A Simplified Reference Sensor-Embedded Low-Noise Fiber-Optic Hydrophone for Improving Deep Learning-Based Underwater Acoustic Perception

Underwater environment denoising and target recognition are vital tasks in underwater acoustic perception. Recent advances in deep learning have greatly improved these tasks, but their performance remains limited by the quality of sensor-acquired data. Fiber-optic hydrophones (FOHs) which are widely used in underwater acoustic data collection suffer from inherent phase noise. This noise introduces random perturbations to the captured data and hence reduces the effectiveness of learning-based methods. Embedding a reference sensor in FOH is a promising solution to reduce FOH phase noise, but is complex to be implemented. To address this challenge, this study proposes a simplified method for implementing a reference sensor-embedded FOH. In the proposed architecture, the signal and reference sensors are made all-passive and share a common input and output via time-division multiplexing. To reduce computational complexity in the demodulation, the conventional ellipse fitting algorithm is replaced with a self-orthogonal component superposition method for nonlinear distortion compensation. Experimental results demonstrate that the proposed FOH can suppress phase noise by up to 9.33 dB compared to conventional FOH, and achieves a Total Harmonic Distortion (THD) of -33.71 dB. Using the proposed FOH, deep learning-based environment denoising is improved by 3.73 dB and underwater target recognition accuracy is enhanced by 6.96%. These findings validate the effectiveness of the proposed FOH and highlight the significance of phase noise reduction in enhancing deep learning-based underwater acoustic perception.