In order to further improve the diagnostic accuracy and diagnostic efficiency of carbon constrained type transformer operating state, a transformer state anomaly diagnosis method is proposed based on the improved k-medoids clustering theory. The transformer operating state is divided into four states including normal, early warning, abnormal and fault as the k-medoids clustering target. The model obtains the running state data of the transformer, and the voiceprint feature index uses the Mel cepstrum feature as the standard to measure the abnormal vibration and sound of the transformer. The transformer operation data is normalized and standardized based on its physical characteristics, which is convenient for clustering model input. Three measures are introduced to improve the traditional k-medoids clustering process, which are the initial clustering center generation based on local density evaluation, the adaptive substitution direction adjustment and the chaotic search of clustering update. The initial clustering center generation is optimized by the improvement measures and the optimization efficiency is improved to avoid prematurely falling into the local optimal clustering point. The historical operation data of a regional power grid transformer are divided into a training group and a test group. The k-medoids clustering network is optimized based on the training group, and the trained k-medoids clustering is used for the transformer operation state clustering and diagnosis of the test group. The correctness and effectiveness of the proposed model are verified.