Wavelet Scattering and Neural Networks for Railhead Defect Identification.

Area of Science:

• Railway Engineering
• Computer Vision
• Machine Learning

Background:

• Railway operational safety relies on accurate railhead inspection.
• Deep learning methods for railhead defect detection face challenges with data requirements and defect size information.
• Existing methods often overlook critical defect size and location data.

Purpose of the Study:

• To develop a machine learning framework for accurate and automatic railhead defect identification.
• To address limitations of current deep learning approaches regarding data intensity and defect size.
• To improve the reliability and applicability of automated railhead inspection systems.

Main Methods:

• A machine learning framework combining wavelet scattering networks (WSNs) and neural networks (NNs) was developed.
• WSNs, parameter-free and akin to deep convolutional neural networks, were used for feature extraction from non-intensive datasets.
• Neural networks were employed to restore defect location and size information.

Main Results:

• The framework achieved high validation accuracies of 99.80% (Type-I) and 99.44% (Type-II) on the RSDD dataset.
• Pixel-level analysis showed precision, recall, and F-measure around 90%, outperforming previous methods.
• Defect-level analysis yielded 100% recall, with precision around 75%, effectively identifying all defects.

Conclusions:

• The developed WSNs and NNs framework is highly effective for identifying railhead defects.
• The approach overcomes data intensity limitations and accurately restores defect size and location.
• This method significantly enhances automated railhead inspection for improved railway safety.