An application in bearing fault diagnosis

Convolutional neural networks (CNNs) have been extensively applied to bearing fault diagnosis, delivering many satisfactory results. However, the features extracted by CNN layers are often abstract, making it challenging to interpret the extracted features in each convolutional layer from physics. To address this issue, this study proposes a layered-training-based CNN. First, the hierarchical labeling is designed to construct three different classification labels based on multirefinement diagnostic tasks, such as normal-fault classification, fault position recognition, and fault size identification. Second, the CNN layers are trained based on hierarchical labels and frozen sequentially layer-by-layer, so that each convolutional layer captures different features for different classification labels. Experimental results demonstrate that the layered-training-based CNN produces clearer and more interpretable feature representations and achieves higher accuracy and stability than the baseline CNN.