acceptodds
Under review as a conference paper at ICLR 2027

Physics-Guided Port-Graph Learning for Fastener Parameter Identification from Vehicle Responses

Abstract

Identifying physical parameters from partial observations of coupled dynamical systems is a fundamental inverse problem, yet remains challenging when target effects are entangled with unknown excitation, unobserved states, and nuisance parameter variation. We study this setting in vehicle–track systems, where segment-shared fastener stiffness and damping are inferred from a single short record of vehicle-side responses without direct rail or slab measurements. We first assess local identifiability under track-irregularity and vehicle-parameter variation using profiled Fisher information. The analysis shows that the two target parameters retain distinct sensitivity directions and that their informative cues vary across frequency. Motivated by these observations, we propose a physics-guided port-graph learning framework that combines frequency-resolved magnitude and phase with the known mechanical interaction structure of the coupled system. A typed port graph propagates observed vehicle information through suspension, wheel–rail contact, and fastener interactions while representing unobserved track components as latent nodes. Experiments on simulated single-pass responses show improved parameter identification over strong spectral and graph-based baselines, with the clearest gains under short observation windows. Ablations further confirm the complementary roles of spectral resolution, phase information, and mechanically structured interaction. These results highlight physical interaction structure as an effective inductive bias for inverse identification from partial and temporally limited observations.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics-Guided Graph Learning, Vehicle–Track Coupled Dynamics, Partial-Observation Inspection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rQKSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper estimates rail-fastener stiffness and damping from a single short record of vehicle accelerations. Rail and slab states are unobserved, while vehicle parameters and track irregularity also affect the response. A nuisance-profiled local information analysis first examines whether the target parameters can be distinguished. The model then retains full-resolution spectral magnitude and circular phase, using a mechanical interaction graph with latent track nodes to predict the parameters. Vehicle–track simulations compare spectral and graph models, ablate components, and vary record duration, noise, and wheelset mass. The main gains occur with short records. Fine spectral resolution mainly helps stiffness estimation, while phase adds information for damping. Under strong noise, the stiffness advantage disappears while the damping advantage remains.

Critical strengths
  1. The short-record gains hold against strong, matched baselines. The comparison includes a parameter-matched spectral Transformer, spectral MLP/CNN models, a graph attention network, complex-periodogram ridge, and graph-bypass controls, with paired bootstrap intervals over records and seeds. At 1 s, the graph reduces in-distribution stiffness log-RMSE from 0.1136 to 0.0957 and damping log-RMSE from 0.1432 to 0.1294 relative to the Transformer, and below the full window it has 16 resolved gains and no resolved losses against it. This is why I consider the paper close to acceptance.

  2. The spectral controls explain which input carries which parameter. Holding the ridge readout fixed shows that full spectral resolution mainly helps stiffness and that phase mainly helps damping, consistent with the two parameters entering different quadratures of the fastener impedance. This ties the representation to the physics rather than to tuning.

Critical weaknesses
  1. The graph ablation does not isolate the physical topology. The ablation in Figure 3(b)–(c), Section 4.3, and the w/o-port-graph row of Table 1 remove interaction computation together with the mechanical connections. The gain could therefore come from added message-passing capacity and latent nodes rather than from the suspension, contact, and fastener wiring. The title and conclusion attribute the gain to physical interaction structure, so this distinction carries the main claim, and it is the main reason for my rating.

  2. The identifiability analysis is not connected to the estimator. In Section 3.2, Equation (8) states the information-retention condition after encoding, but it is checked only for the raw responses and not for the learned projection in Equation (12) or the graph representation. Table 2 supports the choice of spectral inputs; it does not establish that the learned compression preserves the target information identified by the analysis. Theorem 1 itself is a standard linear-algebra equivalence, so the analysis adds to the paper mainly through this link.

Questions for authors
  1. Is Equation (8) evaluated for the trained encoder, i.e., with R the Jacobian of Equation (12) or of the pooled graph representation? If the learned representation loses rank or becomes poorly conditioned along the stiffness–wheelset-mass direction, that would qualify the link between Section 3.2 and the estimator.

  2. The appendix reports local Cramér–Rao standard deviations of about 0.077 for ln k_f and 0.101 for ln c_f at 1 s. Under which matched target state, nuisance basis, noise level, and bias assumptions can these bounds be compared with Table 1? The reported test RMSE averages over a parameter distribution, and the learned estimator can be biased, so a direct numerical comparison alone would not establish proximity to the information limit.

  3. A graph of the same size and node types with shuffled or fully connected edges, trained under the same budget. If the mechanical wiring outperforms it, especially at short record lengths and under the wheelset-mass shift, that would support the topology claim.

3 comments
AuthorsOpenReview ↗

Response to Reviewer rQKS (Part 1/2)

Many thanks to Reviewer rQKS for the insightful review and for recognizing our contributions.

W1 & Q3: Isolating the Contribution of Mechanical Topology

Thank you for this constructive suggestion. To isolate the contribution of the physical connectivity, we added comparisons at 1.0 s between the mechanical port graph, an Erdős–Rényi G(N,M)G(N,M) random graph with the same number of edges, and a fully connected graph. All variants retain the same observed vehicle nodes, latent rail and slab nodes, node types, spectral encoder, hidden width, number of message-passing layers, readout, and training hyperparameters. The results are shown below.

