<p>The objective of this paper is to improve defect detection accuracy by combining information from multi-modality nondestructive evaluation (NDE) sensing. &nbsp;The single modality inspection techniques may not reveal all the characteristics of the damage, which would lead to miss of mostly shallow and sub surface defects and thus providing unreliable decision making. Fusion of data obtained from multiple inspections can lead to improve signal quality, interpretation and decision making thereby reducing uncertainties to a large extent.</p><p>In this paper, we improve the detection accuracy of surface and sub surface defects present in pipeline steel samples by combining the data collected from magnetic flux leakage (MFL), eddy current (EC) and pulse eddy current (PEC) based NDE methods.&nbsp; Both simulation and experimental based studies are performed to obtain the signals obtained from defect having various dimensions. Single sensor systems do not provide accurate estimation and profiling of the shallow and sub surface defects as the perturbed signals due to these defects are weak and are mostly overwhelmed by the background noise. Data fusion among these unrelated physical inspection systems are done at pixel, feature and decision level. Different statistical and deep learning based fusion strategies are developed to provide accurate defect detection.&nbsp; Four data fusion models such as fusion aided with a deep learning framework, fusion with convolutional sparse representation (CSR), fusion based on cross bilateral filter (CBF) and fusion using discrete wavelet transform (DWT) are designed to incorporate NDE data. In both the fusion aided with deep learning and CSR methods the source images are decomposed into two layers: &nbsp;base and detailed layers. Base layers which are extracted from source images by Tikhonov regularization contain the common features and redundant information. In deep learning-based fusion, a deep learning (VGG-19) based framework is used to extract the information from the deep detailed layers. The feature maps and weights of each layers are computed and finally the detailed and base parts are fused to obtain the fused result. In CSR model the fusion of the detailed layers is achieved by mapping the sparse coefficients in spatial domain. In the CBF method all the important information obtained in the individual source images are fused taking into consideration the weights computed from detailed image. The weights are evaluated by measuring the strength of details in a detail image obtained by subtracting the output of CBF filter from the original image. The fourth fusion method (DWT) has been largely used in NDE community for fusion purpose. The approximate and detailed coefficients are obtained by applying discrete wavelet transform on the original images. Then different fusion ways such as max, min and average are considered to fuse the obtained coefficients. Fused image is obtained on performing inverse DWT on the fused coefficients. The purpose of these four data fusion models is to provide better performance than that without fusion for defect detection. We report the efficacy of the proposed methods in detecting defects of various dimensions. Fusion performance, fusion loss and fusion artifacts obtained by using these methods are compared and the best fusion model is reported for NDE based defect detection in steel surfaces.</p>

Improved Non-Destructive Evaluation Based Defect Detection Using Pixel Level Data Fusion

Paper Proceedings: https://asmedigitalcollection.asme.org/qnde/QNDE2021/volume/85529

ASME 2021 48th Annual Review of Progress in Quantitative Nondestructive Evaluation (QNDE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.