The accuracy of the intelligent transformer fault diagnosis method based on DGA data is easily affected by the input characteristics and the difficulty of parameter selection of the extreme learning machine. A transformer fault diagnosis method based on the neighborhood rough set and the adaptive mutation particle swarm extreme learning machine algorithm is proposed. First, the initial feature set of transformer faults is established based on the various DGA fault diagnosis standards, and the key feature indicators with high attribute importance are obtained after the neighborhood rough set analysis; secondly, when optimizing the parameters of the extreme learning machine for the particle swarm algorithm, it is easy to be premature and fall into the local maximum. For excellent defects, an improved particle swarm optimization algorithm with premature self-check mutation mechanism is proposed to optimize the extreme learning machine; finally, through the transformer DGA data example diagnosis, it is compared with the IEC three-ratio method and the diagnostic performance of different combinations of extreme learning machines. It shows that the diagnosis accuracy of the proposed method is higher.