EstimatorID θk\theta_{k}ID θc\theta_{c}Mass shift θk\theta_{k}Mass shift θc\theta_{c}
Mechanical port graph0.0957 ± 0.00300.1294 ± 0.00310.1138 ± 0.00310.1715 ± 0.0031
Erdős–Rényi G(N,M)G(N,M)0.0988 ± 0.00440.1298 ± 0.00270.1222 ± 0.00270.1871 ± 0.0038
Fully connected0.0970 ± 0.00300.1279 ± 0.00250.1164 ± 0.00430.1894 ± 0.0010

The mechanical port graph outperforms the edge-matched Erdős–Rényi random graph on all four metrics. It also outperforms the fully connected graph, which incurs greater message-passing cost per step, on in-distribution stiffness and both wheelset-mass shift metrics. These results support the role of mechanical topology as a physical inductive bias, with particularly strong performance under wheelset-mass shift.

W2 & Q1: Local Identifiability of Learned Representations

Thank you for raising this key question. We further use Eq. (8) to quantitatively diagnose information preservation and disentanglement at the nominal physical state through the input Jacobian R=∂z∂y\mathbf{R}=\frac{\partial\mathbf{z}}{\partial\mathbf{y}} of the trained readout representation z\mathbf{z}.

In evaluating Eq. (8), we account for vehicle-parameter and irregularity-amplitude directions and assess local identifiability after encoding using the rank and Gram condition number of the projected target matrix. We also construct the stiffness–wheelset-mass sensitivity pair Cz=[Rak,,Rbmw]\mathbf{C}_{z}=[\mathbf{R}\mathbf{a}_{k},\\,\mathbf{R}\mathbf{b}_{m_{w}}] and examine its rank, condition number, and directional cosine. The relative nuisance-sensitivity ratio ℓrel\ell_{\mathrm{rel}} in the table measures the nuisance-to-target sensitivity ratio after encoding, normalized by the corresponding ratio in the raw observations. The results are shown below.

EstimatorTarget rankProjected κ(M⊤M)\kappa(\mathbf{M}^{\top}\mathbf{M})ℓrel\ell_{\mathrm{rel}}Stiffness–mass κ(Cz⊤Cz)\kappa(\mathbf{C}_{z}^{\top}\mathbf{C}_{z})Stiffness–mass cosine ∣cos⁡∣\lvert\cos\rvert
Raw observation215.641.00015.050.861
Mechanical port graph210.90 ± 2.730.102 ± 0.0075.60 ± 1.610.672 ± 0.078
Erdős–Rényi G(N,M)G(N,M)215.81 ± 6.680.102 ± 0.0076.19 ± 1.390.710 ± 0.054
Fully connected212.69 ± 6.640.105 ± 0.0087.11 ± 0.500.751 ± 0.015

As shown above, across all training seeds, M\mathbf{M} retains rank 2 for the 1.0-s mechanical port graph, and Cz\mathbf{C}_{z} also has rank 2. The mechanical port graph also achieves the lowest mean values for the projected target condition number and absolute stiffness–wheelset-mass cosine, with relative nuisance sensitivity reduced by approximately 90%. These diagnostics directly connect Eq. (8) to the local identifiability of the trained representations.

AuthorsOpenReview ↗

Response to Reviewer rQKS (Part 2/2)

Q2: Comparing the Local CRB with Test Errors

Thank you for raising this insightful statistical question. First, the appendix's 1.0-s local CRB calculation uses the same observation channels, truncation before resampling, vehicle speed, and additive Gaussian noise level in the observation-frame domain as the main experiment in Table 1, with the noise standard deviation uniformly set to 10% of the training-set reference response RMS. At the nominal operating point, with both fastener and vehicle parameters set to their nominal values, nuisance projection removes nine vehicle-parameter directions and the track-irregularity directions. With these settings aligned, the remaining statistical distinction is that the local CRB gives a lower bound on the conditional standard deviation of a locally unbiased estimator at a single nominal state, while the test log-RMSE in Table 1 measures the estimator's overall root-mean-square error over the parameter and irregularity prior distributions. The latter reflects both sensitivity variation across operating points and the bias–variance trade-off arising from supervised learning with bounded priors.

To examine this physical information reference under a more comprehensive nuisance basis, we use the Landau–Pollak–Slepian (LPS) time–bandwidth scale to further expand the 1.0-s local CRB analysis to K∈2,4,8,16,32,41K\in\\{2,4,8,16,32,41\\} frequency cells. We consider both fixed-phase amplitude uncertainty (F, total nuisance dimension p=9+Kp=9+K) and joint amplitude–phase uncertainty (Q, total nuisance dimension p=9+2Kp=9+2K). For each KK, we partition the spatial-frequency range [0.005,0.5],cyc/m[0.005,0.5]\\,\mathrm{cyc/m} into KK equal cells and use their midpoints as the frequency nodes, so that each KK defines a fresh discretization of the full band. For the same KK, F and Q share the same nominal track-irregularity curve and nominal response. The profiled target information matrix Sθ=A⊤(I−PB)A\mathbf{S}_{\theta}=\mathbf{A}^{\top}(\mathbf{I}-\mathbf{P}_{B})\mathbf{A} and the corresponding CRB standard deviations are reported below.

