轻量化多层感知网络的海洋声纹识别

Marine Acoustic Fingerprint Recognition Using Lightweight Multilayer Perceptron Network

摘要:

为提升海洋声学指纹信号的分类效率与抗噪声干扰能力，同时降低传统深度学习模型的计算成本以适应资源受限的水下设备，本文研究了一种轻量化的海洋声纹识别方法。所提出的方法融合维纳滤波与轻量化多层感知网络，构建了端到端的识别流程。首先提取声学数据的梅尔频率倒谱系数特征，进而通过自适应噪声评估机制对高噪声样本进行滤波降噪，最后利用轻量化分类网络实现高性能信号识别。本文实验验证基于真实海洋生物声学数据的多样性物种数据集，实验结果表明，该方法在整体分类准确率上表现稳定，对不同物种及不同行为状态的声音信号均具有良好的识别效果。同时，本文所提出的方法显著减少了信号处理时间，并在低算力设备上实现了较高的鲁棒性识别能力，为复杂海洋环境下的边缘监测设备提供了可行的技术解决方案。

Abstract:

To enhance the classification efficiency and noise resistance of marine acoustic fingerprint signals while reducing the computational cost of traditional deep learning models for resource-constrained underwater devices, this paper investigates a lightweight method for marine acoustic recognition. The proposed approach integrates Wiener filtering with a lightweight multilayer perceptron network, constructing an end-to-end recognition pipeline. The process begins with extracting Mel-frequency cepstral coefficient features from the acoustic data. An adaptive noise assessment mechanism is then employed to filter and denoise samples with high noise levels, while a lightweight classification network is utilized to achieve high-performance signal recognition. Experiments are demonstrated with a diverse dataset comprising real marine bio-acoustic data from multiple species. The results demonstrate that the proposed method achieves stable overall classification accuracy and exhibits strong recognition performance across different species and their varying behavioral vocalizations. Furthermore, the approach significantly reduces signal processing time and delivers highly robust recognition capabilities on low-computational-power devices. This provides a feasible technical solution for edge monitoring equipment operating in complex marine environments.