M4ANet: Underwater Thruster Condition Monitoring Network Based on Information Fusion

To address the safety hazard of Thrusters in Autonomous Underwater Vehicles (AUVs) becoming entangled with abandoned fishing gear while navigating complex marine environments, this paper proposes a multimodal model, M4ANet, based on optoelectronic information fusion. The model processes underwater optical images and motor electrical signal data separately: the optical backbone network employs a lightweight MobilenetV4, while the electrical signal processing utilizes a self-designed Transformer-based feature extraction network. For feature fusion, a multi-scale feature fusion strategy is adopted, achieving cross-modal correlation analysis between high- and low-dimensional image features and electrical signal features through hierarchical concatenation. Experiments demonstrate that this approach achieves a 95.8% accuracy on a self-constructed dataset, significantly outperforming single-modal methods, thus providing a novel solution for AUV Thruster condition monitoring.