Abstract: Aiming at the core issues of sparse effective samples and poor interpretability of detection results in intelligent sonar systems, an embodied cognition modeling approaches driven by physical model is presented to reveal the mechanism by which the hull-mounted sonar can exploit its "body" to enhance perception, and a comprehensive framework for closed-loop self-learning underwater intelligent detection is introduced to offer new principles and methods for sonar design. Through the validation of experimental data, the proposed method demonstrates significant advantages in source detection, azimuth estimation and localization compared to the traditional methods. Embodied cognition model requires less underwater acoustic data and enhances source detection capability significantly, which lays a foundation for addressing the challenges faced by current artificial intelligence methods in sonar applications and enabling broad application in hull-mounted sonar.