Enhanced defect detection in thermography through temporal denoising and deep feature extraction using a shallow convolution autoencoder

Various stimulation mechanisms and supportive processing techniques promote infrared thermography (IRT) as a powerful tool for inspecting various industrial objects, such as composites, metals and coatings. Frequency-modulated thermography (FMT) is one stimulation scheme offering superior depth scanning and resolution capabilities. However, noise and non-uniform thermal backgrounds make defect detection challenging in infrared thermography. Recent post-processing advancements following NDT 4.0 adapt machine learning and deep learning-based techniques to automate and enhance defect detection procedures. This study offers a shallow convolution autoencoder (SCAE) in FMT with one-dimensional convolution layers to reduce noise in temporal thermal evolution and train high-level features, resulting in improved defect signatures. This method focuses on creating the denoised temporal thermal response at the decoder end and extracting high-level thermal features from the latent space. Experiments were carried out on a mild steel specimen and a carbon fibre-reinforced polymer (CFRP) specimen embedded with flaws of varying depths and sizes. Compared to the shallow autoencoder (SAE) and stacked deep autoencoder (SDAE), the suggested SCAE efficiently denoised thermal profiles, allowing for improved fault signatures. Furthermore, the latent features of the SCAE produced superior defect signatures compared to the above autoencoder models and traditional dimensionality reduction methods such as principal component analysis (PCA), random projection transform (RPT) and partial least-squares regression (PLSR). The thermal profile and defect signalto-noise ratio (SNR) demonstrate the advantages of the proposed method for improved defect detection.