Learning modulation filter networks for weak signal detection in noise. (January 2021)

• Record Type: Journal Article
• Authors: Zhang, Duona; Ding, Wenrui; Zhang, Baochang; Liu, Chunhui; Han, Jungong; Doermann, David
• Is Part Of: Pattern recognition. Volume 109(2021)
• Publication Date: 2021-01
• Subjects: Weak signal detection -- Filter learning -- Attention -- Modulation classification -- Wireless communication; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107590
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: The contributions of our work include: New modulation filters that are employed to refine signal filters, leading to a new architecture for CNNs model. The convolution operation is further improved by the learned filters. LMFNs are highly compressed, yet achieving state-of-the-art performance. Solving LMFNs in an end-to-end framework with a two-stage optimization scheme. These LMFNs outperform all the state-of-the-art models, and can solve the weak signal detection problem under strong and complex background noise with unknown covariance. Establishing a weak signal dataset that contains UAV communication signals in a real-terrain environment. The dataset is rich in attributes and useful for training networks like LMFNs. Abstract: Weak signal detection is a challenging yet significant problem in the field of radio communication. Although hand-crafted filters are widely used in signal processing, they are challenged by the weak signal detection task with unknown background noise especially in the range of 0-5dB. In this paper, we propose the learning modulation filter networks (LMFNs) to improve the detection performance. The approach is based on a two-stage optimization scheme which addresses filter learning, attention mechanism and classification in a unified framework. Modulation filters are built to enhance the capacity of the learned filters, and the attention mechanism further characterizes the saliency properties of the input signal. LMFNs reduce the storage size of the network while achieving the state-of-the-art performance by a significant margin compared to traditional cognitive radio approaches. We establish a weak signal dataset that contains unmanned aerial vehicle (UAV) communication signals in a real-terrain environment. The source code and dataset will be made publicly available soon.