摘  要

在工业物联网与智能制造快速发展的背景下，工业设备运行中源源不断地产生海量数据，这为机械故障监测奠定了数据基础。但在复杂工况下，面对振动、温度等多模态数据时，传统方法在泛化能力和计算效率上暴露出明显不足，很难同时保证诊断的准确性与实时性。

针对这一问题，本文构建了一种基于双模自适应消息传递网络的机械故障监测模型。该模型综合利用振动和温度双模态信号中的时空特征，能够对设备状态进行实时感知，并对异常状况做出诊断。在模型设计上，先搭建了双模自适应网络结构，利用双模自适应消息传递机制对多源数据进行动态融合，使不同模态的信息相互补充；同时采用自适应分箱量化方式处理连续特征，降低了模型对数据分布的敏感性，提升了适应性与稳定性。在此基础上，进一步设计了双模自适应消息传递分类器，并以风力涡轮机齿轮箱的实测数据开展实验，验证其在故障分类任务上的效果。

实验结果方面，在较大数据规模下模型精度约为91%，且计算效率明显优于传统的决策树和支持向量机方法。不过，后续对比分析也发现，双模自适应机制在识别未知故障类型时仍存在局限，说明要继续提升性能，还需要融入更丰富的领域知识来优化模型结构。总体来看，本研究为工业大数据场景下的智能故障诊断提供了一种新思路，模块化的设计使其在多模态融合任务中具有较好的扩展性，具备较高的工程应用价值，有望为工业设备稳定运行和高效维护提供支撑。

关键词：机械故障监测；双模自适应机制；双模自适应消息传递网络；工业大数据；智能诊断

Abstract

The rapid advancement of the Industrial Internet of Things and smart manufacturing has led to an ever-growing stream of data from operating industrial equipment. This data offers a rich foundation for mechanical fault monitoring. Yet conventional methods often fall short when dealing with multimodal data under complex working conditions. Their limited generalization ability and low computational efficiency compromise both detection accuracy and real-time responsiveness.

To address these limitations, this paper introduces a mechanical fault monitoring model based on a dual-mode adaptive message-passing network. The model fuses spatiotemporal features extracted from dual-mode signals—specifically vibration and temperature—to enable real-time perception of equipment status and accurate anomaly diagnosis.

We first designed a dual-mode adaptive network architecture. Within this architecture, a message-passing mechanism dynamically integrates multi-source data, allowing different modalities to interact and complement each other. Continuous features are handled through adaptive binning quantization, which reduces the model's sensitivity to data distribution and markedly improves its adaptability and stability. Building on this, a dual-mode adaptive message-passing classifier was developed. Its fault classification performance was tested on real measurement data from wind turbine gearboxes, and the results confirm its effectiveness.

Experimental findings show that with sufficiently large datasets, the model achieves an accuracy of around 91% and is considerably more computationally efficient than traditional decision tree and support vector machine approaches. Nevertheless, further comparative analysis points to a remaining weakness: the dual-mode adaptive mechanism struggles with previously unseen fault types. This suggests that integrating deeper domain knowledge will be necessary to further improve model performance.

Overall, this study offers a new pathway for intelligent fault diagnosis in industrial big data settings. Thanks to its modular design, the model is highly extensible and adapts well to multimodal fusion scenarios, holding strong practical value for reliable equipment operation and efficient maintenance.

Keywords: Mechanical Fault Monitoring; Dual-Mode Adaptive Mechanism; Message Passing Network; Industrial Big Data; Intelligent Diagnosis