Dynamic Defect Detection in Fast, Robust NDE Methods by Transfer Learning Based Optimally Binned Hypothesis Tests

The structural integrity of safety-critical infrastructures diminishes over time, necessitating periodic inspections. The gathering of low-noise, precise nondestructive evaluation (NDE) readings for defect detection can be both time-consuming and costly. In practical applications, NDE data can be highly noisy due to factors like liftoff/stand-off distances, probe drift, scanning speed, and variations in data acquisition rates. Given the broad spectrum of potential uncertainties that can contribute to noise accumulation, characterizing the noise in such NDE data presents significant difficulties. As a result, the extraction of defect signals through deconvolution becomes challenging. This article introduces an NDE-based dynamic defect tracking framework designed for robust conditions. It incorporates regular periodic monitoring of the material under test (MUT) using a cost-effective, easy-to-implement magnetic flux leakage (MFL) probe. The framework proposes a dynamic setup that includes occasional, highly accurate, expensive MFL scans among regular, low-cost, noisy MFL scans. The objective is to detect and precisely track defect formation in metallic pipes over time, thereby enabling the efficient alarming and localization of defects before they reach detrimental sizes. A key characteristic of the proposed dynamic monitoring method is its noise type agnosticism in the low-cost MFL scans. It remains effective even when the underlying noise-producing uncertainties fluctuate significantly over time, a feature achieved through the use of transfer learning. The method transfers accurate information regarding the locations and sizes of existing corrosions from the detailed MFL scan to the subsequent noisy MFL inspections. With this supportive information from defective scan-points, it estimates the degradation of the signal-to-noise ratio (SNR) in the noisy MFL scans. The framework suggests binned hypothesis tests in noisy MFL scans, setting the level of aggregation based on the SNR estimate in the scan data. This proposed binned hypothesis test optimizes defect coverage and maintains control over the false discovery rate (FDR). Experimental MFL data, gathered under a wide range of uncertainties and defect types, illustrates the highly heterogeneous nature of noise distributions in low-cost MFL scans and the significant variance in SNR deteriorations. The necessity of using binned hypothesis tests in these low-cost MFL data is highlighted by the poor performance of pointwise multiple hypothesis tests. Finally, the high effectiveness of the proposed transfer learning-based binned hypothesis testing method is demonstrated. Although the method was applied to MFL data in this instance, it is also suitable for other NDE applications. The applicability of the algorithms has also been confirmed through validation on experimentally generated Eddy Current (EC) data.