计算机科学与探索

• 学术研究 •    下一篇

伪标签驱动的类内域对齐:跨域开放集故障诊断

江德鸿, 卢佳慧, 李岩   

  1. 1. 哈尔滨理工大学 计算机科学与技术学院,哈尔滨 150080
    2. 哈尔滨工业大学 计算学部,哈尔滨 150006

Pseudo-Label Driven Class-wise Domain Alignment: Cross-Domain Open-Set Fault Diagnosis

JIANG Dehong,  LU Jiahui,  LI Yan   

  1. 1. School of Computer Science and Technology, Harbin University of Science and Technology, Harbin 150080, China
    2. Faculty of Computing, Harbin Institute of Technology, Harbin 150006, China

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,本文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,进一步对齐类内域的特征空间,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,本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

Abstract: At present, domain adaptation fault diagnosis methods are extensively applied and developed. These methods generally assume that the training and testing datasets share an same label set. However, this assumption often proves invalid in practical applications, where unknown fault categories may arise in the testing environment.. To address this challenge, a cross-domain open-set fault diagnosis algorithm based on pseudo-label driven class-wise domain alignment is proposed. In the process of domain adaptation, the pseudo-label mechanism in the target domain is utilized to further align the feature space within the same class. Additionally, a weighted adversarial learning approach is employed to construct decision boundaries for unknown classes. To minimize the impact of noisy pseudo label, the entropy of the pseudo-labels is used to reassign sample weights, so as to construct a more accurate decision boundary between unknown classes and known classes. The experimental results on three bearing datasets demonstrate that the proposed method outperforms mainstream approaches in terms of classification accuracy for both known and unknown classes, thereby validating its effectiveness and advancement.

Key words: Transfer learning, Domain adaption, Open set domain adaption, Fault diagnosis