To solve the problem of low recognition accuracy of radar new active jamming in strong noise environment， a KPCA-SAE-BP network algorithm is proposed. The features of jamming signal in time domain， frequency domain， waveform domain， wavelet domain and bispectral domain are extracted to construct 67-dimensional input space， and the high-dimensional data is nonlinear dimensionality reduction and reconstruction through kernel principal component analysis （KPCA）. A SAE-BP neural network is then used for classification and recognition. Simulation experiments show that in strong noise environment where JNR is greater than -1dB， the recognition accuracy of the KRPA-SAE-BP network algorithm for six new active Jamming is more than 90%， and the training and recognition time is less than 0.7 seconds. Compared with classical BP neural network， SAE-BP network， PCA-BP network and GA-BP network， it has better detection and recognition performance.