【摘要】 迁移学习在故障诊断应用中面临数据隐私保护与标签数据匮乏的双重难题，为此提出融合对比学习的无源领域自适应迁移学习故障诊断算法。通过源域数据完成模型训练，使所得源域模型具备提取故障信号高维特征的核心能力。采用无源领域自适应方法，在完全不调用源域样本数据的前提下，仅依托预训练源域模型与目标域样本构建无监督损失函数体系，结合数据增强后的目标样本开展对比学习，并引入类别混淆最小化机制，显著提升目标样本故障诊断的准确率与结果可信度。构建多尺度注意力科尔莫戈罗夫-阿诺德网络（KAN）卷积神经网络结构，强化模型对故障信号中非线性特征的挖掘与提取能力。基于两个航空高速轴承数据集的验证结果表明，所提故障诊断算法具备优异的鲁棒性，可有效适配数据隐私受限场景下的航空轴承故障诊断需求。更多还原

【Abstract】 Dual challenges of data privacy protection and scarcity of labeled data are faced by transfer learning in fault diagnosis applications, and a contrastive learning-based source-free domain adaptation transfer learning algorithm for fault diagnosis is proposed to address these issues. A source domain model is trained using source domain data, endowing it with the core capability to extract high-dimensional features of fault signals. The source-free domain adaptation method is adopted, and a system of unsupervised loss functions is constructed relying solely on the pre-trained source domain model and target domain samples without invoking any source domain sample data. Contrastive learning is conducted with augmented target samples, and a category confusion minimization mechanism is introduced to significantly improve the diagnostic accuracy and result reliability of target sample fault diagnosis. A multi-scale attention-based KolmogorovArnold network(KAN) convolutional neural network structure is constructed to enhance the model’s capability to mine and extract nonlinear features in fault signals. Validation results based on two high-speed aero-engine bearing datasets demonstrate that the fault diagnosis algorithm exhibits excellent robustness and can effectively adapt to the aero-engine bearing fault diagnosis requirements in scenarios with restricted data privacy.更多还原