Low-frequency noise removal and acoustic spectral distribution assessment method for high-voltage power transformers with varying service lifetimes

Abstract

The acoustic signals of power transformers serve as key indicators for assessing operational statuses and detecting internal mechanical issues. However, low-frequency noise, such as fan noise, often obscures critical features. This study introduces an adaptive low-frequency denoising algorithm based on the filtered-x least mean square (FxLMS) model. By optimising the step size and convergence factor, the algorithm resolves the peak offset issue in traditional LMS methods, enhances denoising performance and computational efficiency, and demonstrates effectiveness in practical scenarios. Using denoised acoustic data, quantitative analysis based on information entropy (frequency complexity analysis (FCA)) evaluates changes in mechanical properties. The analysis indicates that healthy transformers exhibit lower FCA values, whereas aged transformers show values approximately double those of healthy units, reflecting the mechanical changes associated with normal ageing. Further analysis of transformers commissioned in 2004 reveals that FCA values of abnormally aged transformers exceed three times those of typically aged units, indicating severe mechanical degradation. These findings demonstrate that combining the FxLMS algorithm with FCA analysis effectively extracts denoised acoustic features and distinguishes between healthy states, normal ageing, and abnormal ageing in transformers.