In order to select the most predictive features from network sample data for fault diagnosis, a novel Adaptive Immune Clonal Selection Algorithm (AICSA) is proposed. By simulating the mechanisms of biological immune system such as immune memory, clone selection and self-adaptation, AICSA achieves the dynamic control of evolution process, which realizes global optimal computing combined with the local searching. Compared with traditional evolutionary algorithms, the proposed algorithm not only has higher convergence rate and better searching ability, but also can avoid prematurity and degeneration phenomenon. The experimental results show that feature selection for machine learning is necessary, and AICSA can efficiently reduce the number of features while improving the performance of network fault diagnosis based on SVM.