PD-SOVNet: A Physics-Driven Second-Order Vibration Operator Network for Estimating Wheel Polygonal Roughness from Axle-Box Vibrations
arXiv:2604.06620v1 Announce Type: new Abstract: Quantitative estimation of wheel polygonal roughness from axle-box vibration signals is a challenging yet practically relevant problem for rail-vehicle condition monitoring. Existing studies have largely focused on detection, identification, or severity classification, while continuous regression of multi-order roughness spectra remains less explored, especially under real operational data and unseen-wheel conditions. To address this problem, this paper presents PD-SOVNet, a physics-guided gray-box framework that combines shared second-order vibration kernels, a $4\times4$ MIMO coupling module, an adaptive physical correction branch, and a Mamba-based temporal branch for estimating the 1st--40th-order wheel roughness spectrum from axle-box vibrations. The proposed design embeds modal-response priors into the model while retaining data-driven flexibility for sample-dependent correction and residual temporal dynamics. Experiments on three
arXiv:2604.06620v1 Announce Type: new Abstract: Quantitative estimation of wheel polygonal roughness from axle-box vibration signals is a challenging yet practically relevant problem for rail-vehicle condition monitoring. Existing studies have largely focused on detection, identification, or severity classification, while continuous regression of multi-order roughness spectra remains less explored, especially under real operational data and unseen-wheel conditions. To address this problem, this paper presents PD-SOVNet, a physics-guided gray-box framework that combines shared second-order vibration kernels, a $4\times4$ MIMO coupling module, an adaptive physical correction branch, and a Mamba-based temporal branch for estimating the 1st--40th-order wheel roughness spectrum from axle-box vibrations. The proposed design embeds modal-response priors into the model while retaining data-driven flexibility for sample-dependent correction and residual temporal dynamics. Experiments on three real-world datasets, including operational data and real fault data, show that the proposed method provides competitive prediction accuracy and relatively stable cross-wheel performance under the current data protocol, with its most noticeable advantage observed on the more challenging Dataset III. Noise injection experiments further indicate that the Mamba temporal branch helps mitigate performance degradation under perturbed inputs. These results suggest that structured physical priors can be beneficial for stabilizing roughness regression in practical rail-vehicle monitoring scenarios, although further validation under broader operating conditions and stricter comparison protocols is still needed.
Executive Summary
PD-SOVNet introduces a novel physics-guided gray-box framework for the quantitative estimation of wheel polygonal roughness spectra (1st-40th order) from axle-box vibrations. Unlike prior work focused on detection or classification, this method offers continuous regression, explicitly addressing challenges with real operational data and unseen wheels. By integrating second-order vibration kernels, a MIMO coupling module, an adaptive physical correction branch, and a Mamba-based temporal branch, PD-SOVNet leverages modal-response priors while maintaining data-driven adaptability. Experimental results across three real-world datasets demonstrate competitive accuracy, stable cross-wheel performance, and resilience to noise, suggesting the utility of structured physical priors in practical rail-vehicle condition monitoring.
Key Points
- ▸ PD-SOVNet is a physics-guided gray-box model for continuous regression of wheel polygonal roughness spectra from axle-box vibrations.
- ▸ It estimates 1st-40th order roughness, moving beyond traditional detection or classification approaches.
- ▸ The architecture combines shared second-order vibration kernels, a 4x4 MIMO coupling module, an adaptive physical correction, and a Mamba-based temporal branch.
- ▸ The model embeds modal-response priors while retaining data-driven flexibility for sample-dependent corrections and residual temporal dynamics.
- ▸ Evaluated on three real-world datasets, PD-SOVNet shows competitive accuracy, stable cross-wheel performance, and improved noise robustness due to its Mamba temporal branch.
Merits
Novelty in Quantitative Regression
The article addresses a significant gap by focusing on continuous, multi-order roughness spectrum regression, which is more granular and informative than mere detection or classification, offering superior diagnostic capabilities.
Physics-Guided Gray-Box Architecture
The integration of physical priors (modal-response) with data-driven components (adaptive correction, Mamba temporal branch) creates a robust and interpretable model, balancing theoretical grounding with empirical flexibility.
Real-World Data Validation
Testing on three distinct real-world datasets, including operational and fault data, significantly enhances the credibility and practical applicability of the proposed method, particularly its performance on 'unseen-wheel conditions'.
Robustness to Noise
The demonstrated ability of the Mamba temporal branch to mitigate performance degradation under noise injection is a critical advantage for real-world deployment where sensor data is inherently noisy.
Demerits
Limited Scope of Validation
While three datasets are used, the abstract acknowledges the need for 'broader operating conditions and stricter comparison protocols,' suggesting potential limitations in generalizability beyond the tested scenarios.
Absence of Direct Comparative Benchmarks
The statement 'competitive prediction accuracy' lacks specific, named comparative benchmarks, making it difficult to precisely gauge its superiority against leading alternative methods in the field.
Interpretability of 'Adaptive Physical Correction'
The mechanism and specific contribution of the 'adaptive physical correction branch' are not fully elucidated in the abstract, potentially limiting understanding of its role and design choices.
Computational Overhead
The complexity of integrating multiple specialized modules (SOV kernels, MIMO, Mamba) might introduce significant computational overhead, which is not discussed but is crucial for real-time monitoring applications.
Expert Commentary
PD-SOVNet represents a significant methodological advancement in rail-vehicle condition monitoring, strategically pivoting from qualitative fault detection to quantitative spectral regression. The gray-box approach, marrying physics-driven priors with data-driven adaptability, is particularly commendable. This hybrid design addresses a fundamental tension in engineering AI: leveraging domain knowledge without sacrificing empirical flexibility. The inclusion of a Mamba-based temporal branch for noise resilience is a pragmatic and well-justified choice, reflecting an understanding of real-world operational challenges. However, the abstract's call for 'broader operating conditions and stricter comparison protocols' is a critical self-assessment. To truly establish its preeminence, the paper must articulate concrete benchmarks against state-of-the-art purely data-driven methods and provide a deeper dive into the computational efficiency and interpretability of its adaptive components. This work lays a strong foundation, but further rigorous validation is essential for widespread adoption and trust in a safety-critical domain.
Recommendations
- ✓ Conduct comprehensive comparative studies against leading purely data-driven deep learning models (e.g., advanced CNNs, LSTMs, Transformers) to quantitatively demonstrate PD-SOVNet's superior performance, especially in generalization and data efficiency.
- ✓ Provide a detailed analysis of the computational complexity and inference speed of PD-SOVNet, assessing its suitability for real-time, on-board monitoring systems and potential for edge deployment.
- ✓ Elaborate on the design principles and internal mechanisms of the 'adaptive physical correction branch,' offering insights into its learning process and how it dynamically refines physics-based priors.
- ✓ Expand validation to include a wider range of wheel types, vehicle speeds, track conditions, and environmental factors to rigorously test the model's generalizability and robustness across diverse operational scenarios.
Sources
Original: arXiv - cs.LG