Spectral Neural Network for Specific Emitter Identification

Abstract

The existing ResNet models used in specific emitter identification (SEI) typically use global average pooling (GAP) to reduce feature dimensions. However, this results in a substantial loss of key subtle information. In particular, the recognition performance often fails to meet SEI requirements when unbalanced and weakly labeled samples are present. This study uses the characteristics of radar emitter signals and proposes an approach for SEI based on frequency-domain pooling, fast Fourier transform (FFT) pooling, and wavelet transform pooling. First, a detailed mathematical derivation of FFT pooling and wavelet transform pooling was performed. Next, low-frequency (LF) and high recognition accuracy (HRA) selection criteria were used to select the corresponding retained frequency components. Finally, the new pooling method and frequency-component selection criteria were employed to construct a spectral neural network (SNN) framework for recognizing specific radar emitters, using ResNet as the foundation. Experiments were conducted using a real radar radiation-source dataset, and the results indicated that the proposed algorithm improved the recognition performance by nearly 5%, compared to the GAP-based algorithm, under the same conditions. Moreover, the proposed algorithm exhibited superior recognition performance and stronger robustness than the GAP method under the conditions of sample imbalance and few shot.