The fault of one of the key systems in artificial satellites, the Control Moment Gyroscope (CMG), can lead to significant economic losses and irreparable consequences. Therefore, it is crucial to diagnose its faults promptly. Traditional fault diagnosis methods, however, face challenges such as local feature traps and difficulty in feature extraction when dealing with CMG vibration signals, making it hard to meet the requirements for accuracy and robustness. Hence, it is essential to design a high-accuracy model to assess the health status of CMG on time. To address these issues, a fault diagnosis method that combines the Joint Attention Mechanism (JAM) with one-dimensional dilated convolutional networks and residual connections is proposed. The method efficiently learns feature information through the JAM, effectively addressing the time-varying characteristics of vibration signals and focusing more on fault-related features. The influence of rotational speed on the model is overcome to some extent through JAM. The three rotational speeds are mixed as datasets, and the model achieves high accuracy. The proposed method significantly enhances the accuracy and robustness of CMG fault diagnosis. Experimental results on a self-collected dataset demonstrate that the proposed method achieves excellent accuracy (98.14%) and robustness in CMG fault diagnosis.