Hossein Haghbin

Fault diagnosis in rotating machines is an important topic in condition monitoring and predictive maintenance of industrial equipment. Faults in such machines may lead to unexpected shutdowns, reduced efficiency, and increased maintenance costs. Among different monitoring signals, vibration signals are widely used for machine health assessment because they are highly sensitive to mechanical changes and fault-related patterns. In this research, several convolutional neural network-based models are investigated and compared for fault diagnosis in rotating machines. The main objective is to evaluate the ability of these models to learn effective features from raw vibration signals and to examine their stability under noisy conditions. For this purpose, the baseline CNN model and the GACNN, DSICNN, and DPCCNN models are analyzed from different aspects, including diagnostic accuracy, robustness against noise, and computational complexity. The overall results show that using more optimized architectures can improve fault diagnosis performance and increase model stability under noisy conditions. In addition, the comparison of the models indicates that accuracy alone is not sufficient for designing practical condition monitoring systems. The number of parameters, computational cost, and model size should also be considered. Therefore, selecting an appropriate architecture plays an important role in achieving a balance among accuracy, robustness, and computational efficiency.