Improved Deep Learning for Defect Segmentation in Composite Laminates Inspected by Lock-in Thermography

Abstract

In the context of aerospace metrology, reliable inspections of production yields are mandatory for clear safety reasons. If defects are fastly detected and, above all, segmented, further economical benefits may emerge since repairing strategies can be set up. Given this scenario, this paper presents an automatic methodology to process data from lock-in thermography inspections of composite laminates. Thermal analysis works on area scans of the target surfaces and, consequently, leads to complete information of the whole structure in a relatively short time. A deep learning network has been trained to analyze amplitude and phase maps to detect and segment defective inclusions. The reliability of this analysis is typically undermined by actual experimental issues, due to the homogeneity of the input excitation. For this reason, this study proposes a preliminary processing to manage the inhomogeneity of the input excitation, which typically adds a non-negligible bias to the amplitude and phase maps. The improved reliability has been proven through experiments performed on actual samples, coming from aerospace production lines. The outcomes of these experiments have proven a final per-pixel accuracy of 84.38% in the segmentation of buried defects.