Key Takeaways & Executive Findings
- •• • Achieves 99.1% average detection accuracy, a 4.29% improvement over baseline YOLOv5, directly reducing false negatives in safety-critical wheelset inspections. • • Maintains 15 ms per image inference speed, enabling real-time deployment on edge devices for continuous subway tread monitoring without operational delays. • • Outperforms YOLOv12 in convergence speed, detection accuracy, and inference speed, demonstrating that targeted architectural modifications can surpass newer generic detectors. • • Integrates W-MPDIoU loss function to overcome limited labeled data and annotation inaccuracies, accelerating model convergence and improving robustness under data-scarce industrial conditions.
China Advanced Materials & Deep-Tech Radar
Get verified English translations, SEM micrographs & open-access PDF alerts from China's leading state key laboratories delivered to your inbox every Monday at 08:00 EST.
Abstract
Wheelset tread defects in subway locomotives present critical safety hazards, yet manual inspection remains prevalent, suffering from inefficiency and human error. This study proposes an enhanced YOLOv5-based detection algorithm tailored for subway wheelset tread defects. A multi-head self-attention module is integrated to capture long-range dependencies within global feature maps, improving small-target detection. A weighted bidirectional feature pyramid network (BiFPN) enables balanced multi-scale feature fusion and efficient cross-scale integration. To mitigate limited labeled data and annotation inaccuracies, a novel loss function, W-MPDIoU, is introduced to accelerate convergence. Experimental validation using real defect data and simulated experimental data yields an average detection accuracy of 99.1%, a 4.29% improvement over the original YOLOv5, with a detection speed of 15 ms per image. The model also outperforms YOLOv12 in convergence speed, detection accuracy, and inference speed. Despite these gains, limitations persist in defect variety and dataset size, necessitating further refinement for broader generalization. The proposed method enables real-time tread defect detection, enhancing safety and operational efficiency in urban rail transit maintenance.
1. Introduction
Subway wheelset tread defects—fatigue cracks, delamination, abrasion, and foreign object indentation—arise from continuous wheel-rail contact and directly compromise operational safety. Current maintenance practices rely predominantly on manual visual inspection, a method that is inefficient, error-prone, and incapable of meeting the demands of intelligent, high-frequency urban rail service. While machine learning approaches have been explored, unsupervised clustering methods suffer from indeterminate cluster counts and poor robustness to noise, temperature fluctuations, and track-state variations inherent to subway environments.
This study addresses these bottlenecks by enhancing the YOLOv5 framework with a multi-head self-attention module for long-range dependency capture, a weighted bidirectional feature pyramid network for balanced multi-scale fusion, and a novel W-MPDIoU loss function to counteract limited labeled data and annotation inaccuracies. The resulting model achieves 99.1% average precision at 15 ms per image, surpassing both the original YOLOv5 and YOLOv12 in convergence, accuracy, and inference speed, thereby delivering a viable real-time detection solution for subway wheelset maintenance.
Loading authentic research manuscript (Pages 1–5)...
WANG Jindong, XIE Chengsheng, LIU Tao, ZHOU Haonan, LI Ao (2026). A Novel Method for Subway Wheelset Tread Defect Detection with Improved Self-Attention and Loss Function. Railway Engineering Science (铁道工程科学). https://doi.org/10.1007/s40534-025-00415-2
Research & Educational Purpose Only: The translations, structured abstracts, analytical annotations, and data reports provided by SinoTechIntelare intended exclusively for academic research, internal corporate R&D, and educational benchmarking. They do not constitute formal engineering, chemical safety, legal, or professional advice.
Copyright & Intellectual Property Notice: Original copyright of the underlying source articles and experimental data remains with the respective authors, institutions, and original publishing journals. SinoTechIntel claims intellectual property only over its proprietary translations, analytical syntheses, and AEO structured enhancements in accordance with international fair use and academic citation principles.
Frequently Asked Questions
What specific failure mechanisms in subway wheelsets does the proposed detection method target, and how does it perform under varying operational conditions?
The method targets fatigue cracks, delamination, abrasion, and foreign object indentation on tread surfaces. It achieves 99.1% average precision across these defect types, with a 4.29% improvement over baseline YOLOv5, and maintains 15 ms per image inference speed, ensuring reliable detection despite noise and temperature fluctuations typical of subway environments.
