CEEMDAN-MRAL Transformer Vibration Signal Fault Diagnosis Method Based on FBG

• Published in: Photonics Vol. 12; no. 5; p. 468
• Main Authors: Jiang, Hong, Wang, Zhichao, Cui, Lina, Zhao, Yihan
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.05.2025
• Subjects: Accelerometers; Accuracy; Algorithms; Analysis; Bragg gratings; CNN; Demodulation; denoising algorithm; Electric transformers; Electromagnetic interference; Electromagnetism; Fault diagnosis; FBG; Fourier transforms; Methods; Neural networks; Signal processing; Signal to noise ratio; Time dependence; Time series; Venus; Vibration; Vibration analysis; Wavelet transforms
• DOI: 10.3390/photonics12050468
• ISSN: 2304-6732

In order to solve the problem that the vibration signal of transformer is affected by noise and electromagnetic interference, resulting in low accuracy of fault diagnosis mode recognition, a CEEMDAN-MRAL fault diagnosis method based on Fiber Bragg Grating (FBG) was proposed to quickly and accurately evaluate the vibration fault state of transformer.The FBG sends the wavelength change in the optical signal center caused by the vibration of the transformer to the demodulation system, which obtains the vibration signal and effectively avoids the noise influence caused by strong electromagnetic interference inside the transformer. The vibration signal is decomposed into several intrinsic mode functions (IMFs) by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), and the wavelet threshold denoising algorithm improves the signal-to-noise ratio (SNR) to 1.6 times. The Markov transition field (MTF) is used to construct a training and test set. The unique MRAL-Net is proposed to extract the spatial features of the signal and analyze the time series dependence of the features to improve the richness of the signal feature scale. This proposed method effectively removes the noise interference. The average accuracy of fault diagnosis of the transformer winding core reaches 97.9375%, and the time taken on the large-scale complex training set is only 1705 s, which has higher diagnostic accuracy and shorter training time than other models.