Abstract
Online monitoring systems continuously collect massive multivariate time-series data from converter transformers. Accurate outlier detection in these data is essential for identifying sensor faults, communication errors, and incipient equipment failures, thereby ensuring reliable condition assessment and maintenance decisions. However, the complex characteristics of transformer monitoring data-including non-Gaussian distributions from diverse operational modes, high dimensionality, and multi-scale temporal dependencies-render traditional outlier detection methods ineffective. This paper proposes a Multi-View Clustering-based Outlier Detection (MVCOD) framework that addresses these challenges through complementary data representations. The framework constructs four complementary data views-raw-differential, multi-scale temporal, density-enhanced, and manifold representations-and applies four detection algorithms (K-means, HDBSCAN, OPTICS, and Isolation Forest) to each view. An adaptive fusion mechanism dynamically weights the 16 detection results based on quality and complementarity metrics. Extensive experiments on 800 kV converter transformer operational data demonstrate that MVCOD achieves a Silhouette Coefficient of 0.68 and an Outlier Separation Score of 0.81, representing 30.8% and 35.0% improvements over the best baseline method, respectively. The framework successfully identifies 10.08% of data points as outliers with feature-level localization capabilities. This work provides an effective and interpretable solution for ensuring data quality in converter transformer monitoring systems, with potential applications to other complex industrial time-series data.