How does the W-MPDIoU loss function address annotation inaccuracies and limited labeled data, and what convergence benefits are observed?
W-MPDIoU redefines the loss to accelerate model convergence by mitigating the impact of imprecise bounding boxes and scarce labels. Experimental results show faster convergence compared to YOLOv12, contributing to the 99.1% accuracy and enabling effective training with real and simulated defect data.
What are the scalability bottlenecks for deploying this model across diverse subway networks with varying defect types and dataset sizes?
The current model is limited by the variety of detectable defects and dataset size, as acknowledged in the conclusions. While it achieves 99.1% accuracy on the existing dataset, generalization to broader defect categories requires expanding the dataset and refining defect types, which may impact inference speed and accuracy trade-offs.
How does the improved YOLOv5 model compare to YOLOv12 in terms of key performance indicators, and what architectural modifications drive these differences?
The improved YOLOv5 outperforms YOLOv12 in convergence speed, detection accuracy (99.1% vs. unspecified YOLOv12 baseline), and inference speed (15 ms per image). These gains stem from the multi-head self-attention module for long-range dependencies, weighted BiFPN for multi-scale fusion, and W-MPDIoU loss for accelerated convergence.
What are the computational and parameter requirements for real-time deployment, and how do they affect integration into existing subway maintenance systems?
The model achieves 15 ms per image detection speed with reduced parameters compared to the original YOLOv5, enabling real-time operation on edge devices. This efficiency facilitates integration into intelligent maintenance workflows without significant hardware upgrades, though further optimization may be needed for large-scale deployment across multiple inspection points.
Related Chinese Research & Cross-Citations
Harmonic Impedance Studies of Swedish Railway Power System Using Wide-Area Modeling Approach
This paper presents a comprehensive framework for wide-area modeling and harmonic resonance assessment in the AC single-phase Swedish electric railway power system (ERPS). The methodology integrates the modeling of key elements of the catenary system, such as synchronous generators, transmission lines, transformers, and filters, while addressing the dynamic behavior of rolling stocks and inherent system uncertainties. Study cases, including Monte Carlo simulations, are developed to evaluate probabilistic scenarios and impedance variations across the network using nodal admittance modeling and frequency scanning. Key contributions include a method to model moving loads, a comprehensive approach to harmonic resonance analysis based on meshed grid characteristics of the ERPS, and an uncertainty assessment framework that highlights insights for system planning and mitigation actions. Results indicate that filters installed at traction converter stations (TCS) introduce resonances typically between the 10th and 25th harmonics, with amplification factors exceeding five times at certain locations. The wide-area modeling approach is essential due to continuous power supply and influence of distant network sections. Uncertainties related to moving loads, number of rolling stock units, and remote system areas cause variations in both frequency and magnitude of series and parallel resonances. Monte Carlo simulations effectively capture these variations, providing probabilistic information for resonance frequency location and impedance variation. The findings offer critical insights for system planning and mitigation actions in ERPS.
Drive-by interlayer damage detection methodology for heavy-haul railway bridge using axle box acceleration
Interlayer degradation in heavy-haul railway (HHR) bridges under rising axle loads and transport volumes threatens structural safety. Traditional visual inspection and fixed-sensor structural health monitoring are impractical for large bridge inventories. This paper proposes a drive-by inspection methodology that combines vertical axle box acceleration (ABA) with hybrid filtering for rapid interlayer damage detection in multi-span HHR bridges. The framework introduces a Hilbert-transform-based instantaneous amplitude quartic index (IAQI) to enhance damage localization accuracy. The hybrid filtering integrates bandpass filtering targeting sleeper-passing frequency components to suppress track irregularity effects, and a statistical diagnostic tool to discriminate interlayer damage from sleeper-related driving components. Numerical analyses and a field test on an 18-span, 609.5-m simply supported HHR bridge validate the method. Results demonstrate effective detection under combined beam damage, irregularity, and noise. The field test identified five interlayer damage locations requiring on-site confirmation. The method offers a new strategy to improve inspection efficiency and ensure operational safety of HHR bridges.
