Disentangling Damage from Operational Variability: A Label-Free Self-Supervised Representation Learning Framework for Output-Only Structural Damage Identification
Disentangling Damage from Operational Variability: A Label-Free Self-Supervised Representation Learning Framework for Output-Only Structural Damage Identification
Authors: Xudong Jian, Charikleia Stoura, Simon Scandella, Eleni Chatzi Date: 2026-04-21 Paper ID: openalex:2604.19658
Summary
This paper introduces a label-free, self-supervised representation learning framework designed to isolate structural damage signals from confounding environmental and operational variability. The proposed model utilizes a dual-latent autoencoder architecture augmented with VICReg for invariance learning and a specific frequency-domain constraint to ensure reconstruction fidelity. By eliminating the need for labeled damage or environmental data, the framework provides a robust method for structural health monitoring that generalizes well across real-world vibration-based assets.
Key Contributions
- Proposes a self-supervised autoencoder architecture that uses VICReg to disentangle damage-related latent features from operational variability.
- Introduces a frequency-domain constraint to ensure latent space fidelity to input power spectral densities.
- Demonstrates robustness and generalization on real-world vibration datasets (bridge and gearbox) without requiring damage or environmental labels.
Open Questions & Future Work
Key Concepts
- disentangled-representation-learning-damage-identification: A self-supervised framework using dual-latent autoencoders and invariance regularization to separate damage-related features from operational/environmental noise.
Archivist Review
I approved the central mechanism of disentangling damage from operational noise as a standalone concept due to its broad applicability in industrial anomaly detection. I also approved the open question regarding sequence-level decision strategies as it represents a critical bridge between local representation learning and system-level monitoring reliability. Other candidates were rejected for being either too generic or performance-oriented rather than mechanism-oriented.
Approved Concepts
- Disentangled Representation Learning for Damage Identification: It introduces a method to separate physical damage signals from operational noise using VICReg and spectral constraints without requiring labeled data, which is highly reusable in industrial anomaly detection.
Approved Open Questions
- Robust temporal decision strategies for SHM: This addresses the fundamental reliability gap between local model inference and global structural state assessment in SHM.
Rejected Candidates
- [concept] Self-Supervised Label-Free Disentangled Representation Learning (
self-supervised-label-free-disentangled-representation-learning) - other: The title is overly generic; renamed to a more specific and descriptive term for the vault. - [open_question] Improving weak damage detection (
improving-weak-damage-detection) - low_impact: This is a performance-based goal (better detection of weak signals) rather than a clear technical bottleneck or mechanism-level research question.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.19658
- paper_id
- 2604.19658
- paper_source
- openalex
- domain
- time-series
- tags
- time-seriesself-supervised-learningrepresentation-learninganomaly-detectionstructural-health-monitoring
- architectures
-
- datasets
-
- concept_slugs
- disentangled-representation-learning-damage-identification
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-24T05:52:41Z