Study On Detection And Characterization Of Deep Defects In Multi-layered Structures Using Eddy Current Methods

As it is known to all, detection and quantitative evaluation of deep defects in multi-layered structures is an essential task in a number of industries ranging from aerospace industry to atomic engineering. It is widely recognized as one of the difficult inverse problems and principal challenges in the research field of eddy current nondestructive testing (ECNDT). Now, for ECNDT, the qualitative study in multi-layered structures has been presented. However, the quantitative evaluation problem in multi-layered structures remains to be dealt with. Supported by the National Natural Science Foundation of China under grant 50505045 (2006-2008), we carry out the research of detection and quantitative evaluation of deep defects in multi-layered structures from eddy current (EC) signals. The obtained results in this paper are important for the continued scientific research and application in this field. The main works and innovations of the paper are as follows:1. The field analysis based ECNDT method is investigated. The experimental ECNDT system based on giant magnetoresistive (GMR) probe for estimating dimensions of deep defects in multilayered structures is developed successfully. Here, a new EC testing technique with a GMR sensor is used to enhance the sensitivity and spatial resolution of the measurement. The GMR based EC probe can perform better than the conventional probe for low-frequency applications, i.e., when detecting defects deep buried in multi-layered structures.2. Signals preprocessing techniques are studied. Signals preprocessing techniques for noise elimination including time-domain methods, frequency-domain methods, and time-frequency domain methods are studied, respectively. We present a general robust procedure for signals preprocessing. Firstly, the original signals are preprocessed by median signal tracking consists of evaluating the distance between consecutive samples in the complex plane for the impulse noise suppression. Then, the wavelet packet analysis method with Shannon entropy threshold is implemented for EC signals de-noising. The experimental results show that the presented method is superior for EC signal de-noising.3. Signal feature extraction techniques are studied. In ECNDT, the interpretation of EC signals which are produced by a pick-up probe is very difficult since the electromagnetic field is nonlinear. We present a novel EC signal feature extraction method by using kernel principal component analysis (KPCA). By benefiting from a kernel perspective, KPCA is more powerful and applicable to nonlinear processing in a simpler and nicer way. It is shown by extensive experiments that KPCA is proper for signal feature extraction in ECNDT.4. The recognition and classification methods of signals from deep defects in multi-layered structures are studied. A novel approach for EC signal classification using the method integrated KPCA and support vector machine (SVM) has been proposed and investigated. SVM can be carried out well in the classification of low amounts of EC signal records due to its remarkable characteristics such as the structural risk minimization principal. Experimental results show that this approach can provide promising classification performance of different types of EC signals.5. The quantitative evaluation models for estimating dimensions of defects in multi-layered structures are studied by using regression algorithms. In the paper, a kernel partial least squares (KPLS) regression method has been proposed to construct the inversion model of defect dimension estimation. It can be used to model nonlinear EC data relations. When the KPLS regression is used to estimate the dimensions of the rectangular defects, the relative errors are less than±20%.6. The quantitative evaluation models for estimating dimensions of defects in multi-layered structures are studied by using Bayesian networks (BNs). In ECNDT, the inherent uncertainty and stochastic nature of inspection need to be dealt with. BNs are applied to quantitatively evaluate the realistic multi-dimensional characteristic parameters of defects by probability inference. In this paper, these ideas in context with continuous variables and dependencies are used. We mainly discuss how to construct BNs from the domain knowledge and the real research data, and how to perform probability inference in BNs. Firstly, we considered a simple problem that only diameter of defects in the multi-layered structures needed to be determined. The relative errors are less than±9%. Secondly, a more complex example was considered. The signals collected from multi-layer samples with defects varying diameter and depth are analyzed. The relative errors are less than±11%.