Anomaly detection in smart power grids with graph-regularized MS-SVDD: a multimodal subspace learning approach
Anomaly detection in smart power grids with graph-regularized MS-SVDD: a multimodal subspace learning approach
Authors: Thomas Debelle, Fahad Sohrab, Pekka Abrahamsson, Moncef Gabbouj Date: 2026-04-24 Paper ID: openalex:2502.15793
Summary
This paper introduces a generalized Multimodal Subspace Support Vector Data Description (MS-SVDD) framework designed for anomaly detection in complex, heterogeneous smart grid data. By incorporating Laplacian graph-embedded regularization, the model maps multiple sensor modalities into a shared low-dimensional subspace while maintaining modality-specific relational structures. Empirical evaluation on a multi-modality smart grid dataset shows that this structural approach significantly enhances robustness over traditional one-class classification methods. The work provides a systematic methodology for embedding relational graph priors into subspace-based anomaly detection models.
Key Contributions
- Proposed a Generalized Multimodal Subspace Support Vector Data Description (MS-SVDD) model for anomaly detection in smart grids.
- Integrated Laplacian-based graph regularization into MS-SVDD to explicitly preserve multimodal structural dependencies.
- Demonstrated improved event detection robustness on a three-modality smart grid dataset compared to conventional subspace learning approaches.
Open Questions & Future Work
Key Concepts
- graph-regularized-multimodal-subspace-learning: A framework for anomaly detection that embeds modality-specific structural priors via Laplacian regularizers into a shared low-dimensional subspace.
Archivist Review
The proposed model represents a specific, reusable approach to multimodal anomaly detection by integrating graph priors into subspace learning. I have approved this as a concept. I have also approved the open question regarding real-time deployment scalability after renaming it to be more specific to critical infrastructure. Other potential candidates were rejected for being either too generic, local to the specific implementation, or redundant.
Approved Concepts
- Graph-Regularized Multimodal Subspace Learning: This represents a novel, reusable approach to integrating graph priors into subspace-based anomaly detection, which is broadly applicable to heterogeneous time-series data.
Approved Open Questions
- Real-time infrastructure deployment scalability: Crucial for transitioning from high-dimensional latent modeling to reliable, time-sensitive operation in real-world critical infrastructure.
Rejected Candidates
- [open_question] Real-time deployment and scalability (
real-time-deployment-scalability) - duplicate_existing: The title is too generic and conflicts with existing patterns in the vault.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2502.15793
- paper_id
- 2502.15793
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
-
- concept_slugs
- graph-regularized-multimodal-subspace-learning
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-27T06:02:21Z