Speaker
Description
Pulsed Thermography (PT) has established itself as a robust non-destructive evaluation (NDE) technique for inspecting industrial materials. However, the automated detection and characterization of subsurface defects in polyvinyl chloride (PVC) components present distinct technical challenges compared to metals or carbon fiber composites. PVC is characterized by low thermal conductivity, high reflectivity, and variations in emissivity, which frequently result in thermograms with low signal-to-noise ratios and blurred defect edges due to lateral thermal diffusion. Consequently, relying on manual interpretation by human experts is often inefficient and prone to errors for this material class. To address these limitations, this study proposes an automated defect segmentation framework that integrates advanced signal processing with a Transformer-based neural architecture.
The proposed methodology begins with data preprocessing using Thermographic Signal Reconstruction (TSR). The raw temporal temperature sequences for each pixel are fitted in the logarithmic domain using polynomials and converted into first-order derivative images. This transformation effectively suppresses temporal noise and mitigates the effects of non-uniform heating, significantly enhancing the thermal contrast of internal anomalies. To capture complex dependencies, spatio-temporal features are extracted pixel-by-pixel using spatial windows (e.g., 3×3) across the temporal evolution. This approach allows the model to analyze the correlation between a central pixel and its surrounding neighbors, compensating for the "blurring" effect caused by heat diffusion in the PVC structure.
The core of the framework is a Transformer network that employs a multi-head self-attention mechanism to dynamically weigh the importance of global and local features, enabling precise segmentation of defect morphology. Recognizing that deep learning performance is highly sensitive to hyperparameter configuration, this study moves beyond traditional manual tuning or random search methods. Instead, a Bayesian optimization stage is implemented to automatically tune the network. This probabilistic approach efficiently navigates the search space to identify the optimal combination of learning rate, batch size, and attention heads, thereby maximizing the model's generalization capability while reducing computational costs.
Experimental validation was conducted on PVC specimens containing artificial flat-bottomed holes of varying diameters and depths. The results demonstrate that the proposed framework, optimized via the Bayesian approach, achieves superior performance in terms of robustness and precision. The method yields high values in Pixel Accuracy (PA) and Intersection over Union (IoU) metrics, outperforming architectures based solely on Convolutional Neural Networks (CNN) or Recurrent Neural Networks (RNN/LSTM). These findings confirm the framework's potential for automated quality assurance in the manufacturing of PVC components.
