Integrating Convolution and Attention Mechanisms for Weak Structured Signal Enhancement and Detection in Noise-Flooded Multi-Sensor Environments

Weak structured signal enhancement and detection in multi-sensor environments remains challenging due to severe noise interference and the heterogeneity of sensing modalities, which often renders traditional signal processing and conventional deep learning models ineffective. To address these limitations, this study proposes the Convolutional Attentional weak structured signal enhancement and detection Network (CA-WSDN), an end-to-end framework that integrates multi-scale 1D convolution for hierarchical temporal feature extraction with a SE-style cross-channel attention mechanism for adaptive multi-scale feature enhancement across heterogeneous sensor channels. The multi-scale branches capture transient and long-range temporal patterns, while the attention module dynamically emphasizes informative channels and suppresses noise-dominated features, thereby enhancing weak fault-related components. Experiments on simulated medical-equipment monitoring data under ultra-low SNR conditions (−5 dB to 0 dB) demonstrate the model’s robustness and generalization capability. CA-WSDN achieves an SNRI of 8.12 dB at ultra-low SNR experimental setting SNR and 95.1% diagnostic accuracy, outperforming all baseline algorithms. The results indicate that CA-WSDN provides an effective and scalable solution for weak structured signal enhancement and detection in complex noise-flooded multi-sensor systems, offering strong potential for industrial and medical monitoring applications.