MMEVAE: A Time-Varying Rolling Bearing Data Enhancement Method Based on a Multimodal Mechanism Enhanced Variational Autoencoder and Self-Attention Module

Variational autoencoder (VAE) is widely used as a data enhancement technique. However, it faces challenges with inaccurate potential spatial distribution and poor reconstruction quality when dealing with time-varying fault signals. Therefore, we propose a method that combines multimodal mechanism-enhanced VAE with a self-attention module (SAM) to enhance time-varying rolling bearing data. First, we transform the collected vibration signals using a short-time Fourier transform to obtain the time-frequency distribution matrix of the signals. Then, we calculate and normalize the entropy of spectra in each time window to derive the normalized time-frequency entropy spectrum (NTFES), which serves as input for our model. Next, the Gaussian mixture model (GMM) is introduced to redefine the potential space by evaluating and selecting its number of distributions based on Bayesian information criterion (BIC) and mutual information (MI). This approach helps improve both image reconstruction quality and model complexity. Second, SAM is incorporated into our model to effectively capture correlation information between layers and enhance image generation quality. Finally, an actual variable speed-bearing dataset and a public dataset are used to verify the effectiveness of the proposed method.