StackedFormer: A 1D-transformer model for bearing fault diagnosis

Given the variability and complexity of fault diagnosis in rotating machinery, traditional convolutional architectures exhibit deficiencies in feature representation, leading to suboptimal fault identification rates that fail to fully reflect the operational state of the machinery. To address these issues, this paper introduces a novel one-dimensional Transformer model for fault diagnosis—StackedFormer (SF). Firstly, we refine the computation of self-attention within the traditional Transformer framework. Subsequently, the StackedFormer Block (SFB) is employed to incrementally extract features from the raw vibration signals. Finally, the efficacy of our approach is validated using bearing datasets from the University of Paderborn in Germany and Case Western Reserve University. Experimental results demonstrate that, compared to other methodologies, our novel approach achieves higher fault diagnosis accuracy and superior performance in fault identification.