Simulador fenomenológico-aprendizado de máquina de ruído acústico submarino para aplicações em SONAR passivo

Abstract

Underwater SONAR systems play a fundamental role in various civil and military applications. However, the development and validation of signal-processing techniques for such systems face significant challenges, mainly due to the limited availability of experimental acoustic data, especially under different environmental conditions. In this work, we present a Submarine Acoustic Signal simulator (SAS) that integrates phenomenological models of environmental noise (sea state and rain) and radiated noise (cavitation and machinery). To enhance the realism of the synthetic signals, a recurrent neural network of the NARX (Nonlinear Autoregressive Exogenous) type is incorporated, designed to learn and reproduce nonlinear characteristics observed in experimental measurements. The proposed hybrid approach enables the generation of signals that more precisely represent the complex structure of real underwater acoustic environments, especially regarding spectral energy distribution and harmonic content. The applicability and scalability of the SAS are expanded, allowing, from a reference experimental signal, the generation of extended temporal scenarios (simulated signals of the same vessel with longer duration and under different ambient-noise conditions), fleet modeling (generating signals for distinct vessels of the same class by leveraging the NARX network’s extrapolation capability), and operational-dynamics simulation (simulating propeller-speed variations during operation). Validation performed with real data indicates that the simulated signals preserve high similarity to experimental ones, both in broadband characteristics and narrowband components associated with vessel activity. The integration of phenomenological modeling with NARX networks resulted in a hybrid method that significantly improves the correspondence between the phenomenological and experimental signals. The analysis of machinery noise illustrates this improvement: the Mean Absolute Error between amplitudes decreased from 992.06\% (phenomenological model) to 10.33\% (hybrid model), and the Wasserstein Distance was reduced from 30.51 to 2.42. For cavitation, the NARX network achieved a Wasserstein Distance of 0.88~dB, outperforming the phenomenological model’s value of 2.43~dB. Consequently, the proposed method stands as an effective solution to support the design, validation, and evaluation of signal-processing algorithms applied to passive SONAR systems. Its main advantage lies in the ability to create a reproducible and controllable test environment, which drastically reduces the need for field data-collection campaigns, traditionally costly and time-consuming.