Frequency cells KKNuisance dimension (F / Q)Target rank (F / Q)κ(Sθ)\kappa(\mathbf{S}_{\theta}) (F / Q)CRB SD θk\theta_{k} (F / Q)CRB SD θc\theta_{c} (F / Q)
211 / 132 / 215.64 / 13.200.0769 / 0.07760.1010 / 0.1020
413 / 172 / 214.64 / 12.780.0751 / 0.08130.1042 / 0.1084
817 / 252 / 28.76 / 9.610.0721 / 0.07640.1145 / 0.1306
1625 / 412 / 29.00 / 8.890.0689 / 0.07140.1057 / 0.1217
3241 / 732 / 210.50 / 10.070.0644 / 0.07920.0987 / 0.1252
4150 / 912 / 212.78 / 11.900.0638 / 0.08700.1022 / 0.1374

As shown above, the target information matrix Sθ\mathbf{S}_{\theta} retains full rank 2 in all 12 configurations. At K=41K=41, both the F and Q nuisance matrices retain full column rank. When all 82 irregularity amplitude–phase directions are included, the local CRB standard deviations for stiffness and damping rise to 0.0870 and 0.1374, respectively. This expanded analysis specifies the operating point, nuisance dimension, and unbiasedness assumptions associated with each CRB reference and confirms that both target parameters retain robust local identifiability at the nominal operating point even in the 91-dimensional nuisance subspace covering the full time–bandwidth scale.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer rQKS,

We appreciate your insightful feedback. We have carefully considered your comments and provided a point-by-point response, and we hope it addresses your concerns.

We are very keen to clarify any remaining points and look forward to your feedback.

Best regards,

Authors

Reviewer_u4pWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper combines spectral magnitude and phase with a mechanically structured graph network to estimate fastener stiffness and damping from simulated vehicle accelerations. A local sensitivity analysis examines identifiability under nuisance variation.

Critical strengths

Short-window gains over capacity-matched baselines are supported by paired uncertainty estimates. Spectral and graph-removal ablations help explain the empirical improvements.

Critical weaknesses

The contribution appears primarily application-specific. Theorem 1 is an elementary rank criterion, while Table 5 reports no statistically resolved advantage over untyped edges in any of four comparisons. The added value of mechanically typed interactions is therefore not established. The identifiability analysis uses only 2, 8, or 16 irregularity cells, although Appendix A estimates a nominal time–bandwidth scale of approximately 41 cells at one second. The highlighted information bounds use two cells. Since enlarging the nuisance subspace can reduce profiled information, the identifiability conclusions remain conditional on these restricted bases. Empirical validation also remains within one simulator configuration.

Questions for authors
  1. Clarification: How should the claimed benefit of typed interactions be interpreted given the unresolved comparisons in Table 5?
  2. Additional validation: Does identifiability persist with a richer irregularity basis matching the stated time–bandwidth scale?
3 comments
AuthorsOpenReview ↗

Response to Reviewer u4pW (Part 1/2)

Many thanks to Reviewer u4pW for the insightful review and for recognizing our contributions.

W1 & Q1: Physics-Guided Design and Mechanical Topology

Thank you for the insightful comment. We wish to explain that our model design is guided by the physical mechanisms underlying parameter identification. The real and imaginary parts of the fastener impedance Zf(ω)=kf+iωcfZ_{f}(\omega)=k_{f}+i\omega c_{f} are governed by stiffness and damping, respectively. This property directly motivated our spectral features, which retain both full-resolution magnitude and circular phase. As shown by the fixed-readout ablation study in Section 4.4, full-resolution spectral magnitude and circular phase primarily drive the improvements in stiffness and damping identification, respectively.

We agree that Theorem 1 adopts a standard local rank criterion. To connect this analysis to the proposed method, we further apply Eq. (8) to evaluate the trained readout representation z\mathbf{z} through its input Jacobian R=∂z∂y\mathbf{R}=\frac{\partial\mathbf{z}}{\partial\mathbf{y}}. We quantitatively diagnose information preservation and disentanglement using the nuisance-projected target matrix M=(I−PRB)RA\mathbf{M}=(\mathbf{I}-\mathbf{P}_{RB})\mathbf{R}\mathbf{A}, the relative nuisance-sensitivity ratio ℓrel\ell_{\mathrm{rel}}, and the stiffness–wheelset-mass sensitivity pair Cz=[Rak,,Rbmw]\mathbf{C}_{z}=[\mathbf{R}\mathbf{a}_{k},\\,\mathbf{R}\mathbf{b}_{m_{w}}].

