Stochastic Subspace Identification

Stochastic subspace identification (SSI) estimates modal properties from measured structural response time histories by fitting a stochastic state-space model. It is an output-only method: the excitation is treated as unknown process input rather than a measured force.

A common discrete-time stochastic model is:

where \mathbf{x}_{k} is the state vector, \mathbf{y}_{k} is the measured response vector, \mathbf{w}_{k} represents process excitation and \mathbf{v}_{k} represents measurement noise. SSI estimates a realization of the system matrices from output data, then converts the poles of \mathbf{A} into modal frequencies and damping ratios.

Engineering Role

SSI is used when engineers need a time-domain OMA method that can handle multi-channel response data, closely spaced modes, damping estimation and stabilization diagrams. It is common in bridge monitoring, building floor vibration, tower testing, offshore structures, aircraft ground or taxi response studies and large equipment foundations.

A practical SSI workflow usually includes:

• sensor layout and calibration review;
• anti-alias filtering and sampling-rate selection;
• detrending, segmentation and stationarity checks;
• covariance-driven or data-driven subspace identification;
• repeated model orders and stabilization diagrams;
• conversion of stable poles into frequency, damping and mode-shape estimates;
• comparison with finite-element modes, frequency-domain decomposition results or baseline tests.

The method is powerful because it uses the time structure and spatial structure of multi-sensor data. It is also easy to overfit. Model order, block rows, data length, noise, nonstationary excitation and harmonic operating orders can all create convincing but false poles.

Worked Example: Convert an SSI Pole to Frequency and Damping

An SSI run on bridge acceleration data sampled at:

has a sample interval of:

A stable complex pole appears repeatedly in the stabilization diagram. In polar form, the discrete pole is:

The continuous-time decay rate is:

The damped circular frequency is:

The undamped circular frequency is approximately:

so the modal frequency is:

The damping ratio is:

or:

Engineering comment: this pole is plausible for a lightly damped civil structure if it appears with stable frequency, damping and mode shape across several model orders. The result should not be accepted from one model order alone. The engineer checks neighbouring poles, mode-shape consistency, harmonic sources, sensor coverage, temperature or traffic changes and agreement with independent spectral evidence.

Distinction from Related Terms

Stochastic subspace identification is not operational modal analysis as a whole. OMA is the broader output-only discipline; SSI is a state-space identification method within it.

Stochastic subspace identification is not frequency-domain decomposition. FDD works with response spectral matrices at frequency lines. SSI works with time-domain response data, covariance or projection matrices and identified state-space poles.

Stochastic subspace identification is not a generic state-space model. A state-space model is a representation. SSI is a method for estimating one from output-only data.

Stochastic subspace identification is not a Kalman filter. Kalman filtering estimates states using a model and measurements. SSI estimates a model from measured outputs, although stochastic realization theory and innovation forms are related to Kalman filtering.

Stochastic subspace identification is not a frequency response function. It does not require measured input force and does not produce a force-normalized response/input ratio.

Validation and Common Mistakes

A defensible SSI result states sensor locations, response units, sampling rate, filtering, data length, stationarity checks, identification variant, block rows, model-order range, stabilization criteria, pole-selection rules, uncertainty, mode-shape normalization and environmental or operational conditions.

Common mistakes include:

• accepting poles that stabilize numerically but have nonphysical mode shapes;
• using too little data for the frequency and damping resolution required;
• choosing model order only by visual preference;
• mistaking harmonic forcing, traffic rhythm or machinery orders for structural modes;
• reporting damping without uncertainty or repeatability checks;
• comparing SSI modes with finite-element modes without coordinate and normalization review;
• assuming output-only state-space identification can recover absolute input forces or force-normalized FRFs.