A Position-Free Signal Transformer via Multiband Inner Relationship Extraction for Understanding Information Flow of Machinery Diagnosis

Intelligent fault diagnosis methods have shown obvious advantages in health management of mechanical equipment and have been widely studied by scholars. Furthermore, diagnosis under various nonideal conditions has become a hot spot in current research. However, the existing methods for monitoring signal learning are not comprehensive enough, and the intelligent diagnosis model still has the potential to be unexploited. In this article, a signal transformer (SiT) is proposed to extract the internal correlation information of monitoring signals for machinery fault diagnosis. We divide monitoring signals into signal patches containing different frequency bands and add a class token to input them into the model. The model mainly includes a linear projection layer, a self-attention encoder, a class attention encoder, and a classifier. In addition, we cancel the position encoding in transformer. The method is verified based on datasets from two different experiments and achieved accuracy comparable to that of the literature methods. Furthermore, we visualize the self-attention matrix extracted by the model for each class, and the results could help people understand the information flow of diagnosis in model, which has good practical potential. Finally, the key parameters of the proposed model are discussed in detail.