Here, we define the relative nuisance-sensitivity ratio as ℓrel=∥RB∥F/∥RA∥F∥B∥F/∥A∥F\ell_{\mathrm{rel}}=\frac{\lVert \mathbf{R}\mathbf{B}\rVert_{F}/\lVert \mathbf{R}\mathbf{A}\rVert_{F}}{\lVert \mathbf{B}\rVert_{F}/\lVert \mathbf{A}\rVert_{F}}, where A\mathbf{A} and B\mathbf{B} are the response sensitivity matrices for the target and nuisance parameters, respectively. This metric measures the relative change in the nuisance-to-target sensitivity ratio after encoding, using the corresponding ratio in the raw observations as the reference. A value below 1 indicates reduced sensitivity to nuisances relative to the targets. The results are shown below.

EstimatorTarget rank (M\mathbf{M})Projected κ(M⊤M)\kappa(\mathbf{M}^{\top}\mathbf{M})ℓrel\ell_{\mathrm{rel}}Stiffness–mass κ(Cz⊤Cz)\kappa(\mathbf{C}_{z}^{\top}\mathbf{C}_{z})Stiffness–mass cosine ∣cos⁡∣\lvert\cos\rvert
Raw observation215.641.00015.050.861
Mechanical port graph210.90 ± 2.730.102 ± 0.0075.60 ± 1.610.672 ± 0.078
Erdős–Rényi G(N,M)G(N,M)215.81 ± 6.680.102 ± 0.0076.19 ± 1.390.710 ± 0.054
Fully connected212.69 ± 6.640.105 ± 0.0087.11 ± 0.500.751 ± 0.015

As shown above, across all training seeds, M\mathbf{M} retains rank 2 for the mechanical port graph, and Cz\mathbf{C}_{z} also has rank 2. The mechanical port graph reduces relative nuisance sensitivity by approximately 90% and achieves the lowest mean values for the projected target condition number and absolute stiffness–wheelset-mass cosine. These diagnostics directly connect Eq. (8) to the local identifiability of the trained representations.

We agree that Table 5 has not established an independent, statistically significant benefit from explicit edge typing. However, we wish to clarify that the untyped-edge control in Table 5 retains the mechanical connectivity and latent track nodes, removing only the learnable interaction-type embeddings. To further isolate the contribution of the physical connectivity itself, we added comparisons at 1.0 s between the mechanical port graph, an Erdős–Rényi G(N,M)G(N,M) random graph with the same number of edges, and a fully connected graph, keeping the node count and types, encoder, hidden width, readout, and training hyperparameters identical. The results are shown below.

EstimatorID θk\theta_{k}ID θc\theta_{c}Mass shift θk\theta_{k}Mass shift θc\theta_{c}
Mechanical port graph0.0957 ± 0.00300.1294 ± 0.00310.1138 ± 0.00310.1715 ± 0.0031
Erdős–Rényi G(N,M)G(N,M)0.0988 ± 0.00440.1298 ± 0.00270.1222 ± 0.00270.1871 ± 0.0038
Fully connected0.0970 ± 0.00300.1279 ± 0.00250.1164 ± 0.00430.1894 ± 0.0010

The mechanical port graph outperforms the edge-matched Erdős–Rényi random graph on all four metrics. It also outperforms the fully connected graph, which incurs greater message-passing cost per step, on in-distribution stiffness and both wheelset-mass shift metrics. These results support the role of mechanical topology as a physical inductive bias, with particularly strong performance under wheelset-mass shift.

AuthorsOpenReview ↗

Response to Reviewer u4pW (Part 2/2)

W2 & Q2: Local Identifiability under Richer Irregularity Bases

Thank you for raising this key question. We expanded the 1.0-s local identifiability analysis to K∈2,4,8,16,32,41K\in\\{2,4,8,16,32,41\\} frequency cells, covering the nominal time–bandwidth scale of approximately 41 cells, under both fixed-phase amplitude uncertainty (F) and joint amplitude–phase uncertainty (Q). The two parameterizations are rF(x;η)=∑j=1KAjeηjsin⁡(2πfjx+ϕj)r_{F}(x;\boldsymbol{\eta})=\sum_{j=1}^{K} A_{j} e^{\eta_{j}}\sin(2\pi f_{j} x+\phi_{j}) and rQ(x;η,ξ)=∑j=1KAjeηjsin⁡(2πfjx+ϕj+ξj)r_{Q}(x;\boldsymbol{\eta},\boldsymbol{\xi})=\sum_{j=1}^{K} A_{j} e^{\eta_{j}}\sin(2\pi f_{j} x+\phi_{j}+\xi_{j}), respectively. Including the vehicle-parameter directions gives total nuisance dimensions of p=9+Kp=9+K and p=9+2Kp=9+2K, respectively.

We wish to add that the frequency nodes follow a cell-center rule based on a uniform partition of the full frequency band. For each KK, we repartition the fixed spatial-frequency range [0.005,0.5],cyc/m[0.005,0.5]\\,\mathrm{cyc/m} into KK equal cells and use each cell midpoint as a frequency node. Thus, increasing KK recomputes all frequency nodes across the full band and adjusts the positions of the existing nodes, with each KK defining a complete discretization of the band.

