Research and application of MFCC-LSTM based on improved ZOA algorithm to optimize transformer voice-print diagnosis

Regarding the issue of monitoring the operational status of power transformers, this paper presents an improved power transformer acoustic fingerprint monitoring method based on MFCC. Firstly, through on-site investigations and data collection, acoustic fingerprint samples of transformers under various operational states are obtained, and they are preprocessed and data-augmented to construct an acoustic fingerprint sample set encompassing various noise characteristics. Subsequently, by introducing a weighting optimization method based on the F-ratio, the feature extraction method of Mel Frequency Cepstral Coefficients (MFCC) is enhanced to reduce redundant information in the acoustic fingerprint feature vector. Moreover, the hyper-parameters of the LSTM network are optimized using the improved zebra algorithm to enhance the performance of the acoustic fingerprint recognition model. Finally, the improved acoustic fingerprint recognition algorithm is practically applied and tested at the transformer acoustic fingerprint monitoring test points, verifying the practicability and reliability of the algorithm.