Hydroacoustic signal denoising focuses on extracting valuable target signals by suppressing underwater background noise. Due to the non-Gaussian nature of noise, traditional Gaussian-based denoising methods struggle in complex underwater environments, particularly at low signal-to-noise ratios. This study employs the variational Bayesian Gaussian mixture model (VBGMM) and radial basis function (RBF) network to design a noise detection and denoising framework. First, the short-time Fourier transform (STFT) extracts features from hydroacoustic signals, using its amplitude spectra as the features. Next, the VBGMM models non-Gaussian underwater noise more accurately to capture its statistical properties. Finally, the RBF network identifies the statistical characteristics of noise and separates detected non-Gaussian noise from the target signal, effectively denoising hydroacoustic signals. Specifically, to address the challenge of optimizing Gaussian mixture model (GMM) parameters, this study applies the expectation maximization (EM) algorithm. This approach derives a noise-sensitive multivariate Gaussian mixture distribution. In order to solve the problem of complex initialization of RBF network parameters, this study uses the mean and standard deviation of the noise detection GMM to initialize the RBF activation function parameters. It transfers the noise detection prior distribution to the denoising network. This results in a noise-sensitive RBF network for denoising. Simulation experiments demonstrate that the proposed algorithm effectively removes noise even at low signal-to-noise ratios.