SPIRAL-PO: Symbolic Identification of Partially Observed Nonlinear Dynamics with Application to Rotating Machinery

arXiv:2608.00466 · eess.SY, cs.SY · Submitted 2026-08-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "SPIRAL-PO: Symbolic Identification of Partially Observed Nonlinear Dynamics with Application to Rotating Machinery".

Dev: Here's the gist of the paper. SPIRAL-PO presents a framework that treats hidden-state effects not as nuisances to be eliminated but as structured, physically interpretable signatures in the observable equations of motion <ref:2608.00466#pg2>.

Rosa: The method is designed to identify governing equations from partial observations by iteratively fitting a multi-output residual network and projecting the learned structure onto a physics-constrained candidate library <ref:2608.00466#pg6>.

Dev: This approach is significant because it recovers interpretable, statistically validated models from noisy data where direct state measurement is impractical or impossible <ref:2608.00466#pg4>.

Rosa: Pruning redundant reasoning tokens A recurring challenge in engineering dynamics is that the models we need are governed by states we cannot measure <ref:2608.00466#pg2>.

Paper summary: Rosa: The paper explains that existing data-driven methods like SINDy assume you can measure every state that goes into your library Dev But in rotordynamics, you can't measure the tilt angles alpha and beta because they are hidden, even though those angles are physically coupled to the lateral displacements x and y through the bearing geometry Rosa SPIRAL-PO tackles this by keeping the observable coordinates fixed and projecting those hidden state dynamics onto them as physically labeled terms Dev The thesis is that every discovered term retains a direct mechanical interpretation because it maps to a real mechanism Rosa It’s about treating those hidden state effects not as nuisances to eliminate, but as structured signatures in the equations of motion Dev

Dev: And they do this by using an iterative closed-loop algorithm where they decompose the observable acceleration into two parts: one part that depends only on what we see, and another part that contains all the coupling from those hidden states Rosa The loop then fits a multi-output residual network to this projection residual, and it projects that learned structure onto a physics-constrained library Dev This whole process is designed to recover the governing equations even when you only have partial measurements like x(t) and y(t) Rosa It’s about making sense of the dynamics from just what we observe.

Taro: When you think about this, Rosa, it sounds like a way to build a model that respects physics without needing perfect sensor data for every single variable Dev But I wonder how robust this iterative fitting is when the world misbehaves, you know, when the noise artifacts are really high Rosa Because they rely on projecting onto a library of candidate terms that need to survive some kind of statistical testing before they get added to the model Dev

Rosa: That’s a good point about robustness. They build in this multi-stage gating protocol where candidate terms have to pass two independent statistical gates before they are trusted and added Taro Gate one is basically a per-term significance screen using the Bonferroni t-test, and Gate two is a partial F-test with a validation rollback to make sure the model doesn't hurt predictions on data it hasn't seen Dev It’s about ensuring model trust through this process Rosa

Taro: So, if we look at this from an autonomy research standpoint, how does that gating protocol handle a situation where the underlying physics itself is slightly different than what the library assumes Dev If the system suddenly changes its coupling structure in a way that violates those statistical tests, what happens then? Rosa The framework also uses a frequency domain whiteness guard instead of just using variance thresholds for filtering residuals Dev That’s an extra layer to check if the noise is structured correctly before trusting the fitting process Rosa

Taro: I see. So it’s not just about fitting the best curve, it's about ensuring that curve makes sense in terms of physical structure and statistical significance at every step Dev That seems like a solid approach for handling uncertainty when you can't measure everything directly Rosa So we're looking at how this works for rotating machinery next.

Conclusion: Rosa: So, looking at the whole SPIRAL-PO paper by Ayoubi and Piramoon, the main point is that you can recover a lot of physics from just measuring lateral displacements x and y Dev They show that gyroscopic coupling, which is hard to see, and the Duffing nonlinearity in a bearing, plus that tilt-mediated linear cross-coupling are all recoverable from just those two measurements Rosa The authors confirm this numerically against a synthetic four degree of freedom system Dev This means for someone who only listens to the show, it tells you that even if your sensors are missing crucial information, you can still reconstruct the full motion and understand the physical forces at play Rosa

Dev: What’s interesting is how they then use an extended Kalman filter built on their validated dynamics to actually reconstruct all four states, including those hidden tilt angles alpha and beta Rosa This isn't just a guess; it's reconstructing the hidden trajectory based on the correct equation structure they discovered Dev The implication here for control systems is that you can design controllers that operate on the full state even when your sensors only give you a subset of information Rosa It moves beyond just finding an equation to actually using that recovered model to understand and control a complex physical system.

