Semi-supervised specific emitter identification method based on convnext network

This letter proposes a semi-supervised identification method to address the challenge of ineffective identification of unknown radiation sources in communication emitter individual identification. The method involves multi-feature transformation and feature fusion of radiation source signals, followed by training an optimal closed-set model and feature extractor using a ConvNeXt network combined with an attention mechanism. The K-means clustering algorithm is then employed to classify and identify unknown radiation source signals. The effectiveness of the proposed method is verified through various evaluation metrics and feature visualization. Experimental results demonstrate that the method achieves a classification recognition rate of up to 95% for three types of unknown radiation source signals and exceeds 70% for four types, further confirming its effectiveness in the classification and identification of unknown radiation sources. Semi-supervised specific emitter identification method based on convnext network Dian Lv 1, Zhiyong Yu 1*, Junjie Cao 1,2, and Jiawei Xie 1 1 Rocket Force University of Engineering, Xi’an, 710025, P. R. China 2 College of Information and Communication, National University of Defense Technology, Wuhan, 430035, P. R. China Email: [email protected] This letter proposes a semi-supervised identification method to address the challenge of ineffective identification of unknown radiation sources in communication specific emitter identification. The method involves multi-feature transformation and feature fusion of radiation source signals, followed by training an optimal closed-set model and feature extractor using a ConvNeXt network combined with an attention mechanism. The K-means clustering algorithm is then employed to classify and identify unknown radiation source signals. The effectiveness of the proposed method is verified through various evaluation metrics and feature visualization. Experimental results demonstrate that the method achieves a classification recognition rate of up to 95% for three types of unknown radiation source signals and exceeds 70% for four types, further confirming its effectiveness in the classification and identification of unknown radiation sources.

