Considering a large number of radar interference signals could be built within a short period on the battlefield. Traditional convolutional neural networks face issues due to their large scale. It’s a challenge to deploy jamming recognition system on small-scale equipment. This paper proposes an improved lightweight convolutional neural network to solve the problem by adopting adaptive kernel and batch normalization technology to improve recognition efficiency. By extracting time-frequency character to construct training and testing database for neural network training. Experiment shows that the network achieves over 96% identification accuracy for six kinds of interference signals under -8dB JNR. Compared with other networks， it has a superior accuracy efficiency ratio.