An efficient parametrized optical infrared thermography 3D finite element framework for computer vision applications

Matching infrared thermography (IRT) with deep learning-based computer vision has recently gained a lot of interest for automated defect assessment in materials. One of the remaining bottlenecks concerns the necessity of a large and diverse experimental and/or virtual training dataset in order to achieve a sufficiently generalizable computer vision algorithm. This paper presents a parametrized 3D finite element (FE) framework, implemented in Fortran90, for efficiently simulating optical infrared thermographic inspection of multi-layer anisotropic media and establishing large-scale virtual dataset with sufficient diversity. The interface element is introduced for the modelling of an imperfect thermal contact, allowing to simulate a variety of defect types. The flexibility of the interface element makes it possible to simulate delaminations with different thickness using the same dis-cretized model. Validation is done for two benchmark cases which are representative for a fiber reinforced polymer laminate with delamination-like defects. In order to achieve true-to-nature thermographic simulation data, non-uniform heating conditions are adopted from experiment, and a stochastic morphology generator is introduced for modelling realistic irregular defect geometries. To demonstrate the added value of a large, diverse and true-to-nature virtual database for computer vision applications, a Faster-RCNN model was trained on a generated virtual dataset for the detection of delamination-like defects in fiber reinforced polymer laminates. Application of the trained Faster-RCNN on experimental thermographic data yields excellent inference results, illustrating the high generalization ability of the virtually trained object detector.

Automated summary

This paper proposes a parametrized 3D finite element (FE) framework for simulating optical infrared thermographic inspection of multi-layer anisotropic media and generating a large-scale virtual dataset. The interface element is introduced for simulating various defect types, and non-uniform heating conditions and a stochastic morphology generator are used for realistic simulation. The trained Faster-RCNN model demonstrates excellent performance on experimental thermographic data.