Array-Based Faster R-CNN for Underwater Acoustic Communication Signal Detection

The detection and recognition of underwater acoustic communication (UWAC) signals hold significant importance in both civil and military fields. Employing data-driven neural network object detection algorithms facilitates the estimation of temporal spectral locations and modulation types on spectrogram of UWAC signals. However, due to the significant propagation loss and severe multipath effects inherent in UWAC channels, the detection efficiency is compromised, resulting in a limited monitoring area. To mitigate these challenges, we propose leveraging multi-channel spectrograms of sensor array as input to Faster R-CNN, called array-based Faster R-CNN. This method significantly enhances the detection performance, especially under low signal-to-noise ratio (SNR) conditions. Simulation experiments indicate that utilizing a multi-channel array signal with 8 sensors as input lowers the detection threshold by 4 dB. At an SNR of -16dB, the proposed method achieves a detection F1 score of 0.94 and a modulation recognition accuracy of 0.74, significantly outperforming the single-channel method which yields a detection F1 score of 0.58 and a modulation recognition accuracy of 0.48.