Ship Radiated Noise Recognition Based on Dual Low-Rank Adaptation Training

To address the limited generalization capability of deep learning models in ship-radiated noise recognition, this paper proposes a dual low-rank transfer learning framework that simultaneously optimizes both model weights and feature representations. Specifically, in the weight space, the pretrained weights are frozen, and a lightweight low-rank weight adjustment(WLoRA) module is introduced to construct learnable low-rank increments. This strategy enables efficient fine-tuning with significantly fewer trainable parameters, thereby mitigating the risk of overfitting. In the feature space, considering the inherent low-rank properties of Mel spectrograms derived from ship-radiated noise, a low-rank feature adjustment(FLoRA) module is designed to compress and reconstruct the extracted features. This explicit low-rank constraint encourages the model to learn compact and discriminative representations that better capture the essential structures of ship-radiated noise. By jointly exploiting low-rank optimization in both the weight and feature dimensions, the framework maximizes the potential of pretrained models and improves transfer learning performance. The experimental results on two publicly available underwater acoustic datasets, ShipsEar and Deepship, demonstrate that the proposed method significantly enhances the performance of transfer learning in the classification model of ship-radiated noise compared to direct fine-tuning of pre-trained models. Furthermore, ablation studies validate the effectiveness of the two low-rank modules.