Experimental assessment of polynomial nonlinear state-space and nonlinear-mode models for near-resonant vibrations

Maren Scheel, Gleb Kleyman, Ali Tatar, Matthew R. W. Brake, Simon Peter, Jean-Philippe Noël, Matthew S. Allen, Malte KrackarXiv 2020 · arXiv

TL;DR

This paper compares two distinct methods for identifying nonlinear dynamic models from experimental data: a nonlinear-mode model and a polynomial nonlinear state-space (PNLSS) model. The study highlights their strengths and weaknesses in predicting near-resonant vibrations, revealing that while nonlinear-mode models excel at isolated resonances, PNLSS models offer broader applicability but are highly sensitive to training data and can exhibit unstable predictions.

What Problem Does This Paper Address?

Accurately modeling the dynamic behavior of structures is crucial for safe and efficient design, especially as lightweight designs and complex interactions (like joints) introduce significant nonlinearities. While linear models are well-established, predicting nonlinear vibrations, particularly near resonance, remains a challenge. Existing nonlinear system identification methods have distinct limitations that are not always clearly articulated, making it difficult for users to choose the most appropriate approach for their specific application. This paper aims to bridge this gap by experimentally comparing two prominent nonlinear identification techniques, focusing on their ability to predict vibrations near resonance and under transient conditions.

Methods

The paper investigates two primary methods for nonlinear system identification:

  • Nonlinear-Mode Model: This approach identifies a single-degree-of-freedom oscillator model based on amplitude-dependent modal properties (frequency, damping, and mode shape). It relies on controlled, steady-state excitations near resonance and a phase-locked loop (PLL) controller to isolate and characterize nonlinear modes. This method does not require a predefined functional form for the nonlinearity.
  • Polynomial Nonlinear State-Space (PNLSS) Model: This method identifies a time-discrete state-space model augmented with multivariate polynomial nonlinearities in states and inputs. It utilizes advanced system identification tools and can be trained on broadband random inputs, such as multisines. This "black-box" approach offers the advantage of potentially modeling a wider range of nonlinearities and handling transient excitations.

The study applies both methods to two experimental test rigs: a cantilevered beam with repelling magnets (exhibiting hardening stiffness nonlinearity) and a bolted beam (featuring friction and contact nonlinearities). The identified models are then rigorously tested by predicting their dynamic response to various excitations, including steady-state sine waves and swept-sine signals, and compared against reference measurements.

Datasets

The paper utilizes experimental data collected from two distinct test specimens:

  • Cantilevered Beam with Repelling Magnets: Data was generated through various measurements, including linear modal analysis, nonlinear modal analysis, and PNLSS identification using multisine excitations at different RMS amplitude levels (0.5 N, 2.6 N, and 5.2 N). Reference measurements were also performed using stepped sine signals.
  • Two Beams with Bolted Joints (BRB): Similar to the first specimen, data was collected for linear modal analysis, nonlinear modal analysis, and PNLSS identification using multisine excitations at different RMS amplitude levels (10 N, 20 N, 30 N, and 50 N). Reference measurements included stepped sine signals.

No new benchmark datasets were introduced; rather, the study leverages data from these carefully designed experimental setups to compare the performance of the two identification methods.

Key Insights and Ideas

  • Nonlinear-Mode Models Excel at Resonance: The nonlinear-mode model, by design, provides highly accurate predictions for steady-state vibrations precisely at resonance peaks and in their immediate vicinity, especially for systems with well-spaced modes.
  • PNLSS Models Offer Broader Applicability but Face Challenges: PNLSS models can, in principle, capture a wider dynamic range, including transient behavior, and are less restrictive regarding excitation types. However, their performance is heavily dependent on the training data, and they struggle to accurately represent complex nonlinearities (like friction and contact) with low-order polynomials.
  • Input-Dependence is a Major Hurdle for PNLSS: PNLSS models, particularly those trained on limited or specific excitation levels, exhibit significant input-dependence. This means their predictive capability can degrade substantially when tested with inputs that differ in amplitude spectrum or level from the training data, sometimes leading to unstable predictions even for inputs similar to the training data.
  • Transient Predictions are Tricky for Both: While PNLSS models generally perform better for transient sine sweeps than for steady-state harmonic excitations, predicting these responses accurately can still be challenging. The nonlinear-mode model's accuracy for transient sweeps is moderate, and its predictions can become unstable if the physical mechanism of force drops during sweeps is ignored.
  • Model Extrapolation is Risky: Both model types can suffer from extrapolation errors. For PNLSS models, this is particularly problematic when predicting responses at amplitude levels beyond the training data. For nonlinear-mode models, extrapolation outside the identified amplitude range leads to inaccuracies.

Key Results

  • Nonlinear-Mode Model Accuracy: For steady-state harmonic excitations near resonance, the nonlinear-mode model consistently provided predictions very close to the measured responses for both test specimens, often outperforming PNLSS models.
  • PNLSS Model Performance Varies:
    • For the magnetic beam, PNLSS models trained at 2.6 N RMS and mixed levels offered the best predictions for steady-state responses, though models trained at higher levels (5.2 N RMS) or lower levels (0.5 N RMS) showed significant underestimation of amplitudes and modal frequencies.
  • For the bolted joint specimen, PNLSS models showed more variability. Models trained at 30 N RMS and mixed levels provided reasonable accuracy for transient sweeps, but no single PNLSS model consistently performed well across all excitation levels and types.
  • In general, PNLSS models trained at higher excitation levels tended to be more input-dependent, and models trained with mixed levels sometimes improved the validity range but still struggled with accuracy.
  • Unstable Predictions: Several PNLSS models, particularly those trained at higher excitation levels or mixed levels for the bolted joint specimen, were found to predict unbounded (unstable) responses, even when tested with data similar to their training set.
  • Model Form Error: For the bolted joint specimen, the non-smooth nonlinearities due to contact interactions were not well-approximated by low-order polynomials, leading to model form errors and qualitatively incorrect predictions of steady-state vibrations in some cases.

Limitations and Future Work

  • Nonlinear-Mode Model Limitations:
    • Strictly valid only around isolated resonances and for essentially mono-harmonic excitation.
  • Sensitive to uncontrolled sweeps if the physical mechanism of force drop is ignored.
  • Extrapolation outside the identified amplitude range leads to errors.
  • PNLSS Model Limitations:
    • Highly input-dependent, especially for higher excitation levels.
  • May yield unbounded responses due to extrapolation or model errors.
  • Low-order polynomials struggle to accurately represent complex nonlinearities like friction and contact.
  • Model form error is introduced when polynomials cannot fully capture the physics.
  • Computational cost increases significantly with polynomial order.
  • Future Work Suggested:
    • Investigate if steady-state predictions of PNLSS models improve when trained with steady-state signals, similar to the nonlinear-mode model's approach.
  • Explore increasing the polynomial order in PNLSS models to potentially improve accuracy, while acknowledging the increased computational cost.
  • Further research into defining the valid regime for PNLSS models to prevent instability due to extrapolation.
  • Develop methods to better handle complex nonlinearities like friction and contact within polynomial frameworks or explore alternative model structures.

Summary updated

BibTeX
@misc{scheel2020experimental,
  author = {Maren Scheel and Gleb Kleyman and Ali Tatar and Matthew R. W. Brake and Simon Peter and Jean-Philippe Noël and Matthew S. Allen and Malte Krack},
  title = {{Experimental assessment of polynomial nonlinear state-space and nonlinear-mode models for near-resonant vibrations}},
  year = {2020},
  eprint = {2011.08520},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2011.08520},
}

Loading related content...

Figure