Pattern Recognition for Guided Waves Damage Identification Systems in Beams

Edmund Liew (2008). Pattern Recognition for Guided Waves Damage Identification Systems in Beams PhD Thesis, School of Engineering, The University of Queensland.

       
Attached Files (Some files may be inaccessible until you login with your UQ eSpace credentials)
Name Description MIMEType Size Downloads
n40680583_phd_abstract.pdf n40680583_phd_abstract.pdf application/pdf 94.29KB 0
n40680583_phd_content.pdf n40680583_phd_content.pdf application/pdf 15.56MB 0
n40680583_phd_front.pdf n40680583_phd_front.pdf application/pdf 304.79KB 0
n40680583_phd_totalthesis.pdf n40680583_phd_totalthesis.pdf application/pdf 15.69MB 0
Author Edmund Liew
Thesis Title Pattern Recognition for Guided Waves Damage Identification Systems in Beams
School, Centre or Institute School of Engineering
Institution The University of Queensland
Publication date 2008
Thesis type PhD Thesis
Supervisor Veidt, Martin
Subjects 290000 Engineering and Technology
Formatted abstract
The advent of advanced sensor technologies and accurate measuring instruments has motivated the research and development of guided waves in quantitative nondestructive evaluation and structural health monitoring. Guided waves are ultrasonic waves that are highly sensitive in detecting discontinuities in its path of propagation and can travel large distances, interrogating less accessible and obstructed locations of a structure. Despite the advantages, there remains the challenge to evaluate damages from raw signals alone, which leads to the feasibility investigations of advanced signal processing techniques including pattern recognition. Pattern recognition identifies features in the measured transient wave response signals with the aim to quantify damages from a large parameter space within the structure. The research conducted and documented in this thesis studies and develops the application of supervised regression neural networks for pattern recognition as a signal processing tool for guided waves damage identification in rectangular aluminium beams.
The pattern recognition strategy considers training the neural network with simulation data for practical cost saving reasons to construct an intelligent system for identification of experimentally measured inputs. A program framework to generate simulation data for training the neural network is devised where the damage is modelled as an inhomogeneity while non-dispersive wave propagation is assumed. A representative number of rectangular beam specimens with different artificial step damages are fabricated to obtain measurements of the transient wave response signals for the experimental test database. Feature extraction procedures that include effective signal selection followed by decomposition with the discrete wavelet transform are performed to reduce sampling data for fast processing while preserving essential information in the network input patterns for sensitive evaluation of the damages.
Statistical pattern recognition with supervised neural networks contains a broad range of customisation properties, which are fundamentally studied to optimise the quality of training or generalisation of the damage parameter space considered. Common neural network properties like the resilient backpropagation training algorithm and validation early stopping criterion are selected based on these studies while a systematic approach is developed to
design network parameters that are input dependent. The designed network is also used to verify the selection of the feature extraction procedures for high performance and accurate prediction of damages from identifying experimental input patterns.
The broad generalisation spectrum due to random initialisation of network weights and training patterns affects the sensitivity of the trained network to experimental patterns where an accurate prediction can only be obtained from the average in a test results distribution collected in a number of training and test trials. These results also infer that the quality of damage identification in the trained network is test pattern dependent. Further improvements of the pattern recognition system are achieved through series and parallel network processing methods developed based on this understanding. The series network incorporates a technique called weight-range selection that iteratively reduces the size of the training range based on the statistical analysis of the intermediate test results distribution. The reduced training range size improves in generalisation specific to the test pattern where the series network can be observed to achieve more than 80% increase in accuracy compared to conventional neural networks. In addition to these features, integration of a developed parallel network system improves the robustness and reliability of pattern recognition via distributed monitoring by processing different observations of the parameter space in collaborating networks where data fusion can function to associate intermediate test results and isolate outliers.
The designed and optimised pattern recognition system is then tested in more complex damage identification scenarios including the consideration of interference effects from boundary wave reflections and in a pitting corrosion damage model to provide an insight into the potential of the system. Comparison between equivalent experimental and simulated input test patterns for these case studies reveals substantial discrepancies in a number of the patterns. However, application of the developed pattern recognition system with series and parallel processing shows mitigation of these errors where significant increase in the accuracy of the output predictions from conventional neural networks is achieved.


 
Citation counts: Google Scholar Search Google Scholar
Created: Thu, 10 Jul 2008, 13:19:01 EST by Noela Stallard on behalf of Library - Information Access Service