Intelligent recognition of defects in high‐speed railway slab track with limited dataset

• Published in: Computer-aided civil and infrastructure engineering 2024-03, Vol.39 (6), p.911-928
• Main Authors: Cai, Xiaopei, Tang, Xueyang, Pan, Shuo, Wang, Yi, Yan, Hai, Ren, Yuheng, Chen, Ning, Hou, Yue
• Format: Article
• Language: English
• Subjects: Accuracy; Artificial intelligence; Concrete slabs; Datasets; Defects; Engineers; High speed rail; Lightweight; Machine learning; Maintenance; Mathematical models; Model accuracy; Parameters; Public transportation; Service life; Testing time; Transportation safety
• ISSN: 1093-9687; 1467-8667
• DOI: 10.1111/mice.13109

During the regular service life of high‐speed railway (HSR), there might be serious defects in the concrete slabs of the infrastructure systems, which may further significantly affect public transportation safety. To address these serious issues and fulfill the regular functions of HSR, the traditional methods for railway engineers involve carrying out regular on‐site inspections manually or by semi‐automatic inspection vehicles, and conducting timely corresponding repairing approaches and maintenance, where these methods are time‐consuming and dangerous. In recent years, machine learning methods have been widely applied to the intelligent and automatic detection of severe defects in HSR. Currently, one of the most serious problems is the lack of sufficient high‐quality data for model training, resulting in low recognition accuracy in HSR defects. To solve this problem, this paper proposed an intelligent recognition of defects in concrete slabs of HSR based on a few‐shot learning model, that is, an artificial intelligence model based on limited data size, which recognizes three service conditions of concrete slabs in HSR: cracks, track board gaps, and unbroken state. Lightweight few‐shot learning models specifically designed for HSR detection were proposed. Experiments were conducted to compare the performances of different lightweight‐designed models, including accuracy, parameter quantity, and testing time. Results showed that the optimum model can fast and satisfactorily recognize the defects in HSR with a very limited data size of 10 samples for each training category, with a satisfactory accuracy of 73.9% in the test dataset with 20 samples for each category, parameter amounts of 2.8 million, and a testing time of 2.2 s per image. This study provides a reference for the automatic recognition of defects in HSR by railway engineers with insufficient samples.