Deep learning-assisted thermal image analysis for assessment of cementitious composite flaws

Flaws significantly affect safety by reducing the strength and durability of the structure. Various non-destructive testing techniques, such as infrared thermography, ultrasound, and radiography, have been employed to detect flaws. However, conventional non-destructive testing techniques have limitations, including long inspection times, high costs, and accuracy of results that depend on the skill of the inspector. This study uses the thermal properties of multiwalled carbon nanotubes (MWCNTs) to collect thermal images of cementitious composites and to design a deep learning model for effectively detecting internal defects. The specimens are fabricated using cement paste with 1 wt% MWCNTs, and have a rectangular shape with a base area of 60×100 mm 2 and a height of 20 mm. The artificial internal defects are implemented in sizes ranging from 3% to 25% of the bottom surface area of the specimen. Thermal images are collected through a heat generation test. These thermal images are used as training data for the deep learning model. Two deep learning models are designed: one for detecting the presence of internal defects and the other for classifying their size. The model is optimized by tuning the number of Epochs and Batch size. The defect detection capability of each model is evaluated using a confusion matrix, accuracy, precision, recall, and F1-score. The average performance indicators of the two models exceed 0.97, demonstrating that the designed models achieve high performance in detecting internal defects. • Two deep learning models are designed, with one detecting the presence of internal defects and the other classifying their size. • Thermal images are collected using the thermal properties of MWCNT from specimens with internal defects ranging from 3% to 25% of the bottom surface area. • Thermal images are filtered and used to construct the training dataset for the deep learning model. • The defect presence detection model is designed based on CNN, while the defect size classification model is designed based on DenseNet-121. • The performance of the model is evaluated using a confusion matrix and evaluation indicators.