Digital twin-based automated sonar signal simulation for training datasets

The growing importance of sonar AI in maritime defense and target detection, tracking, and classification has created an urgent need for extensive training datasets. However, data collection in real marine environments remains costly and time-intensive. We present an automated sonar signal simulation based on Digital Twin technology. The system generates scenarios automatically using user-defined parameters, including data volume, environmental conditions, ownship and target maneuvers, target noise, and ambient noise characteristics. Acoustic propagation modeling utilizes oceanographic databases with parameters such as depth, temperature, and salinity profiles to account for the physical characteristics of sonar propagation. Target information databases include vessel characteristics—hull dimensions, screw, and engine parameters—enabling realistic simulation of both broadband and narrowband tonal signatures in targets. This simulation framework facilitates large-scale dataset generation for sonar AI applications. We describe the system architecture and signal generation process, with validation through CNN-based target classification experiments. Two datasets generated by our system achieved classification accuracies of 67.9% and 94.4%, respectively. Through analysis of performance differences based on dataset composition, we demonstrate the system's practical utility for sonar AI development. [Work supported by Korean Research Institute for defense Technology planning and advancement, KRIT-CT-22-023.]