Aiming at the weak adaptive capacity, slow diagnosis and different fault modes for the existing network fault diagnosis system, combines the immune principle and agent technology and uses hierarchical multiple step diagnosis concept to construct diagnosis model. Based on the clone selection theory, proposes a new algorithm to complete detector training. The algorithm selects the detector colony and introduces dynamic optimizing parameter, which avoiding premature convergence and generation of local optimal solution, greatly improves the diagnostic performance. The introduction of detector classification concept accelerates the diagnosis process. Comparing to the traditional fault diagnosis methods, proves that the model in the treatment of complex environment network fault has obvious advantages.