TSFANet: Temporal-Spatial Feature Aggregation Network for GNSS Jamming Recognition

• Published in: IEEE transactions on instrumentation and measurement Vol. 73; p. 1
• Main Authors: Zhong, Wanfu, Xiong, Hailiang, Hua, Yuan, Shah, Danyal Hussain, Liao, Zhiwei, Xu, Yudan
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; attention mechanism; Chirp; Convolution; convolutional neural network; Datasets; Feature extraction; Global navigation satellite system; GNSS jamming recognition; Jamming; Machine learning; Modules; Receivers; recurrent neural network; Spectrogram; Spectrograms; time-frequency analysis
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2024.3375975

Malicious jamming attacks have a destructive effect on GNSS receivers and even make them lose lock. Jamming recognition can raise an alarm for the users and assist the receivers in selecting an appropriate mitigation algorithm. However, traditional techniques can only detect the presence of jamming signals and suffer problems when jamming classification is required. The performance of typical machine learning techniques is subject to the quality of handcrafted features and the local receptive field of convolutional operations. In this paper, a multi-feature GNSS jamming (MFGJ) dataset containing the clean GNSS signal and 11 types of jamming signals is first built. Then, a high-performance and lightweight GNSS jamming recognition model named TSFANet is proposed, which is implemented at the precorrelation stage and fed with spectrograms. With the help of multi-feature labels, TSFANet can not only detect the presence of jamming signals, but also accurately recognize jamming types. TSFANet is composed of the prompt feature aggregation module (PFAM), temporal feature aggregation branch (TFAB), spatial feature aggregation branch (SFAB), and feature fusion module (FFM). TFAB and SFAB aim to aggregate temporal and spatial features, individually. FFM utilizes the squeeze-and-excitation block to fuse features extracted from two branches for further improvements. The experimental results indicate that TSFANet outperforms conventional CNNs and transformer-based networks in terms of recognition performance and memory consumption. TSFANet achieves 97.76% accuracy on the public dataset and 95.39% accuracy on the MFGJ dataset with only 1.007M trainable parameters. Its recognition speed is at a medium level. Moreover, TSFANet shows excellent performance in recognizing low-power jamming signals.