A structure-smoothing regularized variational autoencoder (SSR-VAE) was proposed based on the variational autoencoder (VAE) in order to address the issue of single-feature reliance, limited capacity to capture complex fatigue evolution process, and unclear delineation of fatigue state in existing muscle fatigue analysis method. Time-domain, frequency-domain and nonlinear entropy feature were incorporated to enhance the model’s ability to capture the dynamic change in surface electromyography (sEMG) signal. The decoupling and temporal continuity of the latent variable space were optimized by introducing weighted KL divergence and trajectory smoothing regularization term into the loss function. A correlation-based selection mechanism was used to extract the main fatigue factor from the latent space, which effectively represented the muscle fatigue state. The experimental results show that SSR-VAE outperforms the VAE model in reconstruction performance, and the extracted main fatigue factor can be used to more clearly distinguish different fatigue stage. SSR-VAE achieved a maximum accuracy of 96.211% in muscle fatigue state classification, significantly outperforming other comparison method.