For each KK, F and Q share the same nominal track-irregularity curve and nominal response, with the fastener and vehicle parameters set to their nominal values. Using the same noise conditions as the main experiments, we compute the profiled target information matrix Sθ=A⊤(I−PB)A\mathbf{S}_{\theta}=\mathbf{A}^{\top}(\mathbf{I}-\mathbf{P}_{B})\mathbf{A} and the corresponding CRB standard deviations. The results are shown below.

Frequency cells KKNuisance dimension (F / Q)Target rank (F / Q)κ(Sθ)\kappa(\mathbf{S}_{\theta}) (F / Q)CRB SD θk\theta_{k} (F / Q)CRB SD θc\theta_{c} (F / Q)
211 / 132 / 215.64 / 13.200.0769 / 0.07760.1010 / 0.1020
413 / 172 / 214.64 / 12.780.0751 / 0.08130.1042 / 0.1084
817 / 252 / 28.76 / 9.610.0721 / 0.07640.1145 / 0.1306
1625 / 412 / 29.00 / 8.890.0689 / 0.07140.1057 / 0.1217
3241 / 732 / 210.50 / 10.070.0644 / 0.07920.0987 / 0.1252
4150 / 912 / 212.78 / 11.900.0638 / 0.08700.1022 / 0.1374

As shown above, the target information matrix Sθ\mathbf{S}_{\theta} retains rank 2 in all 12 configurations. At K=41K=41, the F and Q nuisance matrices have full column rank, and the minimum eigenvalues of the target information matrices are 5.620 and 3.102, respectively. These results show that both target parameters remain locally identifiable under richer nuisance bases matching the stated time–bandwidth scale.

W3: Simulation Coverage and Field Validation

Thank you for highlighting this important issue. We wish to add that our experiments systematically vary the fastener target multipliers ([0.7,1.3][0.7,1.3]) and vehicle-parameter multipliers ([0.85,1.15][0.85,1.15]) within the Zhai model family, generate an independent track-irregularity realization for each sample, and introduce a wheelset-mass distribution shift to assess robustness within this model family. Simulation-based fastener identification has established precedents (Chen et al., 2022, Yuan et al., 2021), and the underlying Zhai dynamics have undergone full-scale field validation, as reported in Section 6 of Zhai et al. (2009). Evaluating our joint continuous two-parameter regression task on field data requires onboard acceleration measurements paired with independently calibrated segment-level fastener stiffness and damping. However, existing public onboard datasets such as DR-Train and OnboardEU provide only maintenance records or geometric track-irregularity annotations and lack continuous fastener stiffness and damping reference values. Building on the current simulation benchmark, our future work will improve model generalization to real-world operating conditions and validate it on onboard field measurements in conjunction with field calibration of fastener parameters.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer u4pW,

We appreciate your insightful feedback. We have carefully considered your comments and provided a point-by-point response, and we hope it addresses your concerns.

We are very keen to clarify any remaining points and look forward to your feedback.

Best regards,

Authors

Reviewer_dDNuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses a system identification problem involving railway fastener stiffness and damping from vehicle acceleration measurements (without direct observations of the track).

It investigates local identifiability using nuisance-profiled Fisher information and proposes a GNN combining frequency-resolved magnitude and phase information with a graph representing the mechanical connections between vehicle and track components. Unobserved components are represented by learned embeddings.

Experiments on simulated vehicle--track responses show improvements over spectral and graph-based baselines, more so when observation windows are short.

Critical strengths

The problem is interesting and practically relevant -- it's a nice application paper. It's clearly presented and the combination of spectral information with a known mechanical structure is reasonably motivated. The computational evaluation includes several baselines, ablation studies, and statistical comparisons (across multiple training seeds).

Critical weaknesses

My main concern is the limited methodological novelty. The proposed approach essentially combines established spectral representations with graph message passing, while the identifiability theorem is very standard.

Notably, Appendix B, Table 5, shows no statistically significant improvement over the untyped-edge variant in any of the four main comparisons -- this makes me wonder whether it is really the mechanical interaction structure that helps, or simply the introduction of a GNN.

Finally, the experiments are extensive (they really are) but they all rely on data generated by a single simulator.

Questions for authors

Could the authors extend their experimental evaluation to some real-world vehicle--track measurements?

Could they compare their work against established inverse-identification methods, such as nonlinear least squares or Bayesian model updating, EKF/UKF-based parameter estimation, Gaussian process regression, and neural posterior estimation? The current comparison is against mostly generic NN architectures, ignoring simpler (and more established) system id techniques.

4 comments
AuthorsOpenReview ↗

Response to Reviewer dDNu (Part 1/2)

Many thanks to Reviewer dDNu for the insightful review and for recognizing our contributions.

W1: Physics-Guided Design and Identifiability

Thank you for the insightful comment. We wish to explain that our model design is guided by the physical mechanisms underlying parameter identification. The real and imaginary parts of the fastener impedance Zf(ω)=kf+iωcfZ_{f}(\omega)=k_{f}+i\omega c_{f} are governed by stiffness and damping, respectively. This property directly motivated our spectral features, which retain both full-resolution magnitude and circular phase. As shown by the fixed-readout ablation study in Section 4.4, full-resolution spectral magnitude and circular phase primarily drive the improvements in stiffness and damping identification, respectively.

