滚动轴承作为一种精密的机械元件，已广泛运用于现代工业机械设备中。在轴承运行时，采用合理的方法诊断轴承的故障具有重大的意义。但在实际复杂多变环境下，采集振动信号不仅面临样本量少的问题，还受到噪声干扰、工况变换等因素的影响，导致故障诊断的准确率低。因此，针对噪声干扰和变工况下的小样本滚动轴承故障诊断问题，该文提出了一种基于原型域增强的元学习去噪模型（Meta-DAE）。首先，构造基于时频图的小样本故障样本集，引入深度卷积生成对抗网络并对数据进行预处理，生成相似分布的伪样本集；然后，将故障样本集输入Meta-DAE模型进行自适应特征提取，Meta-DAE模型采用原型域增强策略，使同类别原型点在嵌入空间中凝聚更紧密；同时，构建了具有降噪性能的编码器，设计了基于原型域增强和去噪的目标函数，通过在小样本下进行模型微调，以提高小样本下模型的噪声鲁棒性和分类准确率。噪声及变工况下小样本故障诊断实验结果表明，相比于其他模型，所提模型在-8 dB强噪声干扰下，仅用10个样本微调模型，分类准确率提高了35.78~57.25个百分点，具有较强的噪声鲁棒性。

Rolling bearings, as a type of precision mechanical component, are widely used in modern industrial machinery and equipment. It is of great significance to diagnose bearing faults using reasonable methods during bea-ring operation. However, in the actual complex and ever-changing environment, the collection of vibration signals often faces challenges such as limited sample sizes, noise interference, and operating condition variations, resulting in low fault diagnosis accuracy. To address the problem of small-sample rolling bearing fault diagnosis under noise interference and variable operating conditions, this paper proposed a meta-learning denoising model based on prototype domain enhancement (Meta-DAE). Firstly, a small-sample fault dataset based on time-frequency diagrams was constructed, and a deep convolutional generative adversarial network was introduced for data preprocessing to gene-rate a pseudo-sample set with a similar distribution. Then, the fault sample set was input into Meta-DAE for adaptive feature extraction. Meta-DAE adopts a prototype domain enhancement strategy to make prototype points of the same category more closely clustered in the embedding space. At the same time, an encoder with noise reduction performance was constructed, and a target function based on prototype domain enhancement and denoising was designed. By fine-tuning the model under small-sample conditions, the noise robustness and classification accuracy of the model were improved. Experimental results of small-sample fault diagnosis under noise interference and variable operating conditions show that, compared to other models, the proposed model demonstrates strong noise robustness. Under -8 dB strong noise interference, the model achieves a classification accuracy improvement of 35.78% to 57.25% using only 10 samples for fine-tuning.