Transformer noise anomaly detection method based on seasonal trend decomposition and time series prediction

Aiming at the problem that traditional transformer noise monitoring methods cannot provide early warning of noise anomalies, this paper proposes a transformer noise anomaly detection method based on STL decomposition-Transformer-Fast Fourier Transform and Bayesian change point detection. Firstly, seasonal-trend decomposition (STL) is used to decompose transformer noise data to extract seasonal, trend, and residual components. Then, for each decomposed component, the Transformer model is used to perform time series prediction of future noise changes. For the seasonal component, Fast Fourier Transform (FFT) is applied to detect periodic anomalies, while the trend component uses Bayesian change point detection to identify inflection points and abnormal points in the trend. Finally, the detection results are fused by weighted superposition to output the detection results including prediction results, abnormal intervals, and trend inflection points. Experimental results show that compared with traditional methods, this method can more accurately warn of abnormal situations before noise exceeds the standard, significantly improving the efficiency and prediction accuracy of transformer noise monitoring.