Frequency Response Analysis Interpretation Using Numerical Indices and Machine Learning: A Case Study Based on a Laboratory Model

• Published in: IEEE access 2021, Vol.9, p.67051-67063
• Main Authors: De Andrade Ferreira, Regelii Suassuna, Picher, Patrick, Ezzaidi, Hassan, Fofana, Issouf
• Format: Article
• Language: English
• Subjects: Classification; Condition monitoring; Current measurement; Deformation; Fault diagnosis; Frequency analysis; Frequency measurement; Frequency response; frequency response analysis interpretation; Indexes; Laboratories; Linearity; Machine learning; Mathematical models; Model testing; Neural networks; numerical indices; Numerical models; Power transformer insulation; power transformers; Radial basis function; Reproducibility; Sensitivity analysis; Support vector machines; Transformers; Winding; Windings
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2021.3076154

Frequency response analysis is a powerful tool for mechanical fault diagnostics in power transformers. However, interpretation schemes still today depend on expert analyses, mainly because of the complex structure of power transformers. One of the fundamental shortcomings of experimental investigations is that mechanical deformations cannot be managed on real transformers to obtain data for different scenarios because they are too destructive. To address this issue in a systematic way, the current research used a specially designed laboratory transformer model that allows mechanical defects to be introduced so its frequency response can be evaluated under different conditions. The key feature of this model is the non-destructive interchangeability of its winding sections, allowing reproducibility and repeatability of frequency response measurements. Numerical indices were compared over key performance indicators (linearity, sensitivity and monotonicity). The analysis indicated that comparative standard deviation offered promising results for evaluation of mechanical deformations on the laboratory winding model given its monotonic behaviour, sensitivity and linear increase with fault severity. Additionally, support vector machine learning, radial basis function neural network and the statistical k-nearest neighbour method were used for fault classification with different strategies and configurations. While limited data from different transformers are used in the available literature, the approach discussed here considers 371 measurements from the same transformer model. The test results are supportive and demonstrate great accuracy when machine learning is used for winding fault classification.