An end-to-end deep learning framework combining convolutional, recurrent, and attention layers achieved a 96.10 percent accuracy rate in detecting and classifying internal defects within aerospace materials from infrared thermal wave image sequences. Rather than operating as an isolated classification or regression tool, the unified model simultaneously identifies defect categories and determines their precise spatial locations. The research, conducted by Gang Xu and Yunhan Zhou and published in the journal Aerospace Systems, evaluates continuous temperature field information to improve the stability of non-destructive testing for concealed structural flaws.
In experimental evaluations of defect classification, the proposed architecture outperformed alternative deep learning configurations evaluated under identical conditions. The achieved accuracy of 96.10 percent marked an absolute improvement of 6.00 percentage points over a baseline combining a convolutional network with gated recurrent units, known as CNN-GRU. It also exceeded the performance of a CNN-GRU network equipped with an attention mechanism by 2.30 percentage points, demonstrating the specific advantages of the chosen structural components.
The model maintained high balanced performance across all recorded diagnostic metrics during the classification assessments. The authors documented a precision of 95.95 percent, a recall of 95.64 percent, and an overall F1-score of 95.79 percent for defect categorization. In addition, the area under the receiver operating characteristic curve reached 98.40 percent, reflecting consistent discriminatory power across various decision thresholds when distinguishing between defect classes.
For the task of spatial defect localization, the proposed end-to-end system demonstrated high spatial overlap and geometric precision on the thermal wave sequences. The model achieved an intersection over union metric of 0.928 and restricted the localization error to 1.86 millimeters. Furthermore, to evaluate how accurately the network predicted temporal temperature responses, the researchers calculated normalized prediction errors, reporting a normalized mean squared error of 0.0085 alongside a root mean square error of 0.092.
The development of this unified architecture was motivated by severe physical measurement difficulties encountered during non-destructive evaluations of aerospace components. Subsurface defects in these materials are characterized by strong concealment, low thermal response contrasts, and complex temporal diffusion behaviors that complicate automated visual detection. While infrared thermal wave image sequences provide a continuous stream of temperature field data, extracting reliable signals requires overcoming significant background thermal noise and tracking subtle transient heat dissipation over time.
To address raw image variability before feature extraction begins, the framework routes the continuous thermal frames through a dedicated data preprocessing pipeline. The sequence undergoes normalization, noise reduction filtering, dimensional unification to standard image sizes, and automated screening to isolate keyframes. Following this preparation, a two-dimensional convolutional neural network extracts spatial image features, targeting temperature anomalies, structural edge contours, and localized hot spot patterns that indicate subsurface material damage.
To capture dynamic heat transfer characteristics, the framework subsequently introduces a long short-term memory network designed to learn the temporal dependencies of thermal diffusion and temperature decay. An attention mechanism is integrated alongside the recurrent layers to dynamically calculate and increase the weighting applied to the most informative keyframes and localized regions. The authors concluded that uniting spatial feature extraction, temporal modeling, and attention-weighted fusion effectively suppresses background thermal noise while strengthening the accuracy and end-to-end stability of defect detection in aerospace materials.
Subject of Research: Deep learning defect detection in aerospace materials using infrared thermal imaging
Article Title: End-to-end detection method of infrared thermal wave image sequence defects of aerospace materials based on 2D-CNN-LSTM-attention network
Article References: Xu, G., & Zhou, Y. (2026). End-to-end detection method of infrared thermal wave image sequence defects of aerospace materials based on 2D-CNN-LSTM-attention network. Aerospace Systems. https://doi.org/10.1007/s42401-026-00579-7
Image Credits: AI Generated
DOI: 10.1007/s42401-026-00579-7
Keywords: infrared thermal wave imaging, aerospace materials, defect detection, non-destructive testing, convolutional neural networks, long short-term memory, attention mechanism, defect localization, defect classification, thermal diffusion, temperature decay, keyframe screening, thermal noise suppression, intersection over union, thermal response prediction, temperature field analysis
Cite Scienmag News
Blake Davidson. (November 20, 2026). Deep learning model detects aerospace material defects with 96 percent accuracy. Scienmag. https://scienmag.com/deep-learning-model-detects-aerospace-material-defects-with-96-percent-accuracy/
Blake Davidson. "Deep learning model detects aerospace material defects with 96 percent accuracy." Scienmag, 20 November 2026, https://scienmag.com/deep-learning-model-detects-aerospace-material-defects-with-96-percent-accuracy/. Accessed 20 November 2026.
Blake Davidson. "Deep learning model detects aerospace material defects with 96 percent accuracy." Scienmag. November 20, 2026. https://scienmag.com/deep-learning-model-detects-aerospace-material-defects-with-96-percent-accuracy/








