The core objective of Specific Emitter Identification (SEI) technology is to achieve unique identification of different devices by extracting the inherent hardware fingerprint features contained in the emitted signals of the emitter. At present, SEI technology has been widely applied in the military field for friend-or-foe identification and electronic reconnaissance, as well as in the civilian field for wireless network security defense and illegal equipment detection. In recent years, with the breakthroughs in artificial intelligence technology, the SEI method based on deep learning has shown significant advantages and has become the mainstream research direction of SEI issues at present. However, due to problems such as equipment aging and changes in the working environment of the radiation source, its fingerprint features will also change slightly over time. The models trained by traditional deep learning algorithms have the practical deficiency that the recognition accuracy drops sharply over time. Therefore, it is urgent to propose a feasible solution for the fingerprint drift problem of SEI across time domains in actual scenarios.

This paper proposed a transfer learning algorithm for radiation source identification based on multi-scale feature fusion and bidirectional generative adversarial Network (Bi-GAN). Through multi-level feature extraction and dynamic fusion mechanisms, the recognition performance of the model in complex electromagnetic environments was significantly improved. In the feature extraction stage, the algorithm adopted a multi-scale convolutional neural network architecture to capture the local detail features and global context information of the radiation source signal respectively, and realized cross-scale feature fusion through the feature pyramid structure, effectively enhancing the discriminability of the features and the robustness against noise interference. In the design of the network architecture, the bidirectional Generative Adversarial Network (Bi-GAN) framework was introduced, which contained a bidirectional mapping structure of the forward generator (from the source domain to the target domain) and the reverse generator (from the target domain to the source domain). Through adversarial training, the feature distributions of the two domains were forced to align to the same latent space. This bidirectional constraint mechanism not only ensured that the generated samples retain both the category information of the source domain and have the feature distribution of the target domain, but also significantly inhibited the generation of meaningless samples. To further optimize feature selection, the algorithm embodied an attention mechanism in the feature fusion layer and highlighted key radiation features through adaptive weight allocation, enabling the model to focus on the most discriminative signal components.

1. After being processed by the method in this paper, the recognition accuracies when the corresponding numbers of emitters is 3, 4, 5, and 6 can reach approximately 93%, 89%, 87%, and 86% respectively, and the maximum improvement in accuracy can exceed 40%.

2. After being processed by the method in this paper, based on the verification of the actual radar dataset used in this paper, the time complexity of the algorithm in this paper is reduced by more than 20% compared with the existing algorithms.

Aiming at the problems of poor adaptability of existing transfer learning methods to cross-temporal domain data and low algorithm operation efficiency, this paper proposes a generative model transfer learning algorithm based on multi-scale feature fusion. The hierarchical fingerprint features in the signal frequency domain are extracted through a multi-scale depth-separable convolutional network. Combined with the channel attention mechanism, the key components of the inherent distortion of the hardware are adaptively focused to enhance the discriminability of the key fingerprint features. Meanwhile, the potential spatial alignment of the features of the source domain and the target domain is achieved by using the bidirectional generative adversarial network framework. And the maximum mean difference method is adopted to assist in aligning the feature distributions of the source domain and the target domain. Based on experimental verification, it can achieve a recognition accuracy of about 90%, and the time complexity is reduced by more than 20% compared with the existing algorithms, which can adapt to the requirements of actual application scenarios.