Underwater thruster failures pose a critical risk to the mission reliability of unmanned surface vehicles (USVs), especially during long-term autonomous operations. Conventional fault diagnosis approaches often rely on vibration sensors or controlled laboratory settings, which limits their practicality in dynamic marine environments. To overcome these challenges, a noncontact acoustic fault diagnosis framework that leverages microphone-based sensing combined with deep-learning techniques is presented in this paper. Acoustic signals acquired from underwater thrusters are converted into time frequency spectrograms using short time Fourier transform (STFT), and background noise is suppressed through a spectral-aware denoising convolutional neural network (SA-DnCNN). The denoised spectrograms are then used to train a vision Transformer (ViT-B/16) for the classification of four representative thruster conditions: normal, breakage, entanglement, and biofouling. The proposed SA-DnCNN + ViT model was validated using acoustic data obtained both in a controlled water tank environment and during real-world sea trials. Classification accuracies of 99.27% and 96.69% were achieved. Comparative experiments with CNN, ResNet50, and other denoising-based models confirmed the superiority of the proposed method in terms of robustness and precision. Notably, this study is one of the first experimental demonstrations of microphone-based diagnosis of thruster faults under actual marine conditions and highlights the potential of the method as a scalable and practical solution for real-time health monitoring of USVs. Future work will focus on expanding the acoustic dataset through long-term marine deployments and incorporating multimodal sensor fusion to improve diagnostic generalization and resilience across diverse maritime environments and operational scenarios.