Abstract
In active sonar detection systems, the self-radiated noise from the platform itself is a primary factor that interferes with system performance. The platform self-noise exhibits non-Gaussian characteristics, comprising line spectrum and continuous spectrum, where line spectrum exhibits higher intensity and greater frequency stability. To reduce the ship noise, this study introduces a deep-learning based time-domain model with an encoder-separator-decoder architecture. In detail, the encoder and decoder modules utilize learnable convolutional layers to generate distinguishable features with a symmetrical structure. For separator module, we observe that noise components prominently predominate within the noisy signal in low signal-to-noise ratio (SNR) scenarios. The extraction of noise components is typically more tractable than of signal components. Therefore, in contrast to conventional frameworks that focus on extracting the signal subspace, this module is dedicated to extracting the noise subspace. Besides, the modified scale invariant SNR (SI-SNR) is proposed as the loss function for model optimization, which yields better performance than non-noise extraction structure. The robustness of the model is validated on the DeepShip dataset with four ship categories, demonstrating that the model consistently outperforms baselines, including matched filter, in all ship categories.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have