Aiming at the problems of poor adaptability and low efficiency of existing transfer learning methods to cross-time domain data, this paper proposes a generation model transfer learning algorithm based on multi-scale feature fusion. The multi-scale depth-separable convolutional network is used to extract the layered fingerprint features in the signal frequency domain, and the channel attention mechanism is combined with the adaptive focusing of the key components of the inherent distortion of the hardware to enhance the discrimination of key fingerprint features. At the same time, a Bi-directional Generative Adversarial Networks (Bi-GAN) framework is used to realize the potential spatial alignment of the features of the source domain and the target domain using bi-directional mapping constraints. The Maximum Mean Discrepancy (MMD) method is used to help align the feature distribution of source domain and target domain. Based on the real acquisition radar data set verification, the recognition accuracy can reach about 90%, and the time complexity is low, which can adapt to the requirements of practical application scenarios.