Anomaly detection in smart power grids with graph-regularized MS-SVDD: a multimodal subspace learning approach.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Anomaly detection in smart power grids is challenging due to complex, dynamic, and multimodal sensor data.
• Existing one-class classification methods, like Subspace Support Vector Data Description (SVDD), struggle with multimodal data and fail to exploit cross-modal structural dependencies.

Purpose of the Study:

• To propose a generalized Multimodal Subspace Support Vector Data Description (MS-SVDD) model with graph-embedded regularization.
• To improve the robustness of anomaly detection in smart grids by better utilizing multimodal data.

Main Methods:

• Projecting multimodal data into a shared low-dimensional subspace.
• Preserving modality-specific structure using Laplacian regularizers.
• Employing graph-embedded regularization within the MS-SVDD framework.

Main Results:

• The graph-embedded MS-SVDD demonstrated improved robustness in smart grid event detection compared to conventional methods.
• The approach effectively integrates graph priors with multimodal subspace learning.
• Successful evaluation on a three-modality smart grid event time series dataset.

Conclusions:

• Integrating graph priors with multimodal subspace learning enhances anomaly detection robustness in critical infrastructure.
• The proposed MS-SVDD model offers a more effective approach for complex, high-dimensional, and multimodal anomaly detection.
• This work advances AI by embedding relational information into one-class models for robust learning.