Scholarlink

<p>To alleviate the issues of imbalanced classification samples and low model recognition limits in ship-radiated noise recognition, this paper proposes an improved model called Acoustic Jamba based on the Jamba large model. Firstly, this paper constructs the Acoustic State Space Model module, which considers both local and global dependencies based on the acoustic characteristics of ship radiated noise, to extract and integrate time-frequency features, Next, the Deep Frequency Convolution Module is employed to adaptively select suitable thresholds to eliminate frequency noise, and extracts frequency-space domain features from global and local features to form time-frequency-space three-domain fusion features, Then, a dynamic feature fusion factor is introduced to integrate local and global feature information, forming the Acoustic Mamba module, which allows the model to adapt to changes in weights and data distribution, Then, combining the advantages of weighted balance and Facal Loss, the Weighted Rebalance Facal Loss is designed to learn the bias of the model under imbalanced data, adjusting weight coefficients to alleviate the classification imbalance problem, Finally, a cascading approach is adopted to fully utilize the feature information in ship radiated noise to improve the recognition limits of model. Experimental results show that the Acoustic Jamba model improves classification accuracy by 1.7–12.62% and 1.1–21.51% on the ShipsEar and DeepShip datasets, respectively, compared to state-of-the-art methods, providing a reference for constructing sonar recognition systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AJamba: low upper limit and imbalanced ship-radiated noise recognition with Acoustic Jamba model

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Ming Chen,
  • Feng Wang,
  • Yangze Dong

摘要

To alleviate the issues of imbalanced classification samples and low model recognition limits in ship-radiated noise recognition, this paper proposes an improved model called Acoustic Jamba based on the Jamba large model. Firstly, this paper constructs the Acoustic State Space Model module, which considers both local and global dependencies based on the acoustic characteristics of ship radiated noise, to extract and integrate time-frequency features, Next, the Deep Frequency Convolution Module is employed to adaptively select suitable thresholds to eliminate frequency noise, and extracts frequency-space domain features from global and local features to form time-frequency-space three-domain fusion features, Then, a dynamic feature fusion factor is introduced to integrate local and global feature information, forming the Acoustic Mamba module, which allows the model to adapt to changes in weights and data distribution, Then, combining the advantages of weighted balance and Facal Loss, the Weighted Rebalance Facal Loss is designed to learn the bias of the model under imbalanced data, adjusting weight coefficients to alleviate the classification imbalance problem, Finally, a cascading approach is adopted to fully utilize the feature information in ship radiated noise to improve the recognition limits of model. Experimental results show that the Acoustic Jamba model improves classification accuracy by 1.7–12.62% and 1.1–21.51% on the ShipsEar and DeepShip datasets, respectively, compared to state-of-the-art methods, providing a reference for constructing sonar recognition systems.