Dual-path flow-enhanced noise-aware variational autoencoder for underwater acoustic target signal denoising

Underwater acoustic target (UWAT) signal denoising is challenging due to the complexity of the marine environment. Most current denoising methods suffer from poor low- signal-to-noise ratio (SNR) adaptability, limited generalizability, and high vulnerability to target signal component loss. To address these issues, we propose a dual-path (DP) flow-enhanced noise-aware (NA) variational autoencoder (DFNA-VAE) for UWAT signal denoising. DFNA-VAE employs a time–frequency (T–F) attention-based encoder to learn the T–F feature differences between noise and UWAT signals. Subsequently, the NA probabilistic latent Space of DFNA-VAE models the latent representations with a probabilistic framework, and disentangles the latent representations into signal latent variables and noise latent variables, which are then input into two decoders to generate noise and target signals, respectively. This design enhances the model’s generalization capability and robustness. Meanwhile, a DP inverse autoregressive flow is proposed to enhance the representational capacity of the latent space. Finally, we derivate the variational evidence lower bound of DFNA-VAE to guide the optimization of the model in variational Bayesian inference framework. Experimental results demonstrate that DFNA-VAE can achieve good denoising performance on UWAT signals with low SNRs and exhibits good generalization capability. More significantly, DFNA-VAE can reduce the loss of signal components during denoising, thereby enhancing the interclass differences of UWATs, suitable for the preprocessing of UWAT recognition.