SSRL

Official repository for the paper:

Disentangling Damage from Operational Variability: A Label-Free Self-Supervised Representation Learning Framework for Output-Only Structural Damage Identification

Overview

This repository contains the implementation of the self-supervised representation learning framework proposed in the paper, with the current public code focused on the MCC5 gearbox dataset.

The method learns two latent representations from raw vibration signals:

• z_dmg: damage-sensitive representation
• z_ndmg: nuisance-related representation

The model is trained with:

• time-domain reconstruction loss
• PSD reconstruction loss
• self-supervised VICReg loss on baseline healthy samples

Damage identification is then performed using Mahalanobis distance in the learned latent space.

Demonstratiev Data

Download the processed MCC5 dataset from: https://drive.google.com/file/d/1vgFMbcAKVf_FN38JXinLZzP-fRx1NbYa/view?usp=drive_link Then place the file here: data/MCC5.pt

Train

The default config is stored in: scripts/timeseries_MCC5/configs.py Run training with: python scripts/timeseries_MCC5/train.py

Evaluate

Run evaluation with: python scripts/timeseries_MCC5/test_v2.py

Ablation study

Run ablation study with: python scripts/timeseries_MCC5/run_sweep_all.py

Handcrafted-feature baseline

An optional handcrafted-feature baseline is provided in: scripts/artificial_features_MCC5/evaluation_v1.py

Citation

If you use this repository, please cite:

@article{jian2026ssrl, title={Disentangling Damage from Operational Variability: A Label-Free Self-Supervised Representation Learning Framework for Output-Only Structural Damage Identification}, author={Jian, Xudong and Stoura, Charikleia and Scandella, Simon and Chatzi, Eleni}, journal={To be assigned}, year={2026} }

Contact

Xudong Jian ETH Zurich xudong.jian@ibk.baug.ethz.ch