<p>This study proposes a novel neural network architecture—Attention Refinement Versatile ResNet (ARV-ResNet)—which incorporates three hierarchical attention mechanisms, namely frequency attention, self-attention, and channel attention, into the ResNet residual framework to enhance feature representation and fault diagnosis capabilities. To evaluate the effectiveness of ARV-ResNet in detecting faults in rotating machinery, experiments were conducted using both a self-collected dataset from Feng Chia University (FCU) and a publicly available dataset from Southeast University (SEU). Discriminative features were extracted through multi-scale signal decomposition methods, including Variational Mode Decomposition (VMD), Wavelet Packet Decomposition (WPD), and Empirical Mode Decomposition (EMD). These features were subsequently used to train and validate ARV-ResNet, along with four benchmark neural networks—ResNet, AlexNet, GoogLeNet, and ShuffleNet—to compare their classification performance. Additionally, transfer learning was employed to assess the robustness and generalization ability of the models across datasets. Experimental results demonstrate that ARV-ResNet achieves superior classification accuracy and exhibits strong performance in transfer learning scenarios.</p>