In order to improve the intelligent diagnosis level of an on-load tap-changer (OLTC) mechanical condition,a feature extraction method was proposed based on improved variational mode decomposition (IVMD) and weight divergence. The harmony search (HS) algorithm was used to optimize the parameter selection of the relevance vector machine (RVM). The mechanical vibration signals of OLTC under different conditions were measured by simulation experiments. The OLTC vibration signals were then decomposed into a series of finite-bandwidth intrinsic mode function (IMF) by IVMD. Next,Kullback–Leibler divergence (K-L divergence) of the IMF and original vibration signal was calculated. The K-L divergence was multiplied by the weight coefficient to obtain the weight divergence,which represented the time-frequency domain complexity of the OLTC mechanical vibration signals. Simultaneously,the multi-classification model of RVM was constructed. The selections of kernel function parameters were optimized by HS,and the classification of weight divergence was realized effectively. The experimental and data analysis results show that the proposed integrated model exhibits high fault diagnosis accuracy. This model can accurately extract the characteristics of mechanical condition,and provide reference for the practical OLTC intelligent fault diagnosis.