SMA-optimized SVM transformer state identification method based on acoustic vibration feature differentiation

Aiming at the limitation that the acoustic and vibration features extracted based on Mel frequency cepstrum coefficient (MFCC) cannot clearly describe the energy distribution of the transformer signal itself, and thus the accuracy is not high when applied to the identification of transformer mechanical loosening, a transformer fault identification method that prioritizes the acoustic and vibration feature distinction is proposed. Firstly, based on the XGBoost contribution indexing combined with rough set analysis, the MFCC features were distinguished explicitly and implicitly,and the explicit features were found to contribute more to the state identification. An SMA optimization model with Focal loss was established and the weight ranges were set for the SVM inputs by explicitness and implicitness. Finally, the optimized SVM was used to train and analyze the transformer real samples. The results show that this recognition method achieves an average accuracy of 98. 83%, which is 2. 48% higher than the recognition accuracy of parameter-optimized SVM with fewer variance fluctuation. Compared with PSO, WOA and GOA algorithms, SMA algorithm is more prominent in feature global optimization and loss convergence. In addition, the method has some robustness, with accuracy dropping within 0. 3% after introducing 5% interference samples, thus having anti-interference value in the actual operating environment of transformers. © 2023 Editorial Department of Electric Machines and Control. All rights reserved.