Session: Research Posters

Paper Number: 173553

Rail Guard Ai: A Hybrid Deep Learning Architecture for Real-Time Railway Defect Classification 

Abstract:

Railway infrastructure maintenance is critical for safety assurance and continuity of operation; yet, conventional defect detection techniques continue to be time consuming, expensive, and error-prone. To overcome the above problems, we present Rail Guard AI, named Hybrid RailNet, a new hybrid deep learning model that fuses ResNet, Inception blocks, and a Self-Attention mechanism to enhance the detection of railway defects in real time. The system uses a four-branch processing pipeline capturing fine, medium, large-scale, and global contextual features, thus enabling strong detection over various defect scales. Trained on an augmented dataset, Hybrid RailNet achieved a binary classification accuracy of 96.9%. Among the classifiers tested, the Medium Coarse Gaussian Support Vector Machine (MSVM) and the Linear Support Vector Machine (LSVM) performed the best. The results indicate the promise of Hybrid RailNet to play a valuable role in automated railway inspection systems through enhanced detection accuracy, shorter inspection times, and better reliability compared to conventional techniques. 

Introduction:

The integrity of rail infrastructure has a direct impact on public safety, reliability of services, and economic efficiency of national transport networks. As rail traffic increases and infrastructure ages, the need for automated and intelligent defect detection systems has never been more urgent.

Although conventional inspection methods are based on manual inspection or vision-based systems, they prove to be insufficient for acquiring small-scale defects, particularly in the presence of varying environmental conditions. Such methods are not only time consuming and expensive, but also lack consistency because of human subjectivity. Recent advances in CNNs have improved the capacity of machines to recognize visual patterns. A number of studies have shown the success of CNNs in the detection of rail surface defects [1][2][3].

One major limitation consistently referenced throughout the literature is the reliance on single-stream architectures, which tend to process images at a single fixed resolution. Such models have difficulty preserving performance across varying defect scales, resulting in low recall and high false-negative rates. Recent research indicates the potential of incorporating multi-scale feature extraction and attention mechanisms to mitigate these issues. To overcome these difficulties, we present Hybrid RailNet a multi-branch network architecture that aims to extract fine, medium, and large-scale defect features through the fusion of ResNet, Inception blocks, and Self-Attention. Our network seeks to improve and extend the limitations of previous studies by allowing more adaptive and robust binary defect classification under real-world rail environments. 

Methodology:

The development of the Rail Guard AI system followed a structured methodology approach which consist of constructing, training, evaluation, and validating a model using real-world railway defects datasets. This methodological approach was designed to develop a model that made accurate detections on a variety of scales of defect, while being computationally efficient overall.

1. Model Design and Feature Extraction

The core model, named Hybrid RailNet, was developed to extract deep, scale-aware, and context-rich features from input rail images. The Hybrid RailNet consisted of the following components:

1. Residual Networks (ResNet) for extracting deep features.

2. Inception modules to examine spatial features at different resolutions in parallel, and

3. A Self-Attention mechanism which dynamically concentrates the network's attention on key defect areas.

Additionally, Hybrid RailNet uses a four-branch feature extraction pipeline, aimed at capturing:

1. Fine-grained texture for early-stage cracks,

2. Medium-scale geometries such as flaking and grooves,

3. Gross structural deformations such as squats and shelling.

4. Global context via attention-weighted integration.

This end-to-end design not only boosts classification accuracy but also the interpretability and robustness of the model.

After the Hybrid RailNet feature extraction was completed a group of five classifiers was used for classification: 1-  Linear SVM / 2-  Quadratic SVM/ 3-  Cubic SVM / 4-  Medium Coarse SVM/ 5-  Gaussian SVM

Results:

All classifiers were tested under the same conditions using the same features that were extracted.

The classifiers were evaluated based on: Standard Metrics: Accuracy, Recall, Precision, Specificity, F1-score Advanced Metrics: MCC, Cohen’s Kappa Execution Time: Total inference time on a fixed test set ROC Analysis: AUC per class and total AUC. 

To assess the performance of the proposed Hybrid RailNet in binary railway defect classification, we conducted a series of experiments using features extracted by the network that were then classified by five classifiers based on SVM. The results of our testing are provided in Table 1 and illustrated in Figures 2 and 3. As highlighted in Table 1, the Linear SVM and Medium Coarse SVM classifiers were superior to all other classifiers, achieving the following results:

·       Accuracy: 96.97% / Recall: 95.76% / F1-score: 0.9700 / Precision: 98.26% / MCC/Kappa: 0.9397

·       Inference time: 2.27 s (Linear), 2.10 s (Med Coarse – fastest with an inference time in the order of seconds).

The confusion matrix and the scatterplot show a clear separation between defective and non-defective samples, confirming high precision and distinction between the two classes.

The model predicted the defective and non-defective cases with full confidence (score = 1.00) for both subtle defects and severe defects, demonstrating that the model is ready for deployment. 

Conclusion:

This research introduced Rail Guard AI, an intelligent defect detection framework powered by the proposed Hybrid RailNet architecture.

The model combines ResNet, Inception modules, and a Self-Attention mechanism, with the intention to learn fine-to-coarse visual features and use an adaptive focus for defect-critical areas in railway images.

The model was evaluated using five classical classifiers trained on features learned from Hybrid RailNet. The Linear SVM and Medium Coarse SVM classifiers represented the best results in the binary classification context, achieving: 96.97% accuracy, 97.00% F1-score, 95.76% recall, Overall precision > 98%, Inference time of approximately 2.1 seconds, allowing for real deployment of the model.

The model demonstrated robustness over presenting defects with variances in defection feature presentations such as surface cracking, joint misalignment, and subtle signs of structural degradation providing reliable defect detection in a "real-world" context under railway safety monitoring. In addition to performance, the System architecture allows for scalability supporting incorporation of existing railway inspection workflows. The models high AUC scores and clarity in the confusion matrix demonstrates its effectiveness in reducing false positives and guarantees detection reliability

Presenting Author: Jinan Charafeddine ESILV École d'ingénieurs Généraliste

Presenting Author Biography: Dr. Jinan Charafeddine is a researcher at the De Vinci Research Center in Paris, France. She holds a Ph.D. in Biomedical Engineering and specializes in the fields of artificial intelligence, medical robotics, computer vision, and healthcare data analysis. Her research explores the intersection of engineering and medicine, with a focus on developing intelligent systems for diagnostics, secure data communication, and robotic-assisted interventions.

Dr. Charafeddine has authored several peer-reviewed publications in international conferences and journals, and she actively collaborates with multidisciplinary teams across Europe and the Middle East. Her recent work includes AI-based solutions for medical image segmentation, blockchain for IoT healthcare applications, and signal processing in biomedical systems. She is passionate about leveraging cutting-edge technology to solve real-world problems in healthcare and beyond.

Authors:

Taha Houda Prince Mohammad Bin Fahd Univetsity
Jinan Charafeddine ESILV École d'ingénieurs Généraliste