Because a lot of jamming signals are often included in the bearing vibration signals, the key to bearing diagnosis is to extract fault features efficiently and classify them. Traditional methods of fault feature extraction often need a variety of index sets to represent different faults. In this paper, a method of bearing fault diagnosis based on Laplacian matrix weighted by Mahalanobis distance and improved k-means clustering is proposed. Firstly, the time domain discrete signal of the bearing is mapped to the graph domain to obtain the graph signal, and the set of characteristic indexes representing different bearing fault states is obtained by using Laplacian matrix in an algebraic form of the graph signal weighted by Mahalanobis distance. Then, the improved k-means clustering idea is applied to evaluate and classify the set of characteristic indexes, to realize the classification and recognition of different bearing fault states in the case of single index. The experimental results show that the method of bearing diagnosis based on Mahalanobis distance weighted by Laplacian matrix and improved K-means clustering can effectively extract and precisely classify the characteristic indexes of different bearing faults. At the same time, the accuracy of this method in single index classification is much higher than that of traditional fault feature extraction methods.