An Information-Theoretic Method for Dynamic System Identification With Output-Only Damping Estimation

1. Problem Statement

The paper addresses a critical limitation in structural health monitoring (SHM) and early warning systems: the inaccurate estimation of damping in mechanical systems using output-only data.

• The Core Issue: Current operational modal identification methods (e.g., Frequency Domain Decomposition, Stochastic Subspace Identification) rely on empirical approaches that often yield poor damping estimates.
• The Consequence: Damping is the primary factor determining the duration of vibration events. Inaccurate damping estimation leads to misestimating how long a structure remains in a dangerous state. This compromises alert systems, which rely on vibration amplitudes and duration to trigger warnings for damage or anomaly detection.
• The Gap: Existing frameworks lack a robust method to update models specifically for "alert-type" excitations where input forces are unknown (output-only scenarios), leading to unreliable near real-time monitoring.

2. Methodology

The authors propose a novel Information-Theoretic Method based on Kullback–Leibler (KL) Divergence and Signal Energy to select the optimal dynamic model and estimate damping without input force data.

A. Theoretical Framework

• System Model: The structure is modeled as a standard second-order dynamic system: $M\ddot{x}(t) + C\dot{x}(t) + Kx(t) = f(t)$, where the input $f(t)$ is unknown stochastic excitation.
• Output-Only Analysis: The method utilizes Singular Value Decomposition (SVD) on the power spectral density matrix to extract natural frequencies and mode shapes.
• Residual Analysis: The core of the method involves comparing the measured output $y[k]$ with the model output $x[k]$. The residual $z[k] = x[k] - y[k]$ is analyzed.
• Information-Theoretic Metric:
  • The method treats the residual as a random variable.
  • It assumes a prior distribution ($Q$) of zero mean and unit covariance (representing an ideal, uncorrelated residual).
  • It calculates the Kullback–Leibler (KL) Divergence ($D_{KL}$) between the posterior distribution of the actual residuals ($P$) and the prior distribution ($Q$).
  • Objective: Minimize $D_{KL}(P||Q)$. The damping model that minimizes this divergence is considered the "optimal" model because it produces residuals that best match the statistical assumptions of a well-fitted model.

B. Model Selection Process

1. Scaling: The identified damping matrix is scaled by various factors (e.g., 10% to 500% of the initial estimate) to generate a set of candidate models.
2. Evaluation: For each candidate model, the KL divergence is calculated based on the residual signal energy.
3. Selection: The model with the minimum KL divergence is selected as the optimal model for predicting alert durations.

3. Key Contributions

1. Novel Damping Estimation: Introduces an information-theoretic approach (KL Divergence) to solve the under-determined problem of damping estimation in output-only scenarios, outperforming traditional empirical methods.
2. Alert Duration Optimization: Shifts the focus from merely matching peak amplitudes to accurately predicting the duration of vibration events, which is critical for safety alerts.
3. Robustness to Noise: Demonstrates that the method is robust to variations in the assumed prior distribution (covariance magnitude), relying on the divergence to identify the true system behavior.
4. Computational Efficiency: The method is computationally lightweight, capable of near real-time execution on standard hardware, making it suitable for online monitoring and digital twin applications.
5. Damage Detection Capability: Shows that significant deviations in the KL divergence or event duration metrics can serve as indicators of structural damage or anomalies, even when modal frequencies show negligible changes.

4. Experimental Results

The methodology was validated using two distinct datasets:

A. University of Bath Multi-Axis Simulation Table (MAST)

• Setup: A 6-DOF vibration table subjected to impact-type base excitation.
• Findings:
  • Five damping models (scaled from 10% to 500%) were tested.
  • The KL divergence plot showed a distinct minimum at the optimal damping level, clearly distinguishing it from under-damped and over-damped models.
  • The optimal model accurately matched the experimental event duration (time signal exceeds a threshold) across various alert thresholds (0–40 m/s²).
  • Alternative metrics like Rényi entropy and Jensen–Shannon divergence failed to show a clear minimum, proving the superiority of KL divergence for this specific application.

B. IASC–ASCE Benchmark Structural Health Monitoring Problem

• Setup: A 4-story steel frame structure with 15 accelerometers, subjected to hammer impacts and shaker excitations.
• Findings:
  • The method successfully identified the optimal damping model among five candidates.
  • Sensor Triggering: The model accurately predicted the number of sensors triggered at different amplitude thresholds, matching the experimental distribution.
  • Generalization: The model maintained accuracy even when tested with different excitation types (e.g., shaker vs. hammer) and different damage configurations (Case 1, 7, 9), provided the structural changes were within the training scope.
  • Damage Sensitivity: While the method struggled with extrapolation to completely unseen structural conditions without retraining, it successfully detected anomalies when the deviation from the calibrated model became statistically significant.

5. Significance and Future Directions

• Practical Impact: This method provides a powerful tool for early warning systems. By accurately estimating damping, it ensures that alerts are triggered for the correct duration, preventing false negatives (missing a dangerous event) or false positives (unnecessary shutdowns).
• Digital Twin Integration: The approach facilitates the creation of digital twins by continuously updating the dynamic model based on real-time residual analysis without requiring input force sensors.
• Limitations & Future Work:
  • The current approach requires retraining if the structural dynamics change significantly (e.g., major damage).
  • Future research suggests integrating online learning or Bayesian methods to incorporate prior knowledge and adapt to changing structural conditions in real-time.
  • The method is currently limited to global damping scaling; future work could explore localized damping identification.

In conclusion, the paper demonstrates that minimizing Kullback–Leibler divergence is a highly effective, computationally efficient strategy for refining dynamic system models in output-only scenarios, directly enhancing the reliability of structural safety monitoring and alert systems.