基于对称式对比学习的齿轮箱无监督故障诊断方法

针对跨工况下无监故障诊断特征提取难、模型泛化性弱的问题，提出一种基于对称式对比学习策略的齿轮箱无监督故障诊断方法。首先，利用原始信号构建正负样本集，通过加噪声、序列倒转等数据增强后，分别输入两个结构相同的卷积神经网络提取高维特征；其次，度量正负样本的相似程度进行编码学习数据的隐藏表示，通过对称式自监督对比学习优化正负样本的对比估计损失函数，从而有效利用样本自身标签信息，提升网络从无标签样本中学习判别特征的能力；最后，在齿轮箱数据集上对所提方法开展试验验证，通过聚类准确率、分类系数和划分熵进行综合评估。结果表明，所提方法聚类精度可达98%以上，相比其他方法，呈现了更强的聚类能力和泛化性能。

Unsupervised fault diagnosis of gearbox based on symmetrical contrast learning

Unsupervised intelligent fault diagnosis under different operating conditions is still a challenge task. To obtain high-quality samples and strong model generalization performance, an unsupervised intelligent diagnosis method based on the symmetrical contrast learning framework is proposed for gearbox fault diagnosis. Firstly, a positive sample set and a negative sample set are constructed and enhanced from original signals by adding noise and sequence inversion, which can be fed into two convolutional neural networks (CNN) with the same structure to extract high-dimensional features. Then, a novel symmetrical contrast learning method is proposed to obtain the underline encoding information by measuring the degree of similarity between positive and negative samples. Further, the noisecontrastive estimation loss function is optimized through symmetrical self-supervised learning strategy. In this way, the label information of the sample itself could be effectively used, and the discriminative performance of extracted features from unlabeled samples is improved. Finally, the proposed method is tested and verified on the gearbox data set. Three indicators including clustering accuracy, classification coefficient and partition entropy are constructed for comprehensive evaluation. Experimental results show that the proposed method achieves at least 98% clustering accuracy, which has stronger cluster and generalization ability than other diagnosis approaches.