Leakage Identification of Underground Structures Using Classification Deep Neural Networks and Transfer Learning

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Water leakage in underground structures accelerates aging and compromises safety.
• Early detection of leakage is crucial for structural maintenance and reinforcement.
• Deep learning models require extensive data, which is costly to acquire for leakage identification.

Purpose of the Study:

• To develop a deep neural network method for underground structure leakage identification using transfer learning.
• To overcome the challenge of limited training data in deep learning for this application.
• To compare the performance of different deep learning models with and without transfer learning.

Main Methods:

• A transfer learning strategy was developed for deep neural network models.
• A dataset of underground structure leakage was created for training and evaluation.
• Four classification models (VGG16, AlexNet, SqueezeNet, ResNet18) were constructed and trained.
• Classification performance was comparatively studied under varying training data sizes.

Main Results:

• Transfer learning enhanced the classification performance and stability of VGG16, AlexNet, and SqueezeNet.
• ResNet18 with transfer learning showed similar performance but improved stability.
• SqueezeNet achieved the highest and most stable performance across all metrics.

Conclusions:

• Transfer learning is effective in improving deep learning models for underground structure leakage identification, especially with limited data.
• SqueezeNet is a highly effective model for this task, offering superior performance and stability.
• The findings provide valuable insights for developing robust leakage detection systems.