The speed variation poses great hardships to the intelligent fault diagnosis of mechanical equipment. Existing solutions rarely consider the interpretable representation of fault information, and still suffer from the “black box” issue that weakens their own practicability and credibility. Motivated by these issues, this study mathematically derives the modal response model of fault-impulsive signals of rotating machinery under variable speed conditions. On this foundation, an interpretable architecture collaborating signal processing with deep learning—prior-knowledge-guided mode filtering network (PKG-MFNet) is proposed, which consists of three sub-structures: mode filtering layer, prior knowledge pooling layer and classifier. In the mode filtering layer, multiple FIR filtering kernels are first constructed by an explainable speed fusion strategy. Each filtering kernel has a center frequency and a bandwidth coefficient fitted from the speed, which is used to extract fault-sensitive modes under varying speed conditions as the explainable feature representations of fault information. Subsequently, the extracted modes are pooled into 12 modal prior indicators (MPIs) that represent health status information in the prior knowledge pooling layer. Finally, the classifier employs two fully-connected layers to make the final decision. Different from the conventional interpretable network for speed information characterization, the speed information is innovatively fused into a novel mechanism framework for sensitive information mining with rigorous theoretical support. Extensive experiments are conducted to verify the superiority of the proposed method over six state-of-the-art methods. Particularly, the visualization analysis not only demonstrates the mode filtering capacity of FIR filtering kernels under variable speed conditions, but also interprets the guiding significance of MPIs for the final decision.

中文翻译：

---

先验知识引导模式过滤网络可解释变速条件下设备智能诊断

速度变化给机械设备的智能故障诊断带来很大困难。现有的解决方案很少考虑故障信息的可解释表示，并且仍然存在“黑匣子”问题，削弱了其自身的实用性和可信度。受这些问题的启发，本研究从数学上推导了变速条件下旋转机械故障脉冲信号的模态响应模型。在此基础上，提出了一种信号处理与深度学习协同的可解释架构——先验知识引导模式过滤网络（PKG-MFNet），该网络由三个子结构组成：模式过滤层、先验知识池层和分类器。在模式滤波层，首先通过可解释的速度融合策略构建多个 FIR 滤波内核。每个滤波核都有一个中心频率和一个根据速度拟合的带宽系数，用于提取不同速度条件下的故障敏感模式作为故障信息的可解释特征表示。随后，提取的模式被汇集为 12 个模式先验指标（MPI），代表先验知识池层中的健康状态信息。最后，分类器使用两个全连接层来做出最终决策。与传统的速度信息表征的可解释网络不同，速度信息被创新性地融合到一种新颖的敏感信息挖掘机制框架中，并具有严格的理论支持。进行了大量的实验来验证所提出的方法相对于六种最先进方法的优越性。特别是，可视化分析不仅展示了FIR滤波核在变速条件下的模式滤波能力，而且解释了MPI对最终决策的指导意义。