Speaker
Description
Abstract:
Carbon fiber reinforced polymer (CFRP) composites are widely used in aerospace, but manufacturing and in-service processes can introduce hidden defects that threaten structural safety. Infrared thermography [1] is suitable for large-area non-destructive testing, but it is affected by non-uniform heating and noise, making it difficult to reliably identify weak thermal anomalies. This paper proposes Long short-term memory Autoencoder Principal Component Thermography (LAPCT) to model the spatiotemporal information of thermal image sequences in an end-to-end manner. This method utilizes the temporal dynamic features of each pixel within the sequence as its core input. By maintaining complete spatial relationships at the pixel level, similar temporal modes of neighboring pixels can enhance each other, thereby providing spatial contextual support for weak dynamic signals. Simultaneously, low-dimensional temporal features focus on the dynamic trajectory of each single pixel, avoiding interference from irrelevant spatial information. By introducing optical flow consistency loss [2], inter-frame dynamic features are further enhanced. Principal component thermography (PCT) [3] is applied to the reconstructed sequence to extract discriminative features. This method significantly enhances defect characterization and improves contrast-to-noise ratio (CNR) on aerospace CFRP samples, achieving more robust automatic detection compared to the traditional PCT model.
Sample:
This study is based on a honeycomb sandwich CFRP specimen. It is used in the side panels and floor of aircraft cargo holds. The actual sample image is shown in Figure 1.
Figure 1. Experimental Sample Case
To simulate the leakage conditions of in-service components, holes in different diameters were drilled on the back of the specimen. Water and aviation oil were injected into these holes to simulate water and oil leakage scenarios. The experimental specimen dimensions were 24.5 cm × 15.8 cm × 1.0 cm. Hole diameters decreased sequentially from 1.3 cm to 1.0 cm, 0.8 cm, and 0.7 cm. Specific details are shown in Figure 2.
Figure 2. Defect Distribution Diagram
Significant references
1. Liu Y, Yao Y, Wang F, et al. Review of unsupervised machine learning methods in active infrared thermography for defect detection and analysis. Quantitative InfraRed Thermography Journal, DOI: 10.1080/17686733.2025.2592191.
2. Jonschkowski R, Stone A, Barron J T, et al. What matters in unsupervised optical flow. European conference on computer vision. Cham: Springer International Publishing, 2020: 557-572.
3. Fleuret J, Ebrahimi S, Castanedo C I, et al. On the use of pulsed thermography signal reconstruction based on linear support vector regression for carbon fiber reinforced polymer inspection. Quantitative InfraRed Thermography Journal, 2023, 20(2): 39-61.
