实际工程中需要大量数据支撑的常规诊断方法难以有效进行小样本条件下的离心泵故障诊断，为此，将深度学习中的残差网络（residual network, ResNet）与膨胀卷积相结合，并拓展为孪生网络，构建膨胀残差孪生网络（dilated residual siamese network, DRSN）。将膨胀残差网络作为孪生网络的特征提取模块，强化了模型的特征提取能力；构造正负样本对，从每个样本中提取更多的信息，更有效地利用有限的数据；两个子网络共享参数，减少自由参数的数量，降低样本不足时过拟合的风险。提出的网络模型缓解了训练样本不足的问题，提升了数据利用的效率，实现了小样本条件下的离心泵故障分类。研究结果表明：在样本最匮乏的情况下，该模型在离心泵试验数据集上的准确率仍能达到82.20%，相较其他模型，准确率至少提升了8.8个百分点。

Conventional diagnostic methods that require a large amount of data support in practical engineering are difficult to effectively perform centrifugal pump fault diagnosis under small sample conditions. Therefore, the residual network (ResNet) in deep learning was combined with dilated convolution and extended into a siamese network to construct a dilated residual siamese network(DRSN). The dilated residual network was used as the feature extraction module of the siamese network, which enhanced the feature extraction ability of the model. Positive and negative sample pairs were constructed to extract more information from each sample, and make more effective use of limited data. The two sub-networks share parameters, the number of free parameters and lowering the risk of overfitting was reduced when the sample was insufficient. The proposed network model alleviated the problem of insufficient training samples, improved the efficiency of data utilization, and realized the fault classification of centrifugal pump under the condition of small samples. The research results show that even in the most sample-scarce situation, the accuracy of the model on the centrifugal pump test dataset can still reach 82.20%, which is at least 8.8 percentage points higher than other models.