Enhanced Spectral Ensemble Clustering for Fault Diagnosis: Application to Photovoltaic Systems

Abstract

The role of clustering in unsupervised fault diagnosis is significant, but different clustering techniques can yield varied results and cause inevitable uncertainty. Ensemble clustering methods have been introduced to tackle this challenge. This study presents a novel integrated technique in the field of fault diagnosis using spectral ensemble clustering. A new dimensionality reduction technique is proposed to intelligently identify faults, even in ambiguous scenarios, by exploiting the informative segment of the underlying bipartite graph. This is achieved by identifying and extracting the most informative sections of the bipartite graph based on the eigenvector centrality measure of nodes within the graph. The proposed method is applied to experimental current-voltage (I-V) curve data collected from a real photovoltaic (PV) platform. The obtained results remarkably improved the accuracy of aging fault detection to more than 83.50%, outperforming the existing state-of-the-art approaches. We also decided to separately analyze the ensemble clustering part of our FDD method, which indicated surpassing performance compared to similar methods by evaluating commonly used datasets like handwritten datasets. This proves that the proposed approach inherently holds promise for application in various real-world scenarios that are indicated by ambiguity and complexity.

Keywords

• Enhanced spectral ensemble clustering (ESEC)
• bipartite graph partitioning
• eigenvector centrality
• neural networks
• fault detection and diagnosis (FDD)
• photovoltaic (PV) system