P91 steel is widely used in thermal power plants and Type IV creep cracking is the main failure model of P91 steel structure. Since the type IV crack propagates rapidly, it is significant to evaluate the creep state of P91 steel before crack forms. MBN is a promising nondestructive testing method. In order to accurately identify the creep state of P91 steel, an IPSO self-adaptive feature selection and SVM model was proposed. P91 specimens with different creep time were prepared and MBN signals were collected. Twelve feature parameters were extracted from MBN signals and used for the improved PSO self-adaptive feature selection and SVM model. The identification rate of the SVM model after feature selection by the improved PSO algorithm is 91.67%, while that by the PSO algorithm is 83.33%, suggesting the IPSO self-adaptive feature selection and SVM model is effective to evaluate the creep state of P91 steel.