Intelligent Damage Classification for Tensile Membrane Structures Based on an Improved DenseNet121 with Attention Mechanism and Transfer Learning

Membrane materials are crucial components that maintain the stability of tensile membrane structures, and damaged membrane material can readily trigger sudden rupture and collapse. To address the strong subjectivity and low efficiency of manual visual inspection methods, this paper proposes an improved dense convolutional classification network framework for damaged membrane material images, namely, DenseNet121-ECA-PTL. Firstly, two tensile membrane test platforms were independently constructed, and 10,000 damaged membrane material images covering five categories (tear, wrinkle, smudge, screen, and scratch) were collected by using smartphones and UAVs to establish a custom dataset. Secondly, an ECA-enhanced DenseNet121 model combined with transfer learning was then developed to improve complex feature extraction, convergence, and generalization, and its performance was validated through comparative and ablation experiments. The proposed model achieved overall testing accuracy, precision, recall, specificity, F1score, and Matthews correlation coefficient of 98.87%, 98.69%, 98.69%, 98.68%, 99.67%, and 98.35%, respectively. The minimum average prediction time of the proposed model was 0.0472 s for each damaged membrane material image. Finally, model interpretability analyses based on T-SNE and Grad-CAM were used to comprehensively examine the significant feature learning and attention-focusing capabilities of the proposed model for multiple damage types of membrane material images.