Intelligent recognition and measurement for fatigue cracks in orthotropic steel decks: A comparative study of algorithms

Abstract

Orthotropic steel bridge decks are highly susceptible to fatigue cracking under cyclic loading. To overcome the limitations in precision and efficiency of existing crack identification technologies, this study systematically analyzed the performance of four traditional machine learning algorithms, twelve deep learning algorithms, and forty-eight hybrid algorithms with various combinations in the task of fatigue crack image recognition. An innovative comprehensive evaluation method was proposed to assess these algorithms. It integrated an intelligent crack image recognition method with an automatic crack measurement technology to design an application. The results indicate that the combination of deep learning algorithms with K-Nearest Neighbors (KNN) or Support Vector Machines (SVM) algorithms significantly enhances recognition precision and efficiency. Among all the algorithms tested, the ResNet-101+SVM hybrid algorithm achieved the highest score of 98.88, with a final test set accuracy of 98.95 %, a transfer recognition rate of 95.97 %, and a training time of only 20.15 s, demonstrating excellent stability. Furthermore, it has successfully developed a low-cost and high-precision crack measurement technology, capable of controlling measurement errors within 2 %. It offers an efficient and reliable tool for intelligent detection and assessment of fatigue cracks in orthotropic steel bridge decks for engineers and researchers.