Abstract: Currently, domain adaptation-based fault diagnosis methods have been widely applied, typically assuming that training and testing data share the same label space. However, this assumption often fails in real-world scenarios where unknown fault types may appear. To address this challenge, a cross-domain open set fault diagnosis algorithm based on pseudo-label-driven intra-class domain alignment is proposed. In the process of domain adaptation, the algorithm utilizes the pseudo-labeling mechanism of the target domain to effectively reduce the distribution differences between samples of the same class in the source and target domains within the feature space by constructing an intra-class domain alignment strategy. This enhances the model’s ability to discriminate between known classes and builds the decision boundary for unknown classes through extending the weighted adversarial learning of the classifier. In order to reduce the impact of pseudo-label decision-making errors, the method redistributes the sample weight through the entropy of pseudo-labels, so as to distinguish unknown and known classes more accurately. Experimental results on three bearing datasets show that the proposed method is better than the mainstream methods in terms of accuracy of both known and unknown classes, which fully verifies its effectiveness and advancedness.

Key words: transfer learning, domain adaptation, open set domain adaptation, fault diagnosis

摘要： 目前，域适应故障诊断方法具有广泛的应用和发展。这些方法通常假设训练数据与测试数据共享相同的标签集。然而在实际应用中该假设往往不成立，且测试环境中可能会出现未知故障类别。为了解决这一挑战，提出了一种基于伪标签驱动的类内域对齐的跨域开放集故障诊断算法。该算法在域适应过程中利用目标域的伪标签机制，通过构建类内域对齐策略，有效缩小源域与目标域同类样本在特征空间中的分布差异，提升模型对已知类的判别能力，并通过扩展分类器的加权对抗学习来构建未知类的决策边界。为减少伪标签决策错误的影响，该方法通过伪标签的熵重新分配样本权重，从而更加准确地区分未知类与已知类。在三个轴承数据集上的实验结果表明，该方法在已知类类别和未知类类别准确率上均优于主流方法，充分验证了其有效性和先进性。

关键词: 迁移学习, 域适应, 开放集域适应, 故障诊断