Signal Bandwidth Estimation for Jamming Detection: Skip Connection and LARS Augmented Mask R-CNN

Accurate bandwidth estimation is crucial for achieving effective jamming localization and suppression in the domain of jamming detection. Traditional methods exhibit low accuracy and a limited bandwidth range when confronted with conditions characterized by large bandwidth and low signal-to-noise ratio (SNR) of jamming signals. To tackle these challenges, we propose a signal bandwidth estimation algorithm based on skip connection and layer-wise adaptive rate scaling (LARS) augmented mask region-convolutional neural network (SLA-Mask). The proposed SLA-Mask progressively implements skip connections in the backbone and employs a stochastic gradient descent (SGD) and LARS optimization strategy during training. The skip connection framework adopts a dual-path design: a feature rebuilding module with two-stage discrete wavelet transform (DWT) enhancement and a dynamic attention gated fusion (DAGF) based fusion module. This architecture can achieve accurate jamming signal detection, segmentation, and bandwidth estimation. Experimental results demonstrate that the proposed SLA-Mask achieves accurate bandwidth estimation and effectively mitigates feature degradation and oscillations during the period of large-scale training. When the SNR is as low as -10 dB, the proposed method enhances estimation accuracy by 98.7% compared to the traditional auto-correlation function (ACF) based method.