Skip to content

Modal Basis Identification

System-identification method — instantiates Independent Generating Set Design

Identifies a system's natural modes — its characteristic shapes of motion, each with its own frequency — as a small, physically interpretable basis for how it behaves.

Modal Basis Identification extracts a physical system's natural modes — its characteristic shapes of motion, each vibrating at its own frequency with its own damping — and uses those mode shapes as a basis for the system's behaviour. Its defining move is that the generators are discovered from the system's own physics and each one means something: a first bending mode, a lateral sway, a torsional twist. Rather than seek a complete abstract basis, it keeps the dominant few modes that actually shape the response, and it maintains a crosswalk between those modal coordinates and the engineer's physical coordinates. Where an algebraic eigendecomposition factors a known matrix into a full spanning set, modal identification targets the handful of interpretable modes of a real, often measured system.

Example

Engineers worry that a new pedestrian footbridge feels lively underfoot. They instrument the deck with accelerometers and record its response to footfall and gusting wind, then identify its lowest modes from that measured response. Out come a small set of named shapes: a first vertical bending mode where the deck heaves at midspan, a lateral sway mode where the whole deck swings side to side, and a torsional mode where it twists. Each carries its own natural frequency and damping.

These few mode shapes are a basis: any complicated wobble of the deck is a weighted sum of them. When a dense crowd sets the deck swinging sideways, the modal decomposition points straight at the lateral sway mode — a real and famous failure mode for footbridges — and tells the engineers exactly where a damper tuned to that frequency will do the most good. A thousand-degree-of-freedom structure has become a story about three named modes.

How it works

  • Excite and observe. Drive or ambiently load the system and measure its response, or — when the governing operator is known — solve the physical eigenproblem directly.
  • Estimate modal parameters. From the response, identify each mode's frequency, damping, and shape (experimental modal analysis is the standard route).
  • Rank and truncate. Order modes by frequency or energy and retain the dominant few that govern the behaviour of interest; map each mode shape back to physical coordinates.

What distinguishes it is the deliberate stop at a sparse, interpretable set of physically meaningful modes rather than a full spanning basis.

Tuning parameters

  • Number of modes retained — the truncation order; more modes add fidelity, fewer give a sparser, more interpretable model. The core sparsity dial.
  • Excitation and sensor placement — where the system is driven and measured; determines which modes are observable and how cleanly they separate.
  • Frequency band of interest — restricting identification to the modes that matter for the use, since low modes usually dominate the response.
  • Mode-separation criterion — how aggressively to split closely-spaced or "veering" modes so two physical modes are not merged into one.
  • Normalization convention — mass-normalized versus unit-norm mode shapes, which fixes how the modal coordinates read.

When it helps, and when it misleads

Its strength is compression into meaning: a high-degree-of-freedom system collapses into a few modes you can name, reason about, and act on — add damping to this mode, stiffen against that one. The basis is not just small but interpretable.

It misleads when modes are hard to separate or the data is noisy. Closely-spaced or heavily-damped modes can be mis-identified or fused; noise can conjure spurious modes with no physical meaning; and truncation silently discards high-frequency content that occasionally matters. The Modal Assurance Criterion[1] is the standard check that two identified mode shapes are genuinely distinct rather than duplicates or artifacts. The classic misuse is keeping only the low modes and then being blindsided by a response the truncated basis cannot represent, or treating a noise-driven mode as real. The discipline is to check mode distinctness, confirm each mode is physically plausible, and retain enough modes to cover the frequency range the system will actually see.

How it implements the components

Modal Basis Identification fills the interpretable, physically-grounded side of the archetype:

  • candidate_generator_pool — the identified mode shapes are the generators, drawn from the system's own dynamics.
  • basis_semantics_crosswalk — each modal coordinate carries a physical meaning and maps back to physical degrees of freedom, giving a dictionary between the modal basis and the engineering frame.
  • sparsity_locality_and_interpretability_preference — the method deliberately keeps a sparse set of dominant, nameable modes rather than the full spectrum.

It does not guarantee a complete spanning basis or read exact rank — Full-Rank Eigendecomposition and Singular-Value Rank Diagnosis do — nor orthonormalize the often non-orthogonal measured modes (Gram–Schmidt Orthonormalization), nor audit conditioning under perturbation (Basis Conditioning and Perturbation Audit).

Notes

Modal identification and Full-Rank Eigendecomposition coincide when the system operator is known exactly — the modes are the eigenvectors of the mass-and-stiffness system. They diverge in modal identification's real habitat: the operator is unknown and the basis must be inferred from measured, noisy response, where identifiability, mode separation, and spurious modes become the governing concerns rather than clean algebra.

References

[1] The Modal Assurance Criterion (MAC) measures the consistency between two mode-shape vectors as a normalized squared correlation; values near one indicate the same mode, values near zero indicate independent modes. It is the standard check that identified modes are distinct rather than duplicates or noise. ↩