We agree that Theorem 1 adopts a standard local rank criterion. To connect this analysis to the proposed method, we further apply Eq. (8) to evaluate the trained readout representation z\mathbf{z} through its input Jacobian R=∂z∂y\mathbf{R}=\frac{\partial\mathbf{z}}{\partial\mathbf{y}}. We quantitatively diagnose information preservation and disentanglement using the nuisance-projected target matrix M=(I−PRB)RA\mathbf{M}=(\mathbf{I}-\mathbf{P}_{RB})\mathbf{R}\mathbf{A}, the relative nuisance-sensitivity ratio ℓrel\ell_{\mathrm{rel}}, and the stiffness–wheelset-mass sensitivity pair Cz=[Rak,,Rbmw]\mathbf{C}_{z}=[\mathbf{R}\mathbf{a}_{k},\\,\mathbf{R}\mathbf{b}_{m_{w}}].

We define the relative nuisance-sensitivity ratio as ℓrel=∥RB∥F/∥RA∥F∥B∥F/∥A∥F\ell_{\mathrm{rel}}=\frac{\lVert \mathbf{R}\mathbf{B}\rVert_{F}/\lVert \mathbf{R}\mathbf{A}\rVert_{F}}{\lVert \mathbf{B}\rVert_{F}/\lVert \mathbf{A}\rVert_{F}}, where A\mathbf{A} and B\mathbf{B} are the response sensitivity matrices for the target and nuisance parameters, respectively.

EstimatorTarget rankProjected κ(M⊤M)\kappa(\mathbf{M}^{\top}\mathbf{M})ℓrel\ell_{\mathrm{rel}}Stiffness–mass κ(Cz⊤Cz)\kappa(\mathbf{C}_{z}^{\top}\mathbf{C}_{z})Stiffness–mass cosine ∣cos⁡∣\lvert\cos\rvert
Raw observation215.641.00015.050.861
Mechanical port graph210.90 ± 2.730.102 ± 0.0075.60 ± 1.610.672 ± 0.078
Erdős–Rényi G(N,M)G(N,M)215.81 ± 6.680.102 ± 0.0076.19 ± 1.390.710 ± 0.054
Fully connected212.69 ± 6.640.105 ± 0.0087.11 ± 0.500.751 ± 0.015

As shown above, the trained graph representations retain full rank 2 across all random seeds and suppress relative nuisance sensitivity by approximately 90%. The mechanical port graph has the lowest mean projected condition number and stiffness–mass cosine. These diagnostics directly connect Eq. (8) to the local identifiability of the trained representations.

W2: Distinguishing Mechanical Topology from Edge Typing

Thank you for raising this key question. We agree that Table 5 has not established an independent, statistically significant benefit from explicit edge typing. However, we wish to clarify that the untyped-edge control in Table 5 retains the mechanical connectivity and latent track nodes, removing only the learnable interaction-type embeddings. To further isolate the contribution of the physical connectivity itself, we added comparisons at 1.0 s between the mechanical port graph, an Erdős–Rényi G(N,M)G(N,M) random graph with the same number of edges, and a fully connected graph, keeping the node count and types, encoder, hidden width, readout, and training hyperparameters identical. The results are shown below.

EstimatorID θk\theta_{k}ID θc\theta_{c}Mass shift θk\theta_{k}Mass shift θc\theta_{c}
Mechanical port graph0.0957 ± 0.00300.1294 ± 0.00310.1138 ± 0.00310.1715 ± 0.0031
Erdős–Rényi G(N,M)G(N,M)0.0988 ± 0.00440.1298 ± 0.00270.1222 ± 0.00270.1871 ± 0.0038
Fully connected0.0970 ± 0.00300.1279 ± 0.00250.1164 ± 0.00430.1894 ± 0.0010

The mechanical port graph outperforms the edge-matched Erdős–Rényi random graph on all four metrics. It also outperforms the fully connected graph, which incurs greater message-passing cost per step, on in-distribution stiffness and both wheelset-mass shift metrics. These results support the role of mechanical topology as a physical inductive bias, with particularly strong performance under wheelset-mass shift.

AuthorsOpenReview ↗

Response to Reviewer dDNu (Part 2/2)

W3 & Q1: Simulation Coverage and Field Validation

Thank you for highlighting this important issue. We wish to add that our experiments systematically vary the fastener target multipliers ([0.7,1.3][0.7,1.3]) and vehicle-parameter multipliers ([0.85,1.15][0.85,1.15]) within the Zhai model family, generate an independent track-irregularity realization for each sample, and introduce a wheelset-mass distribution shift to assess robustness within this model family. Simulation-based fastener identification has established precedents (Chen et al., 2022, Yuan et al., 2021), and the underlying Zhai dynamics have undergone full-scale field validation, as reported in Section 6 of Zhai et al. (2009).