Publisher Correction: Bayesian Multivariate Track Geometry Degradation Modeling and Its Use in Condition-Based Inspection
This publisher correction addresses a typesetting error in Figure 18 of the original article 'Bayesian multivariate track geometry degradation modeling and its use in condition-based inspection' published in Railway Engineering Science. The correction notice, published online on 3 December 2025, provides both the incorrect and correct versions of Figure 18, which is central to the visualization of track geometry degradation predictions. The original article, identified by DOI 10.1007/s40534-025-00394-4, presented a Bayesian framework for modeling multivariate degradation of track geometry parameters—including gauge, crosslevel, alignment, and profile—to support condition-based inspection scheduling. The erroneous figure compromised the interpretation of posterior predictive distributions and inspection thresholds. The corrected figure restores the accurate representation of degradation trajectories and associated uncertainty intervals, ensuring that maintenance decisions derived from the model remain valid. This correction is critical for railway asset managers who rely on the model's outputs to optimize inspection intervals and reduce lifecycle costs. The authors and publisher affirm that the scientific conclusions of the original work remain unchanged. The correction is published under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, and the original article has been updated accordingly.
Multiscale Investigation on Fatigue Crack Growth and Remaining Useful Life of Bogie Frame Materials Under Service-Induced Damage
This study quantifies the degradation of fatigue crack growth (FCG) resistance in high-speed train bogie frame materials after long-term service. Full-scale frame fatigue tests, multiaxial FCG experiments, and finite element simulations were integrated to determine equivalent crack loading conditions. Digital image correlation captured surface displacement fields for stress intensity factor calculation. Comparative testing of as-welded (AW) and base metal (BM) regions before and after service revealed substantial reductions in remaining useful life: 70.54% for AW and 22.31% for BM. Crack-tip strain responses increased by more than twofold in AW and 1.44 times in BM after service, indicating diminished crack growth resistance. Microscopic fracture surface analysis showed more secondary cracks, unstable crack paths, and blurred fatigue striations in post-service materials, particularly in the AW region. Phased array ultrasonic testing detected no macroscopic defects, yet microstructural deterioration was evident. These findings establish a quantitative link between service-induced damage and fatigue performance degradation, supporting region-specific residual life assessment strategies for bogie frames. The experimental protocol, grounded in actual service loading spectra, improves the accuracy of remaining useful life prediction and provides a reliable basis for maintenance decision-making in high-speed rail operations.
Defects detection for railway catenary system with encoder-decoder architecture
This study presents a robust and efficient damage detection methodology for railway catenary systems, employing an encoder-decoder architecture supplemented by residual analysis. A novel signal segmentation strategy based on catenary structural features is introduced, coupled with a quasi-Welch method to mitigate edge effects. The investigation examines the impact of GPS inaccuracies on detection precision and conducts a comprehensive analysis of normalization techniques and their effects on defect identification. Two primary defect types are considered: hard points in the contact wire and periodic short-wavelength irregularities (PSWI), with variations in train speeds and defect magnitudes. A defect detection criterion is developed, enabling rapid and automatic identification of catenary defects. The integrated approach facilitates effective detection and accurate localization, overcoming limitations of previous methods such as the requirement for high sampling frequency. This work advances catenary inspection methodology and contributes to enhancing railway safety and reliability. The innovation lies in integrating the reconstruction capabilities of the encoder-decoder architecture with a residual-based defect detection method, allowing complementary features to synergistically improve detection performance.
Study on Air Quality Improvement in Train Compartments by Coordinated Exhaust of Multiple Air Outlets
Ventilation is one of the most effective ways to improve the air quality in trains. Top exhaust and bottom exhaust are two commonly used modes. The study hopes to switch the exhaust mode to adapt to the indoor requirements of different scenes without changing the layout of the end pipe. In the study, the airflow characteristics, energy consumption, thermal comfort and air quality in the compartment are evaluated based on computational fluid dynamics. The results show that the energy consumption decreases with the increase of the top exhaust air volume in summer conditions, while the opposite is true in winter conditions. In terms of thermal comfort, combining top and bottom exhaust can effectively improve air speed index, temperature difference index and air diffusion performance index. In addition, the draft rate index and percent dissatisfied index are less than 10% and 3%, respectively, which meet the requirements of ISO 7730 standard. In terms of air quality, the average pollutant concentration inside the vehicle decreased, but the longitudinal penetration capacity of the pollutant has increased. The research results can provide some suggestions and help for the ventilation design of high-speed trains.