融合密集卷积网络和迁移学习的张拉膜结构膜面损伤智能识别分类

Intelligent Identification and Classification of Membrane Damage in Tensile Membrane Structures Based on Dense Convolutional Networks and Transfer Learning

摘要: 膜面破损容易引发张拉膜结构发生瞬时破断和溃塌，针对传统方法依赖人工目视检测普遍存在主观性强、效率低等问题，本文提出一种膜面损伤识别分类的密集卷积网络深度学习框架(DenseNet121-ECA-PTL)。首先，以自主构建的典型张拉膜试验平台为基础，建立包含5种典型膜面损伤类型的张拉膜结构膜面损伤图像数据集；其次，为提升对早期微小损伤的表征能力，在DenseNet121模型密集连接块中嵌入高效通道注意力模块，增强模型对损伤区域关键特征的聚焦能力；最后引入迁移学习策略优化模型收敛性能，并进行了模型对比试验用以验证模型改进效果。结果表明：本文提出的DenseNet121-ECA-PTL模型有效提升了基础模型在膜面细微特征高效提取能力，在测试集上的识别准确率达到98%，单张图像推理耗时仅为0.04 s。该算法在保证高精度的同时具备优异的推理效率，为张拉膜结构膜面损伤的智能化监测与移动端实时部署提供坚实的技术支持。

Abstract: Membrane damage can easily lead to instantaneous rupture and progressive collapse of tensile membrane structures. Traditional methods relying on manual visual inspection are commonly characterized by strong subjectivity and low efficiency. This study proposes a dense convolutional network-based deep learning framework (DenseNet121-ECA-PTL) for membrane damage identification and classification. Firstly, the dataset containing five typical types of membrane damages was constructed based on the typical tensile membrane structure experimental platform. Secondly, an Efficient Channel Attention (ECA) module was embedded into the dense blocks of DenseNet121 to enhance the representation capability for early-stage micro-damage. Finally, a transfer learning strategy was introduced to improve convergence performance, and comparative experiments were conducted to validate the effectiveness of the proposed model. The results demonstrate that the DenseNet121-ECA-PTL improves the feature extraction capability of the baseline model for subtle damage characteristics. The proposed model achieves an accuracy of 98% on the test dataset, with an inference time of 0.04 s per image. Furthermore, the proposed model exhibits excellent computational efficiency while maintaining high accuracy, providing strong technical support for intelligent monitoring and real-time mobile deployment of membrane damage detection in tensile membrane structures.