Abstract: In the actual combat environment, complex electromagnetic environments can result in radars that do not have access to the prior information needed to meet performance, and radars are no longer limited to a single mission and operating mode. To achieve multiple performance improvements of radar in electronic warfare, a waveform design method based on attention mechanism (AM) and convolutional neural network-bidirectional long short term memory (CNN-BiLSTM) is proposed in this paper. Firstly, the environmental information is used to build a data set via mutual information (MI) criterion and signal-to-interference-plus-noise ratio (SINR) criterion. Secondly, a CNN-BiLSTM-AM neural network model is established, and the data set is utilized for training and test comparison, so as to achieve the purpose of using the network to generate waveforms. Finally, simulation experiments verify that the proposed method can effectively balance multi-criteria performance, with an average radar comprehensive performance improvement of 1.14 % compared to the MI criterion and 2.97 % compared to the SINR criterion.