A novel context-aware deep learning framework for ship radiated noise recognition

Existing methods for Ship Radiated Noise (SRN) classification primarily focus on feature extraction and model design based solely on acoustic data, which limits further performance improvements. To overcome this limitation, we propose CA-SRNNet, a context-aware deep learning framework that integrates both temporal and semantic information to enhance SRN classification. Specifically, a time-aware varying-coefficient regression module is introduced to dynamically incorporate temporal context, allowing the model to adapt feature representations over time. Additionally, ship category labels are leveraged to guide representation learning via a semantic-aware contrastive loss. We employ Bayes error bound analysis to demonstrate the framework’s superior feature extraction capability and use SHAP to quantify the contribution of the temporal module. Extensive experiments across multiple backbone networks consistently show improved classification performance, confirming the effectiveness and generalizability of the proposed approach.