Abstract
Intelligent fault diagnosis of high-speed train bogies faces significant challenges as speeds increase: diagnostic models trained on lower-speed data often fail at high velocities, and labeled fault data at high speeds is difficult to obtain. This paper presents a framework for reliably diagnosing unlabeled signals at varying speeds using limited labeled samples through deep transfer learning. The study involves a real-train experiment on a rolling test bed to obtain fault data across a range of speeds (50–350 km/h). By analyzing distributional differences and cross-speed transfer performance, the proposed model predicts the optimal source domain for high-speed monitoring. The feasibility and stability of the method are verified across various signal processing techniques and network architectures.
Introduction
As high-speed trains (HSTs) push the limits of velocity, the mechanical signatures of their components change. A diagnostic model trained to recognize a failing bearing at 100 km/h often fails at 300 km/h because the vibration patterns—the "features" the AI looks for—shift significantly. The core problem is a lack of labeled data; while it is relatively easy to collect and label fault data at lower speeds in controlled environments, doing so at 350 km/h is expensive and rare. This creates a "label-less" problem for high-speed operations. Engineers need a way to take the knowledge gained at lower speeds (the source domain) and apply it to higher speeds (the target domain) without needing a new set of labels for every incremental speed increase.
Proposed Methodology
The authors propose an adaptive transfer learning strategy composed of three main stages:
- Feature Distribution Difference Calculation: A 1D Convolutional Neural Network (CNN) extracts features from vibration signals. The "distance" between these features across different speeds is measured using metrics like Maximum Mean Discrepancy (MMD), Wasserstein Distance (WD), and Cosine Similarity.
- Cross-Transfer Learning: The performance of models trained at one speed is tested at another across various transfer learning (TL) algorithms, including Deep Adaptation Network (DAN), Domain-Adversarial Neural Network (DANN), and Batch Nuclear-norm Maximization (BNM).
- Ensemble Learning Mapping: Gradient Boosting Regression fits the relationship between distribution differences and actual diagnostic accuracy, allowing the system to predict transfer performance before deployment.
Experimental Setup
The study is grounded in high-fidelity experimental data rather than simulations. The platform utilizes a real high-speed train car on a full-scale rolling test bed, employing NTN CRI-2692 double-row tapered roller bearings. Simulated "real-world" damages, including outer ring indentation, peeling, and corrosion, were introduced. Vibration data was sampled at 25.6 kHz across seven distinct speed conditions ranging from 50 km/h to 350 km/h.
Results and Discussion
Transferability and Speed Discrepancy
The research reveals that "transferability" between speeds is non-intuitive. One might assume that a 250 km/h model is naturally the best fit for 300 km/h data, but complex mechanical resonances can mean a much lower speed provides a better feature set for a high-speed target.
Performance Analysis
The BNM (Batch Nuclear-norm Maximization) method emerged as the most robust, often achieving near 100% accuracy when transferring between various speed domains using envelope spectrum data. The proposed framework successfully predicted the "optimal source domain." For example, when diagnosing a 350 km/h target, the model correctly identified that transferring from the 200 km/h or 300 km/h source would yield the highest accuracy. The framework also proved stable across different bogie structures.
Key Insights
- Envelope Spectrum: Using envelope spectral data consistently outperformed raw time-domain data in transfer tasks.
- Negative Transfer: Some popular methods, like "DeepCoral," suffered from "negative transfer," where the transfer process actually degraded the model's performance compared to simpler approaches.
Conclusion and Future Work
This study demonstrates that deep transfer learning can effectively bridge the gap between low-speed labeled data and high-speed unlabeled operating conditions in rail maintenance. While successful, future work will address the computational costs of finding optimal source-target matches and the real-time constraints for on-board deployment. Further theoretical investigation into the uncertainty of transfer between highly disparate conditions is also required.
Summary updated