A lightweight multi-scale attention framework for ship-radiated noise recognition in passive acoustic sensing

Abstract

This study aims to develop a lightweight ship-radiated noise recognition method for resource-constrained passive acoustic sensing applications, with the goal of improving low-frequency analysis and recording (LOFAR) spectrogram classification performance and deployment feasibility. A WT-PA-MobileViT model is proposed by introducing a WTConv wavelet convolution module and a polarity-aware attention mechanism into the MobileViT backbone. LOFAR spectrograms of ship-radiated noise are used as model inputs. WTConv is used to extract multi-scale line-spectrum structures and local frequency-band variations, while polarity-aware attention enhances spectral responses related to class discrimination. Experiments are mainly conducted on the ShipEar dataset, with additional evaluation on the DeepShip dataset. The proposed WT-PA-MobileViT achieves 98.00% accuracy, 98.01% F1-score and a Kappa coefficient of 0.9750 on the ShipEar dataset under clean conditions, with only 0.5935 M parameters. Compared with the MobileViT baseline, the proposed model improves recognition accuracy while reducing the parameter size. On the DeepShip dataset, the accuracy increases from 70.50% to 73.25%, showing that the model retains a degree of adaptability across different data sources. The study is mainly based on LOFAR spectrogram classification using the ShipEar and DeepShip datasets. Although the proposed model shows good performance under clean and moderate-noise conditions, recognition accuracy still decreases under extremely low-SNR conditions. Future work will consider noise-robust feature enhancement, low-SNR spectrogram restoration and cross-dataset adaptation to improve model stability in more complex marine environments. The proposed model has a small parameter size and low computational cost, making it suitable for resource-constrained passive acoustic sensing platforms. It can support ship target recognition in underwater monitoring, maritime surveillance and marine environmental observation systems. The lightweight design also provides deployment potential for edge devices or unmanned underwater platforms where storage and computational resources are limited. Reliable ship-radiated noise recognition can support marine traffic monitoring, underwater environmental protection and maritime safety. By improving the automatic recognition capability of passive acoustic sensing systems, the proposed method may help reduce manual monitoring workload and provide useful information for ocean observation and management. This study proposes a lightweight LOFAR-oriented recognition framework that combines multi-scale wavelet feature extraction and polarity-aware attention, providing a feasible implementation for ship target recognition on passive acoustic sensing platforms.