Multi-View Clustering-Based Outlier Detection for Converter Transformer Multivariate Time-Series Data.

Area of Science:

• Electrical Engineering
• Data Science
• Artificial Intelligence

Background:

• Converter transformers generate massive multivariate time-series data requiring accurate outlier detection.
• Traditional methods fail due to non-Gaussian distributions, high dimensionality, and multi-scale temporal dependencies in transformer data.
• Identifying outliers is crucial for detecting sensor faults, communication errors, and equipment failures.

Purpose of the Study:

• To propose a novel Multi-View Clustering-based Outlier Detection (MVCOD) framework.
• To address the challenges of complex transformer monitoring data for effective outlier detection.
• To improve the reliability of condition assessment and maintenance decisions for converter transformers.

Main Methods:

• Constructing four complementary data views: raw-differential, multi-scale temporal, density-enhanced, and manifold representations.
• Applying four outlier detection algorithms (K-means, HDBSCAN, OPTICS, Isolation Forest) to each view.
• Utilizing an adaptive fusion mechanism to dynamically weight 16 detection results based on quality and complementarity.

Main Results:

• MVCOD achieved a Silhouette Coefficient of 0.68 and an Outlier Separation Score of 0.81.
• Demonstrated significant improvements of 30.8% and 35.0% over the best baseline methods.
• Successfully identified 10.08% of data points as outliers with feature-level localization.

Conclusions:

• The MVCOD framework offers an effective and interpretable solution for ensuring data quality in converter transformer monitoring.
• The proposed method significantly outperforms existing techniques in complex time-series data analysis.
• Potential applications extend to other industrial time-series data requiring robust outlier detection.