| 李嘉恒,常致远,樊薇,陈超,许桢英.融合SSM分词器的轻量高斯注意力轴承RUL预测[J].电子测量与仪器学报,2026,40(1):169-179 |
| 融合SSM分词器的轻量高斯注意力轴承RUL预测 |
| Rolling bearing remaining useful life prediction based on a SSM tokenizer andlinear multi-scale gaussian attention |
| |
| DOI： |
| 中文关键词: 滚动轴承 剩余寿命预测 Transformer网络 高斯分布 自注意力机制 状态空间模型 |
| 英文关键词:rolling bearings RUL prediction transformer Gaussian distribution self-attention SSM |
| 基金项目:煤炭精细勘探与智能开发全国重点实验室开放研究课题(SKLCRSM24KF009)、国家自然科学基金（51905218）、第八届中国科协青年人才托举工程（2022QNRC001）、江苏大学青年英才培育计划、江苏省自然科学基金青年项目（BK20210772）资助 |
|
|
| Author | Institution |
| Li Jiaheng | School of Mechanical Engineering, Jiangsu University, Zhenjiang 212013,China |
| Chang Zhiyuan | School of Mechanical Engineering, Jiangsu University, Zhenjiang 212013,China |
| Fan Wei | School of Mechanical Engineering, Jiangsu University, Zhenjiang 212013,China |
| Chen Chao | School of Mechanical Engineering, Jiangsu University, Zhenjiang 212013,China |
| Xu Zhenying | School of Mechanical Engineering, Jiangsu University, Zhenjiang 212013,China |
|
| 摘要点击次数: 655 |
| 全文下载次数: 507 |
| 中文摘要: |
| 轴承作为旋转机械系统的关键传动部件，其运行状态直接关系整体设备的安全性与运行效率。因此，实现轴承状态实时监测与剩余使用寿命（remaining useful life，RUL）的准确预测，对于预防潜在故障具有重要意义。当前基于深度学习的自注意力机制模型虽在寿命预测中得到广泛应用，但由于其主要依赖特征嵌入与位置编码机制，难以有效捕捉退化过程中的关键微观特征变化。尽管嵌入式高斯掩码能显著提升模型对局部细微退化特征的捕捉能力，但在处理数据时计算复杂度3次方增长，限制了使用效率。针对上述问题，提出融合状态空间模型（state-space model，SSM）与注意力机制的协同预测框架：通过将小波变换和倒谱滤波融入状态空间，构建新型特征分词器替代传统嵌入模块，提升退化特征表征能力；基于门控机制的动态筛选算法实时分析特征参数的单调性演变、趋势波动及抗干扰特性，实现关键退化指标的智能提取；结合振动信号的局部特性和全局退化规律，设计轻量化多尺度注意力模块，有效降低计算负荷并实现寿命映射解码。对比实验基于PRONOSTIA轴承寿命数据集工况1、2及江苏联益友仪器测控技术有限公司000A1-3轴承全寿命试验数据，结果表明所提方法在寿命预测平均绝对误差（MAE）和均方根误差（RMSE）指标上分别实现11.4%、20%、15.4%和15.2%、18.5%、27.4%的提升，消融实验表明关键模块可提升计算效率达55.6%。 |
| 英文摘要: |
| Bearings are critical transmission components in rotating machinery whose operating conditions directly affect equipment safety and efficiency, making real-time monitoring and accurate prediction of remaining useful life (RUL) essential for preventing failures. Although deep learning-based self-attention models are widely used for life prediction, their reliance on feature embeddings and positional encoding hinders the capture of subtle degradation changes. Embedded Gaussian masks improve detection of delicate local degradation features, but their cubic computational complexity with data length limits practical efficiency. To overcome these issues, this study proposes a collaborative framework that integrates state-space model (SSM) with attention mechanisms. By incorporating wavelet transforms and cepstral filtering into the state-space process, the new feature tokenization module replaces traditional embeddings to enhance degradation representation. A gating-based dynamic selection algorithm then analyzes feature evolution, trend fluctuations, and noise resistance in real time to intelligently extract key degradation indicators, while a lightweight multi-scale attention module decodes life mapping by merging local vibration characteristics with global degradation patterns and reducing computational load. Comparative experiments on the PRONOSTIA dataset (conditions 1 and 2) and full-life test data from Jiangsu Lianyy Measurement and Control Technology Co., Ltd. show MAE improvements of 11.4%, 20%, and 15.4% and RMSE enhancements of 15.2%, 18.5%, and 27.4%, with ablation studies confirming up to a 55.6% boost in computational efficiency. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
