Detection and Classification of Defects Using ECT and Multi-Level SVM Model

In this work, classification of cracks in a non-ferromagnetic material (aluminum 1050) is presented using two supervised machine learning techniques called support vector machines (SVM) and decision tree. The classification performances between the two classifiers were evaluated using confusion matrices and miss-classification errors. In order to improve the classification performance, a multi-level SVM model is proposed wherein each level implements SVM using different features and kernels. The experimental tests were performed in two aluminum plates containing 18 defects with different depths and lengths. The 18 defects were inspected in the surface and sub-surface locations, resulting in a study of 36 different crack cases. The experimental data used to extract features for supervised machine learning were obtained using an eddy current testing (ECT) probe that includes a planar coil and a tunnel magnetoresistance (TMR) sensor. The probe was excited by a multi-frequency signal with four testing frequencies to evaluate the dimensions and the locations of the defects. Due to the experimental noise existing in the measurement signals, it is important to pre-process the signals before feature extraction. Signal processing techniques (such as moving average filter, spline interpolation and symmetry corrections) were used before extracting the features for classification.

I. Introduction

In the non-destructive testing (NDT) field, eddy current testing (ECT) is used in a wide variety of industrial applications for the evaluation of the condition of metallic components [1]–[6]. For example, in the aerospace industry, a human operator is responsible to decide if the sample under test is acceptable or needs to be substituted. The evaluation is usually performed by the visualization of the bi dimensional signature obtained by the ECT probe [7]. An inadequate human examination can cause a wrong evaluation of the condition of the aircraft structure. Hence, it is important to rely on a mathematical model that aids the evaluator to make a correct judgment on the condition of the sample under test. The progress in machine learning tools has been attracting researchers to use them for the predictive maintenance in the eddy current NDT field. Machine learning is used on supervised learning or unsupervised learning. In the case of supervised learning, a training model is developed on known input and output data so that it can predict future outputs. On the other hand, unsupervised learning finds hidden patterns in the input data without the knowledge of the actual pattern. In the case of supervised learning, classification and regression techniques are used to develop predictive models. In [8], Bernieri et al. compared the performance of two machine learning techniques, artificial neural network (ANN) and support vector machine (SVM), for the regression of the dimensions of cracks in aluminum structures. The results showed that the SVM approach for regression provided the best solution to solve crack characterization problems in eddy current testing. In [9], Kim et al. proposed the use of Principal Component Analysis (PCA) and k-means algorithms to detect and classify sub surface cracks with different lengths at rivet sites in aircraft structures. In [10], Rosado et al. used nonlinear regression and ANN to estimate the parameters of defects. In [11], Bernieri et al. presented a biaxial ECT probe with a multi frequency excitation and an optimized support vector machine for regression (SVR) to estimate the size of cracks in a thin aluminum plate of 2 mm. In [12], Chen et al. proposed a Fisher linear discriminate analysis to get automated feature extraction and selection for defect classification in multi-layer structures using pulsed eddy current technique. In [13], D’Angelo et al. performed a defect classification using the Lissajous signatures of the defects obtained by the ECT probe. Several features such as length, width, area and inclination angle of the Lissajous signatures were considered to perform the classification using Naïve Bayes, C4.5/J48 Decision Tree, and Multilayer Perceptron neural network classifiers. In [14], Pasadas et al. used multi frequency excitation and SVM algorithm to classify the location and depth of sub surface defects in multilayer aluminum plates.