Intelligent characterization and robustness quantification of frozen soil strength images using a multi-module fusion strategy

Abstract

In frozen wall engineering, traditional detection methods are difficult to implement for real-time monitoring of strength parameters, due to their intermittent nature. This makes it difficult to provide timely warnings and effective responses to potential risks of frozen walls, highlighting the urgent need for a method that can continuously and accurately monitor the strength of frozen walls. An image data-driven intelligent identification method for frozen soil strength based on convolutional neural networks is proposed. By capturing images of cured samples from multiple angles and conducting uniaxial compressive strength tests, the collected strength data were divided into 12 categories, creating an image dataset for deep learning model training. A Resnet-34 deep learning model combined with global attention and downsampling showed excellent performance in comparison with a series of basic models, with an accuracy of 93.3 % and no overfitting. The accuracy of the deep learning model was assessed using adversarial attack and defense strategies, serving as an indicator of robustness. The improved model exhibited better robustness in comparative analyses. Using SHapley Additive exPlanations (SHAP) value analysis, the feature extraction process of the convolutional neural network in recognizing frozen soil strength was investigated and clarified, confirming the model's ability to identify and extract key features in frozen soil images, such as particle size, texture, cracks, and ice crystal distribution patterns. This technology offers a cutting-edge method for real-time tracking of frozen wall conditions and early disaster warnings, significantly enhancing the safety and success rate of construction projects under freezing conditions.