SPECTRA: A Physics-informed Digital Twin for Real-time Structural Anomaly Inference under Operational Variability

Technical Summary: SPECTRA – A Physics-Informed Digital Twin for Real-Time Structural Anomaly Inference

Problem Statement
Structural health monitoring (SHM) is transitioning from simple damage detection to real-time decision support for critical infrastructure. A persistent challenge in this domain is distinguishing true structural anomalies from benign environmental and operational variability (EOV), such as temperature fluctuations, traffic loads, and wind. Classical SHM methods often rely on statistical pattern recognition, where deviations from a reference condition trigger alerts. However, as demonstrated by studies like the Z24 bridge, environmental effects can induce modal changes comparable to or larger than those caused by damage. Consequently, purely statistical novelty detectors frequently generate false alarms when operational conditions change, undermining their utility for safety-critical management. Furthermore, existing methods often lack physical interpretability, making it difficult for engineers to trust automated decisions.

Methodology: The SPECTRA Framework
The paper introduces SPECTRA (Self-adaptive Physics-informed Eigen-Compressed Twin for Real-time Anomaly inference), a framework designed to infer anomalies based on the disagreement between measured responses and a physics-informed "healthy" digital twin, rather than statistical novelty alone. The workflow consists of five integrated components:

1. Physics-Informed Healthy Twin: A reference model representing the expected undamaged behavior of the structure. Crucially, this twin incorporates environmental compensation (e.g., temperature-dependent stiffness multipliers) to account for benign global variations, ensuring that changes due to EOV are not misinterpreted as damage.
2. Eigen-Compressed Dynamic Representation: The system response is projected onto a reduced basis of dominant modes. This allows for efficient real-time inference while retaining the essential dynamic behavior of the structure.
3. Twin Innovation Signal: The core metric is the "innovation," defined as the difference between the measured structural response and the response predicted by the healthy twin. This signal quantifies the physical disagreement between the real asset and the healthy model.
4. Residual-Augmented Feature Vector & KPCA: A feature vector is constructed combining spectral descriptors (centroid, bandwidth, energy) and the twin innovation. These features are analyzed using Kernel Principal Component Analysis (KPCA) to generate two monitoring indices:
   • $T^2$ (Hotelling statistic): Measures variation within the learned feature manifold.
   • SPE (Squared Prediction Error): Measures novelty outside the manifold.
5. Persistent Decision Rule: To prevent isolated statistical excursions from triggering false alarms, the framework employs a gating mechanism. A raw alarm is only generated if the innovation index (physical disagreement) exceeds its limit and is corroborated by KPCA novelty ($T^2$ or SPE). This raw alarm must persist for a defined number of consecutive windows ($p$) to trigger a final structural anomaly flag.

Key Contributions
The primary contribution of this work is the integration of a physics-informed twin, eigen-compressed dynamics, twin-innovation features, and KPCA monitoring into a single anomaly inference framework. Unlike methods that rely solely on data-driven novelty, SPECTRA enforces a decision structure where statistical novelty serves as corroborating evidence, but the final decision is gated by physical inconsistency with a healthy model.

The paper validates this framework through seven numerical benchmarks designed to test specific failure modes and confounding factors:

• Nonlinear Drift: A smooth Duffing-type nonlinear drift to test specificity against benign nonlinearity.
• Abrupt Stiffness Loss: A sudden 30% stiffness reduction.
• Gradual Degradation: A linear ramp of stiffness loss over time.
• Breathing Stiffness: A bilinear, non-smooth stiffness mechanism simulating a breathing crack.
• EOV-Confounded Damage: Local stiffness loss occurring simultaneously with temperature-driven global stiffness changes.
• Damping Loss: An 82% loss of local damping, a mechanism often invisible to frequency-based methods.
• Operational-Only Negative Control: A scenario with changing temperature and loading but no structural damage.

Results
The numerical results demonstrate that SPECTRA successfully detects abrupt, gradual, nonlinear, and damping-related damage mechanisms with finite detection delays while maintaining high specificity:

• Damage Detection: The framework detected all damage cases (stiffness loss, breathing cracks, damping loss) with zero pre-damage persistent false alarms. The detection delay was determined by the monitoring window length and persistence requirement, not by threshold tuning.
• Specificity in Negative Control: In the operational-only negative control case, the KPCA SPE index rose above its upper control limit due to the change in operating regime. However, because the physics-informed twin (compensated for temperature) continued to explain the measured response, the innovation index remained within limits. Consequently, the persistent decision flag remained at zero. This confirms that SPECTRA can distinguish between feature-space novelty (caused by EOV) and structural anomaly.
• Damping Sensitivity: In the damping-loss case, where frequency-based features (centroid) remained stationary, the innovation channel provided the decisive signal, demonstrating the framework's ability to detect faults that do not manifest as frequency shifts.
• EOV Robustness: In the confounded damage case, the framework successfully attributed global stiffness changes to temperature compensation while isolating and flagging the local damage.

Significance and Claims
The paper claims that SPECTRA provides a reproducible route for testing physics-informed anomaly inference before deployment. Its significance lies in shifting the paradigm from "statistical novelty detection" to "physics-space disagreement inference." The authors argue that a structural anomaly should not be inferred from statistical novelty alone; rather, it must be interpreted as a persistent disagreement between the measured response and a healthy, physics-informed twin.

The study explicitly acknowledges limitations: the benchmarks use idealized twins with exact reference matrices and controlled excitations, meaning the absolute magnitudes of innovation signals are optimistic compared to field deployments where model fitting errors exist. Additionally, the study is confined to numerical shear-building models. However, the authors assert that the core contribution is the gated decision logic, which remains valid even if the twin is fitted from data, provided the physics-informed structure is maintained. The framework offers a technically interpretable route from measured response to anomaly decision, supporting the move toward resilient, real-time infrastructure management.