Detection of hidden corrosion by pulsed eddy current using time frequency analysis

This thesis reports development of an automatic inspection method incorporating pulsed eddy current testing to determine the position and size of flaws buried in an aluminum double-layer structure and also to measure thickness variation in an aluminum single-layer system. Three main aspects are presented; building up a pulsed eddy current setup with all the components to acquire signals from defected samples, applying new signal processing and treatment methods using time-frequency distributions to reveal hidden information from the acquired pulsed eddy current system and applying artificial intelligence methods to minimize human interference in estimating and determine the defect distribution in aircraft fuselages. Inspection in the aerospace industry is a sensitive case and a non-destructive method must be accurate and reliable for aircraft inspection.;In the first section of this thesis, several types of metal losses with a wide range of depths were fabricated on surfaces of an aluminum plate to simulate corrosion and defect conditions. A complete pulsed eddy current setup was built and programmed to acquire precise signals from defected and non-defected surfaces.;In the second section, pulsed eddy current signals were processed using a time-frequency analysis method. Internal and external noise effects were simulated using small defects placed far from the surface, inter-layer gap and lift off. Noise resources cause a reduction in efficiency of signal representation in the time domain and a time-frequency analysis is therefore implemented to improve representation of signals and reveal more hidden information from defects. In this step the acquired pulsed eddy current signals, which are represented by a voltage-time series (in time domain), are converted to represent signals in three dimensions (time-frequency-amplitude). Over ten different time-frequency distributions were programmed and implemented. These distributions use Short Time Fourier Transforms (STFT) as a linear time-frequency distribution and bilinear time-frequency distributions such as Wigner-Ville distribution, Smooth Wigner-Ville distribution, Pseudo Wigner-Ville distribution, Born-Jordan distribution, Rihaczek distribution, Choi-Williams distribution, Zhao-Atlas-Marks distribution, Butterworth distribution and spectrogram. The Rihaczek distribution was chosen as superior among all the above distributions to convert signals from time domain to time-frequency domain. Two type of bilinear time-frequency distributions are reported in this thesis which are Rihaczek distribution for hidden defect detection in double layer structures and spectrogram for thickness variation detection in single layer. The Rihaczek distribution has shown minimum interference and cross term effects compared to other time-frequency distributions. Additionally, using the real part of the energy in the Rihaczek distribution prevents the need to abandon any sort of analogy to physical phenomena with negative values for energy. In this thesis, the Rihaczek distribution was used to represent signals which come from a double-layer system. In the case of thickness variation measurements, a spectrogram was applied to represent signals from a single-layer system in three dimensions.;The final point studied in this research was to implement and apply a feature extraction method and a classifier for automatic defect detection. Based on a mathematical model for extracting features from time-frequency representation data, principal component analysis (PCA) was implemented and applied to remove redundant data, reduce the size of the data set and also extract some new parameters as a unique specification of each type of synthetic defect and input of classifiers. The efficiency of classification methods is usually determined by misclassification error. Two types of discriminative and probabilistic classifiers had minimum misclassification error among several classifiers which were studied during this project. These classifiers are K-Mean Clustering as a discriminative method and Expectation-Maximization (EM) algorithm as a probabilistic method. To calculate misclassification error in each classifier, several unknown samples were tested to determine the reliability of classification and also amount of misclassification error.;Combination of all the above steps in this work provides an automatic inspection tool (hardware and software) which has high accuracy and reliability for defect detection in aircraft fuselage and metallic multilayer structures. (Abstract shortened by UMI.).