Deep learning has emerged as a formidable tool in bearing fault diagnosis, yet its effectiveness is often hampered by the opaqueness of feature interpretation and the scarcity of labeled data under varied industrial conditions. In response to these challenges, this paper introduces a mechanism-driven domain adaptation model with interpretability tailored for bearing fault diagnosis across various operating conditions. Specifically, a customized autoencoder driven by bearing fault mechanism is developed for unsupervised extraction of interpretable features from vibration signal, aiming to guide the fault diagnosis model capture the mechanism features of bearing faults. Furthermore, the mechanism features of each sample signal are constructed into a physical parameter matrix composed of natural frequency, damping ratio, amplitude, etc, transforming the input into meaningful physical indicators of bearing condition rather than mere original data points. Ultimately, the physical parameter matrix of all samples is input into the domain adversarial neural network to adaptively diagnose bearing faults under different operating conditions. The feature extraction capability of the mechanism-driven model is verified through various simulation analyses, and three sets of experiments and comparison methods are utilized to verify the effectiveness and advantages of the proposed method in terms of bearing fault diagnosis and interpretability.

中文翻译：

---

迈向增强的可解释性：用于跨工作条件下进行轴承故障诊断的机制驱动的域适应模型


深度学习已成为轴承故障诊断的强大工具，但其有效性经常受到特征解释的不透明性和不同工业条件下标记数据的稀缺性的影响。为了应对这些挑战，本文引入了一种机制驱动的域适应模型，该模型具有可解释性，适用于各种运行条件下的轴承故障诊断。具体来说，开发了一种由轴承故障机构驱动的定制自编码器，用于从振动信号中无监督提取可解释特征，旨在指导故障诊断模型捕获轴承故障的机构特征。此外，将每个样本信号的机理特征构建成由固有频率、阻尼比、振幅等组成的物理参数矩阵，将输入转化为有意义的轴承状态物理指标，而不仅仅是原始数据点。最终，将所有样本的物理参数矩阵输入到域对抗神经网络中，以自适应地诊断不同工况下的轴承故障。通过各种仿真分析验证了机构驱动模型的特征提取能力，并利用三套实验和对比方法验证了所提方法在轴承故障诊断和可解释性方面的有效性和优势。