Abstract
As railway traffic volumes and train speeds increase, rail maintenance is becoming more crucial to prevent catastrophic failures. This study developed an artificial intelligence (AI)-based solution for automatic rail flaw detection using ultrasound sensors to overcome the limitations of traditional inspection methods. A hierarchical classification model was proposed, integrating ultrasound B-scan images and machine learning. The two-stage approach—Model A for fuzzy classification followed by Model B for fine-grained diagnosis—utilized the EfficientNet-B7 architecture. Experimental results demonstrated the model's ability to accurately detect rail flaws, achieving 88.56% accuracy and a processing speed of 0.45 s per ultrasound image sheet. This AI-based solution shows promise for automated, rapid, and reliable rail flaw detection to support safer railway infrastructure maintenance.
The Challenge of Internal Rail Flaws
As global railway networks expand and train speeds rise, ensuring the structural integrity of rails is paramount. Internal flaws, such as transverse cracks, bolt hole cracks, and rail bottom defects, are often invisible to the naked eye but can lead to catastrophic derailments if left undetected.
Traditionally, the industry has relied on "rail-walking" with manual ultrasonic tools or human interpretation of B-scan data. These methods are labor-intensive, slow, and prone to subjective human error. While automated inspection vehicles (like the one shown below) collect vast amounts of data, the challenge lies in processing this information accurately and rapidly.
A Hierarchical AI Approach
To address these challenges, researchers developed a hierarchical classification framework designed to mimic the logical progression of a human expert. The system employs a two-stage cascading model:
- Primary Classification (Model A): This stage performs "fuzzy" classification, identifying the broad category of the flaw (e.g., transverse flaw, bolt hole crack, or rail bottom defect).
- Fine-grained Classification (Model B): Once the broad category is identified, this stage provides a detailed diagnosis, specifying the subtype and severity of the flaw along with a confidence score.
After benchmarking several state-of-the-art architectures—including ResNet, Swin Transformer, and Conformer—the researchers identified EfficientNet-B7 as the most effective "engine" for this hierarchy. Despite the current popularity of Transformers in computer vision, the convolutional neural network (CNN) architecture of EfficientNet-B7 provided a superior balance of accuracy and computational efficiency for the distinct, line-based patterns found in ultrasonic B-scans.
Dataset and Training
The model was trained and tested using a robust dataset of severe rail damage records collected between 2012 and 2022. The data was categorized into four main groups based on international standards:
- D1: Rail transverse flaws (further divided by location: inside, central, outside).
- D2: Bolt hole cracks.
- D3: Rail bottom flaws.
- D4: Other flaws (including horizontal and diagonal cracks).
The dataset was split into a 4:1 ratio for training and testing to ensure the model's ability to generalize to new, unseen data.
Performance and Key Results
The hierarchical EfficientNet-B7 model demonstrated significant improvements over standalone architectures:
- High Accuracy: The system achieved an overall accuracy of 88.56%, significantly outperforming standalone EfficientNet-B7 (83.64%), ResNet (78.45%), and Conformer (76.36%).
- Rapid Processing: The model can analyze a single ultrasound image sheet in just 0.45 seconds, making it suitable for real-time or near-real-time applications.
- Targeted Precision: The model was particularly effective at identifying severe bolt-hole cracks and rail bottom flaws, with the latter reaching nearly 98% accuracy.
Future Directions
While the results are promising, the authors highlight several areas for future work. Expanding the dataset to include more diverse geographical regions will improve the model's robustness against environmental variables like temperature and humidity. Additionally, there is a focus on "Green AI," balancing model complexity with energy consumption. The ultimate goal is the seamless integration of this AI framework into fully autonomous track maintenance systems, including robotics and drones.
Summary updated