An unsupervised few-shot fine-grained fault diagnosis method based on contrastive variational autoencoder

Deep learning-based fault diagnosis methods typically require abundant labeled data and often focus on coarse-grained fault categories, limiting their applicability to fine-grained diagnosis under data-scarce conditions. This paper proposes an Unsupervised Few-Shot Fine-Grained Contrastive Variational Autoencoder (UFF-COVAE) framework for fault diagnosis. The model adopts a task-specific Gaussian mixture prior and constructs a set-level variational posterior to characterize latent class distributions. A momentum encoder and a negative-sample queue are introduced to construct InfoNCE-based contrastive regularization, which increases the computable lower bound on the mutual information between input samples and latent variables, thereby reducing latent information degradation and mitigating posterior collapse. During unsupervised meta-training, an expectation-maximization (EM) algorithm iteratively estimates the mixture weights, means, and covariance matrices of the task-specific prior, while the encoder and decoder learn transferable fault representations using a joint objective comprising reconstruction, Kullback–Leibler (KL) divergence, and contrastive losses. During meta-testing, semi-supervised EM exploits labeled support samples and unlabeled query samples to estimate task-specific class distributions and establish the correspondence between Gaussian components and fault labels. Experimental results on three datasets demonstrate the superior performance of the proposed method in fine-grained fault diagnosis. UFF-COVAE effectively improves diagnostic accuracy under unsupervised few-shot conditions and achieves superior overall performance across the evaluated settings, demonstrating promising potential for engineering applications.