Abstract: The presence of jamming signals can cause radar signals to overlap in both frequency and time domains, exhibiting nonlinear cyclic characteristics. Different types of jamming possess complex features. It is difficult to find the classification results corresponding to the jamming types for the suppression process, and it is hard to accurately adjust coefficients for Wiener filters, resulting in poor jamming suppression performance. First, empirical mode decomposition method is employed to decompose radar signal into intrinsic mode function (IMF) components with distinct time-frequency features for adaptive decomposition of complex radar signals. Then, taking IMF components as network inputs, a self-attention densely connected convolutional network is constructed to extract complex signal features and realize accurate classification of suppressive jamming, deceptive jamming, multipath jamming and clutter. Finally, the parameters of the Wiener filter are dynamically adjusted based on the classification results. For suppressive jamming, the filter bandwidth is expanded and frequency-domain weights are optimized; for deceptive jamming, an adaptive notch filter is designed; for multipath jamming and clutter, suppression is realized by combining a multipath channel model with space-time filtering. The experimental results show that this method can significantly improve the radar anti-jamming capability and has favorable practicability.