Fault diagnosis methods based on clustering analysis can cluster the fault samples into a certain class according to their similarities without supervision, and thus become one type of effective fault diagnosis strategy. To overcome the problem that traditional clustering methods are susceptible to the initial clustering centers, and thus poor local optima is easily obtained, a MinMax kernel K-means clustering algorithm is introduced. In the proposed method, clusters are assigned weights relative to their variances. And kernel trick is introduced to deal with linear inseparable problem in input space. The proposed method is compared with the traditional K-means and K-means++ in some international standard datasets. The comparison results show its effectiveness and advantage. Then, a fault diagnostic model based on MinMax kernel K-means clustering algorithm is proposed. At last, the fault diagnosis model is applied in fault diagnosis for hydro-turbine generating unit. The results illustrate the effectiveness of the proposed method.