Abstract:This paper proposes a transformer vibration signal feature extraction method based on Grey Wolf Prey (GWO) optimized Variational Mode Decomposition (VMD) and SOM neural network algorithm.Firstly, the grey wolf predation algorithm is used to search for the optimal combination of influencing parameters for VMD. Based on the search results, the penalty factor and number of components of the VMD algorithm are set, and used to decompose the transformer vibration signal, screen for the best components, and further extract six representative feature quantities in the time and frequency domains, for a total of 120 sets.Using SOM neural network for feature classification and recognition.The method was applied to analyze the vibration signals of experimental transformers under four states: no-load, short circuit, normal load, and load fault. The results showed that the optimized VMD can avoid excessive decomposition and correctly extract the best components, with a state recognition accuracy of up to 99.37%. Finally, the method was applied to the fault diagnosis of in-service transformers, and the results showed that the method can effectively identify faults in in-service transformers.