Specific Emitter Identification (SEI), a cutting-edge technique in wireless communication, enables the discrimination of individual radio-frequency (RF) emitters through the extraction and analysis of their unique device-specific RF fingerprints [1]. This paradigm shift from conventional signal recognition to fine-grained emitter identification has significant implications for spectrum management, secure communications, and electronic warfare. Traditional SEI methods rely on handcrafted feature extraction based on domain expertise, followed by machine learning-based emitter classification. However, this approach inherently depends on prior knowledge and suffers from limitations such as poor generalizability, time-consuming implementation, and low efficiency [2]. Recent advances have seen deep learning (DL) techniques applied to signal recognition, demonstrating remarkable superiority in addressing SEI challenges and significantly enhancing emitter identification accuracy. While DL offers substantial advantages, most studies remain confined to closed-set scenarios, with limited exploration of open-set recognition tasks involving unknown emitter instances. Recent advances have seen deep learning (DL) techniques applied to signal recognition, demonstrating remarkable superiority in addressing SEI challenges and significantly enhancing emitter identification accuracy. While DL offers substantial advantages, most studies remain confined to closed-set scenarios, with limited exploration of open-set recognition tasks involving unknown emitter instances. To address the aforementioned challenges, this work proposes a ConvNeXt-based semi-supervised approach for communication SEI. During preprocessing, energy detection algorithms are employed to intercept critical signal segments containing RF fingerprint information, followed by multiple time-frequency transformations for feature extraction. An attention-enhanced ConvNeXt network is designed to process fused features, with the closed-set recognition network trained on labeled data to retain the optimal model. The test set containing unknown emitters is fed into the trained feature extractor, followed by K-means clustering for open-set classification. Validation is performed on a Bluetooth dataset to demonstrate effectiveness. Emitter signal processing: Bluetooth signals received by the receiver typically comprise both the desired signal and noise components. When valid signals are present or absent in the channel, a distinct energy difference occurs. To extract the valid signal component and reduce data redundancy, an energy detection algorithm is applied to process the captured Bluetooth signals. The processing formula of the energy detection algorithm is as follows: The original signal x(i) is segmented into multiple windows, where E j denotes the total energy detected in the j -th window, with window length S, empirical threshold th, and total window count N . After processing via the energy detection algorithm, the result is illustrated in Figure 1. (a) Original Signal (b) Pre-processed signal Fig. 1 Original vs. Pre-processed signal comparison diagram Signal time-frequency processing: In SEI scenarios, relying solely on single features for emitter identification yields insufficient information, making it challenging to maintain reliable recognition rates in complex electromagnetic environments. To address this, the study proposes applying multi-time-frequency transformations to pre-processed signals, integrating complementary time-frequency features to enhance identification accuracy. The Wigner-Ville Distribution (WVD) analyzes the interaction between time and frequency in signals to reveal instantaneous frequency variations and temporal characteristics, offering advantages such as high time-frequency resolution and clear physical interpretation [3]. The specific formula for WVD is provided below: Where x * (t) denotes the complex conjugate of x(t), and f represents the frequency. The Short-Time Fourier Transform (STFT) divides a signal into multiple time segments and applies an appropriate window function to enable local time-frequency analysis of the signal, offering advantages in analyzing non-stationary signals [4]. The specific formula for calculating the STFT of a signal x(t) is as follows: Where is the time lag variable, ѡ is the angular frequency, and h(t) is the window function. To mitigate spectral leakage, this study employs a Hamming window of length 256 as the window function. The Continuous Wavelet Transform (CWT) computes coefficients at various positions and scales by convolving a signal with a set of wavelet bases, revealing local signal characteristics [5]. It offers the advantage of adaptive resolution adjustment for multi-scale analysis. The specific formula for calculating the CWT of a signal x(t) is as follows: Where is the wavelet function, and a and b represent the scale and translation parameters, respectively. The Hilbert-Huang Transform (HHT) comprises two main steps: Empirical Mode Decomposition (EMD) and Hilbert Spectrum Analysis (HSA) [6]. EMD decomposes a complex signal into several Intrinsic Mode Functions (IMFs), each representing signal characteristics at different time scales. These IMFs can be iteratively combined to approximate the original signal. The specific formula for applying HHT to each extracted IMF IMF(t) is as follows: Where u(t) represents the IMF obtained through EMD. Identification System Design: The ConvNeXt network integrates traditional convolutional neural networks with Transformer architecture, enabling improved feature extraction and enhanced flexibility [7]. This study primarily employs the ConvNeXt-Large model as the recognition framework, as illustrated in Figure 2. Squeeze-and-Excitation (SE) can generate weights to modulate the response of each channel by encoding global information [8]. We consider integrating the SE module into ConvNeXt-Large. The SE module mainly consists of two parts: Squeeze and Excitation. Squeeze is responsible for compressing the feature map from W×H×C to a feature vector Z of 1×1×C through global average pooling. In the Excitation step, a neural network containing two fully connected layers and an activation function α is used to perform nonlinear transformation on the feature vector to generate the weight coefficient matrix Q . The main step calculation formula is shown in Equation 6. Where Z c represents the feature vector of the c -th channel, O c (i,j) denotes the value of the c -th channel at position (i,j) in the feature map, and is the rescaled feature map. The K-means algorithm iteratively partitions data into K clusters, ensuring the minimum total distance between each point and its cluster center [9]. It can be applied to the identification of unknown radiation sources. The value of K represents the number of radiation source categories, which is unknown. By reasonably calculating the K value through the algorithm, we can classify unknown radiation sources. Therefore, we consider integrating the K-means algorithm into a closed-set network. The overall system composition is shown in Figure 3. Fig. 2 ConvNeXt-Large Network Architecture Fig. 3 Overall System Architecture Diagram Verification and Result Analysis: To analyze the effectiveness of closed-set network recognition, the effectiveness of several network recognition models was compared under different signal-to-noise ratios. As can be seen from Figure 4a, the algorithm proposed in this paper outperforms other algorithms within the experimental range of signal-to-noise ratios, especially at low signal-to-noise ratios. From Figure 4b, it can be observed that the closed-set network has completed signal classification. (a) Comparison of Different Models (b) Feature Visualization Fig. 4 Effectiveness Diagram of Closed-set Network (a) K=3 (b) K=4 Fig. 5 Recognition Results of Unknown Emitter Signals (a) K=3 (b) K=4 Fig. 6 Line Graph of Recognition Rate for Different K Values In the experiment, the unknown radiation source categories in the test set were set to 3 and 4, respectively, and the clustering results are shown in Figure 5a and Figure 5b, respectively. By analyzing the clustering results in the figures, it can be seen that the signals are clearly divided into 3 and 4 categories, and the number of unknown radiation source categories is successfully identified. To further verify the classification accuracy, three different evaluation metrics were used to evaluate different K values, namely Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), and clustering accuracy (Accuracy) using the Hungarian Algorithm. The results are shown in Figure 6, which presents the recognition rates corresponding to different evaluation metrics under different clustering K values. Figures 6a and 6b show the recognition rate line graphs for true clusters of 3 and 4 categories, respectively. When the recognition rates of the three evaluation metrics are all the highest, the expected number of clusters coincides with the true number of clusters. The results in the figure verify the accuracy of the clustering algorithm proposed in this paper, achieving correct classification and identification of unknown radiation source data.