Evaluating our joint continuous regression task on field data requires onboard acceleration measurements paired with independently calibrated segment-level fastener stiffness and damping. However, existing public onboard datasets such as DR-Train and OnboardEU provide only maintenance records or geometric track-irregularity annotations and lack continuous fastener stiffness and damping reference values. Building on the current simulation benchmark, our future work will improve model generalization to real-world operating conditions and validate it on onboard field measurements in conjunction with field calibration of fastener parameters.

Q2: Comparison with Established Inverse Methods

Thank you for suggesting comparisons with established inverse-identification methods. Following your suggestion, in addition to the complex-periodogram ridge regression already included in Table 1, we added two established inverse-identification baselines, Gaussian Process Regression (GPR) and Neural Posterior Estimation (NPE), using the same experimental settings and data splits as Table 1.

GPR uses two independent exact Gaussian processes with Matérn-5/2 ARD kernels on PCA32 features to estimate stiffness and damping, respectively. NPE uses masked autoregressive flows (MAFs) from the official sbi library with a learnable spectral embedding. For each observation window, the best architecture is selected by validation negative log likelihood, and point predictions are obtained from the posterior mean in log-multiplier space. The results are shown below.

WindowEstimatorID θk\theta_{k}ID θc\theta_{c}Mass shift θk\theta_{k}Mass shift θc\theta_{c}
0.5 sPort graph0.1103 ± 0.00120.1330 ± 0.00220.1238 ± 0.00190.1740 ± 0.0049
0.5 sGPR0.18450.17940.17020.1865
0.5 sNPE0.1287 ± 0.00680.1476 ± 0.00360.1346 ± 0.00570.1681 ± 0.0044
1.0 sPort graph0.0957 ± 0.00300.1294 ± 0.00310.1138 ± 0.00310.1715 ± 0.0031
1.0 sGPR0.17330.16870.15220.1805
1.0 sNPE0.1281 ± 0.02910.1401 ± 0.00310.1373 ± 0.02260.1709 ± 0.0033

At both observation windows, the mechanical port graph outperforms both added baselines on all in-distribution metrics, and NPE is competitive under wheelset-mass shift. Forward-solver-based methods such as nonlinear least squares, Bayesian model updating, and EKF/UKF require explicit joint inference priors and state-space parameterizations for unknown spatial track irregularity, vehicle nuisance parameters, and latent track states. Their accuracy and computational cost depend strongly on these additional assumptions. Our added comparisons therefore focus on kernel regression and amortized posterior estimation, both of which fall within the direct inverse-mapping approach.

Thanks for the answers and the new experiments, which address part of my concerns. The GPR and NPE baselines are welcome, but under the mass shift NPE does as well as the port graph or better on damping, which is where the topology comparison finds its clearest benefit. Since the vehicle and irregularity priors are sampled, a model-based baseline on the same simulator looks feasible, and it would be informative.

In distribution, the mechanical graph does no better than a random or a fully connected one. The gains appear only under the mass shift, at the longer window, so the comparison at short windows would be useful, and the claims would then follow the evidence.

I understand that no field data with calibrated fastener parameters is available; please state this as a limitation. My main concern, on the methodological novelty, stands, and I keep my score.

Many thanks to Reviewer dDNu for reviewing our response and acknowledging that the additional experiments address some of your concerns.

Baseline Comparisons and the Benefits of Mechanical Topology

Thank you for welcoming the additional baselines and for your further analysis of the topology controls. We agree that NPE is competitive in damping estimation under wheelset-mass shift. In the reported 1.0 s topology comparison, the mechanical port graph has a slightly lower mean in-distribution stiffness error, while the three topologies have similar in-distribution performance overall. Under wheelset-mass shift, the mechanical port graph has lower mean stiffness and damping errors than both the edge-matched Erdős–Rényi random graph and the fully connected graph. We will further clarify the experimental conditions under which these topology benefits are observed.

Computational Cost of Model-Based Identification Methods

Thank you for suggesting further comparisons with model-based identification methods using the same simulator. We agree that the available priors for vehicle parameters and track irregularity provide a basis for constructing these methods. However, the unknown vehicle parameters and irregularity still need to be inferred for each observation record, requiring repeated calls to the forward dynamics solver.

We use the same physical model, observations, and noise settings as the main experiments, representing the unknown irregularity with 41 fixed-phase frequency cells covering the nominal time–bandwidth scale at 1.0 s. Including the two fastener target parameters and nine vehicle parameters gives a joint unknown-parameter dimension of D=2+9+41=52D=2+9+41=52. The evaluation covers the test set, and a complete forward solve for a 1.0 s observation window takes approximately 10.4 seconds on one CPU core.

NLS: Consider a joint nonlinear least-squares implementation with a forward finite-difference Jacobian. Each full Jacobian evaluation requires approximately D+1=53D+1=53 forward solves. With a budget of 50 full Jacobian updates per record, this requires 50×53=265050\times53=2650 forward solves.

BMU: Consider joint posterior sampling using classical random-walk Metropolis–Hastings, with the likelihood at the current position cached. Each new proposal requiring likelihood evaluation entails one forward solve. With a budget of 4 chains, each with 1000 burn-in steps and 2000 sampling steps, and including the initial evaluation for each chain, this requires approximately 4×(1+1000+2000)=120044\times(1+1000+2000)=12004 forward solves per record.

