To address the significant impact of small-sample data on the diagnostic accuracy of models in secondary equipment fault diagnosis, a small-sample fault diagnosis method based on a conditional variational autoencoder (CVAE) and an improved convolutional extreme gradient boosting (ConvXGB) model is proposed. First, fault types and corresponding fault feature information of secondary equipment are systematically analyzed to form a fault information feature set. Second, a CVAE is used to perform data augmentation on specific small-sample datasets, generating a balanced dataset. Principal component analysis and t-SNE algorithm are then used to reduce the dimensionality of the CVAE latent space for visualization. Finally, a self-attention mechanism is introduced by adding a squeeze-and-excitation network module after the activation function layer of the ConvXGB model. This enables re-learning of data features and adaptive allocation of feature weights, thereby completing fault feature extraction and fault type diagnosis. Case studies show that, under unbalanced datasets, the proposed method achieves accuracy of 98.86 % and 97.75 % on the training and test sets, respectively, effectively mitigating the influence of small-sample data and enabling rapid and accurate diagnosis of secondary equipment fault types.