It’s difficult to extract the rich fault information from vibration signal by the traditional feature extraction methods of time domain, frequency domain and time frequency domain parameter. In order to solve this problem, a multiscale permutation entropy is proposed to extract fault features and combine the improved multiclass relevant vector machine to fault diagnosis. Since the kernel parameters of the multiclass relevant vector machine do not have the adaptive ability, it has the great influence on the accuracy of fault diagnosis. The multiclass relevance vector machine is improved by a grasshopper optimization algorithm to realize the adaptive fault diagnosis. The experimental data from the University of Western Reserve in the United States show that the proposed optimized fault diagnosis model can realize the fault diagnosis of different types and the identification of different fault degrees. Compared with the fault diagnosis model of particle swarm optimization optimizes multiclass relevant vector machine, the accuracy of proposed fault diagnosis model is 100%.