A CNN-Based Fault Diagnosis Method for AUV Thrusters Integrating Multidomain Data Sequence Reconstruction and Data Channel Extension

Autonomous underwater vehicles (AUVs) are increasingly used in marine exploration and underwater operations, where unexpected thruster faults may hinder mission execution and compromise operational safety. Therefore, efficient and reliable fault diagnosis methods are essential for improving the safety and autonomy of AUV operations. Data-driven fault diagnosis approaches provide an efficient way for detecting thruster faults in autonomous underwater vehicles (AUVs). A typical data-driven method that directly supplies the original AUV dynamic signal sequences into a CNN classifier is adopted as the baseline in this study and is referred to as OriCNN. To further improve the representation of fault-related information in the OriCNN, whose temporal location may vary among finite signal windows, and to enrich the available input channels of the OriCNN, this study proposes a CNN-based fault diagnosis framework integrating multidomain data sequence reconstruction (MDDSR) and data channel extension (DCE). MDDSR constructs alternative representations in the time, frequency, and time–frequency domains to reduce the sensitivity of CNN inputs to variations in the temporal location of fault-related features. DCE employs multiple signal processing techniques to generate enhanced data sequences, treats each sequence as an independent data channel, and optimizes channel arrangements to reduce CNN convergence iteration. Experiments using a prototype AUV in a lake were conducted, in which the thruster faults were software-simulated thrust-loss conditions generated by modifying the thruster control-voltage command. Experimental results show that the proposed MDDSR + CNN effectively mitigated the unstable training behavior observed in OriCNN and the proposed MDDSR + DCE + CNN further reduced the convergence iterations while maintaining robustness against different convolution kernel sizes. These results indicate that the proposed MDDSR + CNN and MDDSR + DCE + CNN frameworks contribute to AUV thruster fault diagnosis and thereby to safer and more autonomous underwater operations.