functions are obtained from
decomposition algorithms such as EMD, EEMD and CEEMDAN. Once the experimental
signals were obtained and the HHT applied for feature extraction, a classifier based on
multilayer perceptron neural network was used and showed a satisfactory performance,
reaching efficiency product above 90% in all considered cases. When EEMD and CE-
EMED were used as decomposition algorithms, the efficiency product achieved 98.5%
and 97.4%, respectively, whereas the EMD algorithm was used, the value reached 95.3%.
When HHT was associated with PCA, the classifier was able to discriminate the different
defect classes with 91.0% for HHT-EMD, 93.0% for HHT-EEMD and 92.7% for HHT-
CEEMDAN. These results are comparable to those found in the literature in which use
other pre-processing techniques, whose highest values obtained for a similar dataset were
94.8% using the DFT and 97.5% for DFT with PCA.