Valid roller fault diagnose plays an important role in improving working efficiency and intelligent plant. In view of complex industrial environment and numerous noise types, first, a difference method is used to eliminate the influence of time trend in audio sequence data, and to extract the characteristics of the roller audio sequence. Secondly, K-Means and spectral clustering algorithms are used to have cluster analysis and roller faults identification. In order to evaluate advantages and disadvantages of proposed clustering model, an average ratio of sub-sequence same labels from an audio sequence is proposed to achieve the above aim. Experimental results show that local diagnostic accuracy can be improved by dynamic selection of parameter values. Roller fault can be effectively identified and diagnosed by two proposed clustering algorithms, but spectral clustering algorithm is superior to K-Means algorithm. By use of the proposed methods, one can see that the efficiency of coal preparation is improved, number of unplanned outages is reduced, and good economic benefits are also produced.