Taro: I think the core contribution, besides just getting the numbers right, is that they proved impossibility for constant speed operation Dev They show that if you keep the speed constant, the disk traces a near circular orbit where all those candidate signatures collapse onto a low order trigonometric span Rosa This causes the signature matrix to lose column rank and makes gyroscopic coupling and Duffing nonlinearity unidentifiable at any sample size or sensor precision Dev

Taro: That impossibility result is pretty strong, Rosa. It means that relying on constant-speed data alone won't let you recover those specific nonlinear effects, no matter how good your sensors are or how long you run the test Dev It sets a clear boundary for what kind of data you need to see to get the full story Rosa

Rosa: Exactly. So while SPIRAL-PO is fantastic at recovering the structure from noisy coast-down data where things are complex, they’ve also given us a hard limit on when that specific type of data is sufficient Dev It shows that we can get very deep into nonlinear dynamics, but we have to be careful about what kind of excitation you provide to get the unique coefficients back Rosa

Dev: So the big picture here is moving away from needing direct measurement of every state and instead focusing on discovering the physical structure itself from what's observable, and then using that structure to reconstruct everything else Rosa That’s how they tackle these hard problems in engineering dynamics.

Department of Mechanical Engineering, Santa Clara University · Laboratory Products Division, Thermo Fisher Scientific

eess.SY, cs.SY

Submitted: 2026-08-01

Updated: 2026-11-04

Comments: Revised version, under review at Nonlinear Dynamics. 32 pages, 6 figures, 14 tables. Code and data: https://doi.org/10.5281/zenodo.23705395

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 81/100

The gist: SPIRAL-PO presents a framework that treats hidden-state effects not as nuisances to be eliminated but as structured…

Key concepts

Hidden-State Effects
These are unmeasured variables that influence the system's motion but cannot be directly observed. SPIRAL-PO treats them not as errors to remove, but as structured physical signatures within the equations of motion. This allows for physically interpretable modeling even without direct access to all states.
Physics-Constrained Library
This is a pre-defined set of mathematical terms derived from dimensional analysis and known physical principles. Instead of fitting arbitrary functions, the method projects learned dynamics onto this library, ensuring that every discovered term has a clear mechanical interpretation in the context of the physical system.
State Reconstruction via EKF
An Extended Kalman Filter is built on the validated dynamics discovered by SPIRAL-PO. This filter uses only the observable measurements (like lateral displacements) to reconstruct all hidden states, such as tilt angles. It proves that discovering the correct equation structure is key to accurately inferring unmeasured variables.

Terminology

Summary

SPIRAL-PO presents a framework that treats hidden-state effects not as nuisances to be eliminated but as structured, physically interpretable signatures in the observable equations of motion <ref:2608. p. 000466#pg2>. The method is designed to identify governing equations from partial observations by iteratively fitting a multi-output residual network and projecting the learned structure onto a physics-constrained candidate library <ref:2608. p. 000466#pg6>. This approach is significant because it recovers interpretable, statistically validated models from noisy data where direct state measurement is impractical or impossible <ref:2608. p. 100466#pg4>.

Pruning redundant reasoning tokens

A recurring challenge in engineering dynamics is that the models we need are governed by states we cannot measure <ref:2608.00466#pg2>. Existing data-driven methods handle the full-state case well, but they share the assumption that every state entering the library is directly measurable <ref:2608. p. 100466#pg3>. SPIRAL-PO differs by keeping observable coordinates fixed and projecting hidden-state dynamics onto them as physically labeled library terms, so that every discovered term retains a direct mechanical interpretation <ref:2608. p. 100466#pg4>.

How the training signal is built

The identification objective is to determine which terms belong in Delta and what their coefficients are, using only the observable trajectory <ref:2608.00466#pg6>. This involves an iterative closed-loop algorithm where at each iteration, the observable acceleration is decomposed as q¨˜obs(t) = f(n)physξ(t); θ(n) + ∆(n)ξ(t) + ε(t), where r(n)(t) = q¨˜obs(t) − f(n)physξ Tθ(n), and the loop advances as f(n+1)phys project --> r(n+1)(NN fit --> f(n+1)poly Gate 1 --> Gate 2 --> f(n+1)phys <ref:2608. p. 400466#pg6>.

Where the gains come from

The gains come from a multi-stage gating protocol that ensures model trust by requiring candidate terms to survive two independent statistical gates and do not harm held-out prediction accuracy <ref:2608.00466#pg6>. Gate 1 is a per-term significance screen using the Bonferroni t-test, while Gate 2 is a partial F-test with validation rollback to ensure out-of-sample improvement <ref:2608. p. 700466#pg6>. The framework also uses a frequencydomain whiteness guard, replacing variance thresholds with the Ljung–Box test on the validation residual <ref:2608. p. 100466#pg6>.

