作者：潘东辉,刘永康,魏艳涛,刘永斌*

出版刊物：Expert Systems with Applications

出版时间：2025年

内容摘要：

Health prognostics based on deep learning (DL) has attracted considerable attention in predictive maintenance of rotating machinery. An effective health indicator (HI) with explicit degradation trend is conducive to the division of health stage and enhance the accuracy of remaining useful life (RUL) prediction. However, the vast majority of existing DL-based RUL prediction works are the end-to-end methods, which fail to provide HI representing the health status of bearings. This paper proposes a novel bearing health prognostic framework that integrates HI construction and RUL prediction tasks. In the stage of HI construction, an adaptive shapelet extraction method is proposed, and embedded features are defined by similarity between raw vibration signals and shapelets. The HI of bearings is obtained through average pooling using embedded features. In the stage of RUL prediction, dynamic graphs are introduced to represent connections between embedded features, and graph structure evolution reflects bearing degradation processes. A spatiotemporal learning model based on graph convolutional networks and bidirectional gated recurrent units is used to extract degradation features. A compensation attention mechanism is proposed to reduce information loss in long-term time series prediction. Finally, traditional network structures are replaced with Bayesian neural networks based on variational inference to map the prediction values more accurately while describing prediction uncertainty. A comparison of PHM 2012 bearing datasets shows that the presented prognostic framework achieves higher prediction precision than other state-of-the-art approaches. Explainability analysis based on graph evolution is provided, which demonstrates the corresponding meaning of changes in graph structure as the degree of degradation deepens.

基于深度学习的健康预测方法在旋转机械预测性维护领域受到广泛关注。具有明确退化趋势的健康指标有助于划分健康阶段并提升剩余使用寿命预测精度。然而现有基于深度学习的RUL预测方法大多为端到端模型，未能提供表征轴承健康状况的健康指标。本文提出一种融合健康指标构建与RUL预测任务的轴承健康预测框架。在健康指标构建阶段，提出自适应形状基元提取方法，通过原始振动信号与形状基元间的相似性定义嵌入特征，并利用嵌入特征的平均池化生成轴承健康指标。在RUL预测阶段，引入动态图表征嵌入特征间的关联关系，通过图结构演化反映轴承退化过程；采用基于图卷积网络与双向门控循环单元的时空学习模型提取退化特征；提出补偿注意力机制以降低长期时序预测中的信息损失；最终基于变分推断的贝叶斯神经网络替代传统网络结构，在提升预测值映射精度的同时描述预测不确定性。通过在PHM 2012轴承数据集上的对比实验表明，该预测框架较现有先进方法具有更高预测精度，并基于图演化过程提供可解释性分析，揭示图结构变化与轴承退化程度加深之间的对应关系。