Fault Diagnosis of Rotating Machinery Based on the Improved Multidimensional Normalization ResNet.

Data-driven-based fault diagnosis methods play an increasingly important role in rotating machinery. Among these methods, deep learning is widely concerned due to its strong nonlinear feature learning ability, wherein ResNet is a very powerful model. However, the diagnostic performance of the model depends on sufficient labeled data samples, which is extremely difficult to achieve under actual complex working conditions. In this article, a multidimensional normalized ResNet is proposed for fault diagnosis of cross-working conditions under limited label samples. First, the collected vibration data under different conditions are preprocessed by computed order tracking to reduce the distribution differences. Second, batch normalization and group normalization are fused together to enhance the feature extraction ability of ResNet. Moreover, the rectified linear unit (ReLU) is replaced by the Gaussian error linear unit to improve the robustness of the trained model. Then, the model is trained by source domain samples, and the obtained parameters will be transferred to the target domain. Finally, the transferred model is fine-tuned through limited samples in the target domain and applied for cross conditions fault diagnosis. Two kinds of datasets are analyzed by the proposed method and existing models to demonstrate their superiority.