基于神经网络的故障诊断方法因其黑盒特性导致可靠性不足,为增加其可信度,本文提出一种基于物理信息增强的轴承故障诊断方法(Physical-Information Probabilistic Deep Neural Network, PIPDN),该方法根据轴承理论故障的特征频率生成物理标签,通过赋予数据更具物理性的标注,提升模型的准确性并降低其预测不确定性。为验证所提出方法的可靠性,通过比较传统深度卷积神经网络、物理信息神经网络及其他诊断方法的智能诊断对比实验。结果表明：PIPDN能够准确辨别轴承故障特征,具有良好的诊断性能和较强的鲁棒性,为轴承故障诊断提供了可靠的技术路径与理论支撑。

To address the reliability limitation caused by the nature of black box of the AI-driven fault diagnosis model, in this paper, a Physical-Information Probabilistic Deep Neural Network (PIPDN) method for bearing fault diagnosis is proposed. This method generates physical labels according to the characteristic frequency of bearing theoretical faults giving more meaningful physical information, which can improve the diagnostic accuracy and reduce the prediction uncertainty. The reliability of the proposed method in fault diagnosis is validated via comparative experiment with traditional deep convolutional neural network, physical information neural network and other di-agnostic methods. The results show that PIPDN can accurately identify the bearing fault characteristics, and has good diagnostic performance and strong robustness, which provides a reliable technical path and theoretical sup-port for bearing fault diagnosis.

TH133

国家自然科学基金项目（面上项目，重点项目，重大项目）