基于类感知对比学习的半监督故障诊断

Fault Diagnosis Based on Class-Aware Contrastive Semi-Supervised Learning

当前基于数据驱动的故障诊断方法依赖于标注完备的训练样本,然而在实际工程活动中标注足量故障样本需要耗费大量人力物力。对此提出一种基于类感知对比学习的半监督故障诊断方法以综合应用少量标注样本以及大量无标注样本进行训练,减少模型训练对于标注样本的需求。首先根据模型最大概率预测值动态赋予无标注样本伪标签以参与模型训练,并结合置信度筛选以减少伪标签中噪声标签所引起的负面影响,同时引入一致性正则化,增强模型对伪标签样本的特征表达能力,构建更为完备的决策边界。随后设计类感知对比学习模块以确保模型特征空间中的各类故障样本间的类内一致性以及类间对比度,实现判别能力增强。实验结果证实,该方法能够在较少标签的条件下取得良好的诊断结果。

The current data-driven fault diagnosis methods rely on well-labeled training dataset, however those datasets in practical engineering activities require lots of resources. In this way, a semi su-pervised fault diagnosis method based on class-aware contrastive learning is proposed to compre-hensively apply a small number of labeled samples and a large number of unlabeled samples for training, reducing the need for labeled samples in model training. Firstly, based on the maximum probability prediction value of the model, unlabeled samples are dynamically assigned pseudo la-bels to participate in model training, and confidence screening is combined to reduce the negative impact caused by noisy labels in the pseudo labels. At the same time, consistency regularization is introduced to enhance the model’s feature expression ability for pseudo label samples and con-struct a more complete decision boundary. Subsequently, a class-aware contrastive learning mod-ule is designed to ensure intra class consistency and inter class contrast among various fault sam-ples in the feature space, achieving enhanced discriminative ability. The experimental results con-firm that this method can maintain great fault diagnosis performance with few labels.