Speaker
Description
Non-Destructive Testing (NDT) is fundamental for guaranteeing industrial structural integrity, with Infrared Thermography (IR) highlighted for its non-invasive detection capacity and no harmful radiation. However, the dependence on manual parameterization and the complexity of analyzing transient data poses a difficulty for the development of automated system for failure detection. This work proposes the development of optimized and generalist models based on Deep Learning (DL) for failure detection using infrared thermography, aiming to reduce human intervention. The developed methodology used the PVC-Infrared dataset, selecting 5 specimens subjected to Pulsed Thermography (PT) containing artificial, i.e. previous known, defects. Raw data was first processed with two well-known techniques - Thermographic Signal Reconstruction (TSR) and Principal Component Thermography (PCT) - for contrast enhancement and noise reduction. From the processed images, 50 Regions of Interest (ROIs) were extracted, balanced between failure and integrity classes. Experiments were performed with these 5 different structures to evaluate method robustness under sample variations. Training was conducted with 36 samples (stratified between training and validation), while the evaluation was performed on 14 samples not used in the adjustment. This separation strategy reinforces evidence of generalization, indicating that the methodology does not limit itself to memorizing patterns, but preserves the capacity for discrimination in unprecedented data. In the testing stage, model calibration indicated an optimal decision threshold of 0.7677. The model reached, in the test set, an Accuracy of 92.86%, Precision of 100% (absence of false positives), and an Intersection over Union (IoU) of 0.8571. These results exceed classic semantic segmentation architectures applied to the same dataset such as U-Net and SegNet and demonstrate competitiveness with the state of the art, validating the effectiveness of the ROI-based approach and advanced pre-processing for automated monitoring applications. The next steps include evaluating the performance of the proposed methodology for different materials, e.g., training with PVC samples and testing with composites and metal samples.
