To efficiently mine the classification model for machine fault diagnosis based on images, a hybrid classification algorithm, which inspired by combining nonnegative matrix factorization and artificial immune system, was put forward. In the algorithm, nonnegative matrix factorization was employed for dimensionality reduction of the time-frequency spectral images. An artificial immune based classification model was constructed by means of training of data samples mapped into low-dimensional space to recognize the machine conditions and diagnose faults. Experimental results on the fault classification of diesel valve train demonstrate the effectiveness of the algorithm. Compared with probabilistic neural network classifiers, the hybrid classifier achieves better fault diagnosis performance.