This letter mainly adopts a semi-supervised approach to classify and identify unknown radiation source individuals. It performs multi-time-frequency feature fusion on radiation source signals and integrates an attention mechanism into the ConvNeXt_Large network. The fused features are then used as system input for closed-set network training. The effectiveness of the closed-set network is verified by feature visualization and comparing the accuracy of different models. Finally, test set samples containing unknown radiation source signals are used as input to the feature extractor, and the K-means clustering algorithm is employed to complete the identification of unknown radiation sources. The clustering results are verified through various evaluation metrics. Experiments show that this algorithm has certain advantages in the field of unknown radiation source identification. 2025 The Authors. Electronics Letters published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. Received: xx April 2025 Accepted: xx March 2025 doi: 10.1049/ell2.10001

1. Zhang, H., Jiang, Y., Zhao, L., Peng, B.: Specific emitter identification under extremely small sample conditions via chaotic integration, Electron. Lett. 60(14), e13269 (2024) 2. Shi, Y., Yuan, M.-D., Ren, J., Xu, S.: Specific radar emitter identification based on two stage multiple kernel extreme learning machine, Electron. Lett. 57(18), pp. 699-701 (2021) 3. He, Y., Wang, K., Song, Q., Li, H., Zhang, B.: Specific emitter identification algorithm based on time–frequency sequence multimodal feature fusion network, Electronics (Switzerland) 13(18), art. no. 3703 (2024) 4. Wang, B., Xie, J., Wang, F.: Specific emitter identification based on ACGAN and STFT, in: 2024 7th International Conference on Advanced Algorithms and Control Engineering, ICAACE 2024, pp. 400 - 403 (2024) 5. Tang, C., Lian, J., Zheng, L., Gao, R.: Specific emitter identification method for limited samples via time–wavelet spectrum consistency, Sensors 25(3), art. no. 648 (2025) 6. Yu, K., Wang, X., Chen, Z.: Radiation source identification by feature fusion, in: Lecture Notes in Electrical Engineering, vol. 873, pp. 316 - 323 (2023) 7. Liu, Z., Hanzi, Mao, C., Wu, C., et al. : A ConvNet for the 2020s, in: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 11966-11976 (2022) 8. Zhao, Y., Wang, X., Huang, Z.: Multi-scale specific emitter identification via self-attention-based feature pyramid network, in: Lecture Notes in Electrical Engineering, vol. 1193, pp. 177 - 185 (2024) 9. Guo, Z., Huang, Y., Li, Y., Chen, Z., Chen, X.: A soft-decision weighted CNN for specific emitter identification in aeronautical monitoring system, in: ACM International Conference Proceeding Series, pp. 573 - 580 (2022) Information & Authors Information Version history Peer review timeline Published Electronics Letters Version of Record5 Jun 2025Published Copyright This work is licensed under a Non Exclusive No Reuse License. Collection

Authors Metrics & Citations Metrics Article Usage 405views 249downloads Citations Download citation dian lv, Zhiyong Yu, Junjie Cao, et al. Semi-supervised specific emitter identification method based on convnext network. Authorea. 06 April 2025. DOI: https://doi.org/10.22541/au.174396799.94565826/v1 DOI: https://doi.org/10.22541/au.174396799.94565826/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu.