Data-driven fault diagnosis for multiple subsystems of deep-sea submersible

The deep-sea submersible requires a reliable fault diagnosis system for safe operation. However, many current fault diagnosis methods focus on single subsystems, lacking the capability to diagnose concurrent failures across multiple systems. We propose a multi-subsystem fault diagnosis method with two key modules: a feature extraction module using 2D convolutions with 1D kernels to preserve sensor-specific temporal patterns, and an attention fusion module incorporating a squeeze-and-excitation mechanism to dynamically weight sensor importance. This architecture enables independent processing of multi-sensor data while maintaining computational efficiency and achieves multi-label fault classification across the electrical, hydraulic, propulsion, and control systems. In experiments conducted on a real-world operational dataset from a deep-sea submersible, our proposed method achieved the highest Macro F 1-Score of 0.920 and demonstrated the fastest convergence among all compared methods.