ABSTRACT

Bearings are widely applied in rotating machinery of nuclear power plants (NPPs). Data-driven fault diagnosis technology is critical to ensuring the reliable operation of rotating machinery. Aiming at the problem of poor model generalization ability caused by discrepant data distribution of monitoring signals under various working conditions, a deep transfer learning method based on fully categorized alignment subdomain adaptation (FCA-SAN) is proposed in this paper. Firstly, the bearing vibration signals of the source and target operating conditions are preprocessed and converted into time-frequency domain images suitable for model input. Subsequently, a pre-trained deep convolutional neural network (DCNN) model is adopted as the feature extractor, which is combined with FCA-SAN to extract transferable features across different working conditions. The subdomain adaptation method reduces the data distribution discrepancy more fine-grained by aligning the feature distribution of different working conditions, thereby effectively improving the model generalization ability. Finally, the experimental results show that, compared with the traditional method, the proposed subdomain adaptation method reaches the highest fault diagnosis accuracy in different transfer tasks, which demonstrates the potential application value in rotating machinery of NPPs.

摘要

轴承广泛应用于核电站（NPP）的旋转机械中。数据驱动的故障诊断技术对于确保旋转机械的可靠运行至关重要。针对各种工况下监测信号数据分布不一致导致模型泛化能力差的问题，提出一种基于全分类对齐子域自适应（FCA-SAN）的深度迁移学习方法。首先，对源和目标工况的轴承振动信号进行预处理，转换为适合模型输入的时频域图像。随后，采用预训练的深度卷积神经网络 (DCNN) 模型作为特征提取器，结合 FCA-SAN 提取不同工作条件下的可转移特征。子域自适应方法通过对齐不同工况的特征分布，更细粒度地减少数据分布差异，从而有效提高模型泛化能力。最后，实验结果表明，与传统方法相比，所提出的子域自适应方法在不同的传输任务中达到了最高的故障诊断准确率，证明了在核电厂旋转机械中的潜在应用价值。