An Improved ViT-LSTM Hybrid Model for Bearing Fault Diagnosis

To address the issue that signal features in rolling bearing fault diagnosis rely on manual extraction and selection, which easily affects the fault classification accuracy, an improved ViT-LSTM model is proposed herein for bearing fault diagnosis. The ViT is utilized to extract features from vibration signals, while the LSTM is adopted to capture time-dimensional dependencies. The ProbSparseAttention mechanism is introduced into the original ViT model, which effectively reduces the calculation amount, supports the model to process longer vibration sequences. Meanwhile, variational mode decomposition is introduced to decompose the vibration signal into multiple components, and then convert them into two-dimensional or multi-dimensional data to adapt to the input of ViT. Experiments show that the fault classification accuracy of the proposed algorithm reaches 98.51%. Moreover, it achieves superior results in multi-classification tasks, paving a new way for the automated and intelligent development of rolling bearing fault diagnosis.