Hierarchical-injected noise-enhanced convolutional neural network for rolling bearing fault diagnosis under limited training samples

In the field of rotating machinery fault diagnosis, convolutional neural networks (CNNs) have received lots of attention due to their wide application, and improving the performance of CNNs has extremely profound significance in engineering practice. Reasonable noise injection can effectively enhance the capability of CNN in extracting mechanical fault features. As a result, developing effective noise injection strategies has become an important issue worthy of in-depth investigation and research. Therefore, this article proposes a hierarchical-injected noise-enhanced CNN (HINECNN) model aimed at improving the fault diagnosis capability of CNN, particularly when processing the limited training samples. Specifically, different noise injection modes are designed in the convolutional and activation layers of the model, thereby enhancing the capacity for extracting complex data features and generalization performance. Furthermore, a noise intensity adaptive determination method combining network parameters and training epochs is designed. It can not only accelerate the early training process of the network but also effectively reduce the potential interference of noise on model convergence in the later stage of training. Finally, to evaluate the ability of the HINECNN model to generalize, two fault datasets of rolling bearing are used to construct testing tasks with different sample sizes. The results illustrate that when compared with traditional methods, the suggested model achieves significant improvements in diagnostic accuracy and generalization capability, particularly when processing the limited training samples. This research can offer a reference for exploring the utilization of noise benefits in mechanical fault intelligent diagnosis.