EKF: Consider a classical augmented-state EKF that jointly estimates displacements, velocities, and static unknown parameters. The physical model contains 830 generalized coordinates, giving an augmented-state dimension of n=2×830+52=1712n=2\times830+52=1712. Constructing the state-transition Jacobian by forward finite differences requires approximately n+1=1713n+1=1713 interval propagations per update interval. Since the integration steps across all update intervals sum to those required for the complete 1.0 s record, the total propagation workload per record is approximately equivalent to 1713 full-trajectory propagations.

UKF: Consider a standard scaled UKF with the same augmented state, propagating 2n+1=34252n+1=3425 sigma points per update interval. As with the EKF, summing the propagation steps across all update intervals gives a workload approximately equivalent to 3425 full-trajectory propagations per record.

Estimated test-set runtimes are shown below.

MethodSolves / equivalent propagations per recordTest-set totalSerial runtimeIdeal 24-core parallel runtime
NLS26501,017,600122.49 days5.10 days
BMU (Metropolis–Hastings)120044,609,536554.85 days23.12 days
EKF1713657,79279.18 days3.30 days
UKF34251,315,200158.31 days6.60 days

These estimates account only for forward dynamics solves in a single run, excluding covariance-matrix operations in the filters, hyperparameter tuning, and repeated runs. The runtimes in the table are therefore conservative estimates, and the computational cost of a complete evaluation would be higher. We recognize the value of these model-based comparisons, but the computational and time costs above prevented us from completing the corresponding implementations and evaluations.

Field-Validation Limitation

Thank you for understanding the constraints on field data. We accept this suggestion and will explicitly state in the revised manuscript that the current experimental validation remains simulation-based and that our method has not yet been validated on onboard field measurements paired with independently calibrated segment-level fastener stiffness and damping. This is a limitation of the present study.

Thank you again for taking the time to review our work and engage in this discussion.

Comments on the paper

2 comments

Dear Reviewers,

Thanks for your time and effort in reviewing this paper. The authors have responded to the comments by providing a rebuttal. Please read the rebuttal and indicate whether your concerns have been resolved. Thanks.

Regards from the AC.

Many thanks to the Area Chair and the three reviewers for their time and suggestions throughout the review and discussion process. To support an overall assessment of our paper, we summarize the additional evidence provided during discussion and explain how it further supports our contributions.

Research Problem and Contributions

We study fastener stiffness and damping identification from short vehicle-side vibration records, with track responses unobserved and target effects entangled with vehicle-parameter variation and unknown irregularity. A nuisance-aware local identifiability analysis examines the information available for distinguishing the two targets. The roles of stiffness and damping in the fastener impedance motivate a spectral representation retaining full-resolution magnitude and circular phase, while a mechanical port graph models interactions between observed vehicle and latent track components. Matched-baseline comparisons and fixed-readout ablations in the paper support improved short-record identification and the different roles of spectral resolution and phase for the two targets.

Additional Analyses and Experimental Results

In response to the reviews, we added the following analyses and experiments.

  • Connecting identifiability analysis directly to trained representations. Using the input Jacobian of the trained readout representation, we extended identifiability checks from raw responses to learned representations. At the nominal operating point, the mechanical port graph's nuisance-projected target matrix and stiffness–wheelset-mass sensitivity pair remain full rank across all evaluated training seeds. These diagnostics establish a direct empirical link between the physical information analysis and the estimator.

  • Local identifiability under richer irregularity bases. We expanded the 1.0 s irregularity basis to 41 frequency cells matching the nominal time–bandwidth scale and further included phase uncertainty. Both targets remain locally identifiable at the nominal operating point across all evaluated configurations, showing that the original finding holds under richer nuisance representations.

  • Isolating mechanical connectivity. We added edge-matched Erdős–Rényi and fully connected controls, keeping the nodes, node types, encoder, hidden width, readout, and training settings identical. In the reported 1.0 s comparison, the three topologies perform similarly in distribution, while the mechanical port graph has lower mean stiffness and damping errors under wheelset-mass shift. This further clarifies the conditions under which the topology benefits are observed.

  • Broadening comparisons with established inverse-identification methods. Beyond the complex-periodogram ridge baseline in the paper, we added Gaussian process regression (GPR) and neural posterior estimation (NPE) with the same observations and data splits. The mechanical port graph has lower in-distribution errors than both added baselines at both evaluated windows, while NPE is competitive in damping estimation under wheelset-mass shift.

Scope of the Conclusions and Overall Value

Current validation remains simulation-based, and field validation using onboard measurements paired with independently calibrated fastener parameters is still lacking. We will explicitly state this limitation in the revised manuscript.

We believe that connecting physical information conditions in partially observed inverse identification to representation design and mechanical interaction modeling, and testing these links through matched baselines, controlled ablations, and trained-representation diagnostics, constitutes a solid contribution. The additional evidence strengthens this empirical basis and clarifies the scope of our conclusions. We hope the Area Chair will consider this evidence in assessing the paper's overall research value.

Thank you again for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.