Impossibility and identifiability

The impossibility result shows that insufficiently rich excitation makes recovery impossible for any estimator <ref:2608.00466#pg2>. Specifically, at constant speed the disk traces a near-circular orbit on which all candidate signatures collapse onto a low-order trigonometric span, causing the signature matrix to lose column rank and rendering gyroscopic coupling and Duffing nonlinearity unidentifiable <ref:2608. p. 000466#pg6>. A run-up sweeps the excitation frequency, restoring full column rank and satisfying the persistent-excitation condition necessary for unique coefficient recovery <ref:2608. p. 600466#pg6>.

Application to Rotating Machinery

The framework is applied to a four-degree-of-freedom vertical flexible rotor with Duffing supports, where only the lateral displacements x(t), y(t) are measured while the tilt angles α(t), β(t) remain hidden <ref:2608.00466#pg6>. SPIRAL-PO correctly identifies the governing equation structure from x, y alone, and an Extended Kalman Filter built on the validated dynamics can subsequently reconstruct all four states—including α and β—with sub-2% relative error against synthetic ground truth <ref:2608. p. 100466#pg6>. The method confirms that gyroscopic coupling, the Duffing nonlinearity, and the translation–tilt cross-coupling are recovered from noisy coast-down data <ref:2608. p. 000466#pg6>.

State reconstruction via extended Kalman filtering

An Extended Kalman Filter is built over the full 8-state system z = qobs U qhid, using (59)–(61) as the process model and only x, y as observations <ref:2608.00466#pg6>. This step reconstructs the hidden tilt states α(t), β(t) from observed displacements alone <ref:2608. p. 000466#pg6>. The filter is not merely echoing the measured channels; it demonstrates that the correct equation structure is discovered, and the hidden-state trajectory is then reconstructed from that model <ref:2608. p. 100466#pg6>.

Comparison with related methods

SPIRAL-PO differs from delay-coordinate SINDy by working in original physical coordinates with a physics-structured library, so its discovered terms are the physical couplings themselves, and the subsequent EKF stage reconstructs the actual tilt trajectories <ref:2608.00466#pg6>. While delay-coordinate methods reconstruct an attractor diffeomorphic to that of the full state, SPIRAL-PO works in original physical coordinates with a physics-structured library, so its discovered terms are the physical couplings themselves, and the subsequent EKF stage reconstructs the actual tilt trajectories <ref:2608. p. 100466#pg6>. SPIRAL-PO’s contributions include an impossibility result specialized to the rotor as Corollary 1, proving that constant-speed data cannot recover gyroscopic and Duffing terms at any sample size or sensor precision <ref:2608. p. 000466#pg6>. The framework preserves physical interpretability by having every admitted term map to a mechanical mechanism <ref:2608. p. 100466#pg6>.

Conclusion

SPIRAL-PO extends the SPIRAL framework to the practically important class of nonlinear dynamical systems where the full state is not measurable <ref:2608.00466#pg2>. By reformulating the identification problem in the observable subspace, building a physics-structured candidate library from dimensional analysis of hidden-variable dynamics, replacing variance-based guards with a frequencydomain whiteness test, and using a multi-output residual neural network, the method discovers interpretable, statistically validated equations of motion from partial observations only <ref:2608. p. 000466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608. p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 1 p. 100466#pg6>. Beyond discovering governing equations, an Extended Kalman Filter built on the SPIRAL-PO-validated dynamics reconstructs the hidden tilt trajectory itself from x(t), y(t) alone, with sub-2% relative error against ground truth on synthetic data <ref:2608. p. 100466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.00466#pg6>. The framework’s recovery of the gyroscopic coupling, the bearing-mediated Duffing nonlinearity, and the tilt-mediated linear cross-coupling from x(t) and y(t) measurements alone is confirmed numerically against a synthetic 4-DOF system <ref:2608.004…

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Projection Residual Formulation: The system can identify hidden-state effects by fitting a multi-output residual network to a projection residual that maps hidden-state dynamics onto observable coordinates as structured candidate library terms.

  2. Physics-Structured Library Construction: The AI system builds a library of candidate terms based on dimensional analysis and conservation laws, ensuring every term carries the same physical units as the observable acceleration, which improves model interpretability.

  3. Iterative Model Refinement: The method uses an iterative loop where it asks whether the current model leaves unexplained structure in the residual, allowing it to determine what physics-consistent terms would best account for it.

  4. Robust Statistical Screening: The system employs a two-gate screening process, using a Ljung-Box test as a whiteness guard and a partial F-test with validation rollback to ensure admitted terms are both significant on their own and improve out-of-sample prediction.

  5. Physical Interpretation of Coefficients: Every admitted coefficient is mapped back to its physical origin, such as gyroscopic coupling or tilt-mediated cross-channel coupling, allowing engineers to check the term against specific mechanical specifications.

  6. State Reconstruction: After identifying the governing observable dynamics, an Extended Kalman Filter is built on this validated model to reconstruct hidden states like tilt angles from only the measured lateral displacements alone, achieving sub-2% relative error.

Sources

Related papers