Abstract:

It is vital for the detection of critical and important signals to suppress background signals at the end of receiver effectively. Current methods inadequately utilize signal context information, rendering it challenging to address the issue of background signal suppression of single channel in noisy environments and time-frequency overlaps. To regard this, a dual-path feature fusion mask-based suppression (DPFF-MS) algorithm is proposed, utilizing mid-frequency time-domain signals. The algorithm utilizes a neural network to fit the mask suppression model, effectively suppressing background signals. The high-dimensional transformation and inverse transformation of signals can be enabled by the employment of a suite of convolutional encoder-decoder networks, diminishing the adverse effects of noise. A dual-path feature fusion (DPFF) module is developed, which leverages long short-term memory (LSTM) networks of varying paths to alternately extract both local features and global contextual information. An iterative attention feature fusion (iAFF) is used to optimize the process of fusing features of different scales, fully exploiting the intra-pulse and inter-pulse information to address the suppression issue in time-frequency overlapping environments. The experimental results indicate that, in comparison to other signal processing approaches and neural network models, the proposed model shows significant enhancements in terms of scale-invariant source-to-noise ratio (SI-SNR) and background pulse suppression rate. Furthermore, it significantly reduces the number of model parameters, making it straightforward to deploy and possess high application value.

Key words: electronic reconnaissance, suppression of background signal, deep learning, multi-scale feature fusion