Enhanced and Automatic Defect Detection in Thermography using Stacked Denoising Autoencoder

Active infrared thermography (AIRT) with frequency-modulated stimuli has emerged as a feasible and cost-effective non-destructive testing method for checking a variety of materials, with improved fault identification and depth resolution. However, assessing this non-stationary thermal response while reducing noise and inhomogeneous backgrounds is difficult in FMT. Conventional thermographic data analysis techniques use signal or image processing methods and supervised learning models, which are known for linear representation or feature extraction of thermographic data. This paper describes a stacked denoising autoencoder (SDAE) in frequency modulated thermography (FMT) that extracts non-linear information from temporal thermal profiles to improve flaw detection in a steel structure. The SDAE trains on temporal thermal profiles in order to generate denoised profiles at the other end. Furthermore, the latent space in the middle of the SDAE decreases dimensionality while highlighting flaws. Finally, the latent space of SDAE and local outlier factor (LOF) were coupled to form SDAE-LOF, which provided automatic fault detection in unsupervised learning passion. Experimentation results on a mild steel structure show an improvement in defect signature by reducing noise in temporal thermal profiles, and the proposed SDAE-LOF achieved an AUC of 0.81 in automatic defect detection.