1 Shijiazhuang Campus, Army Engineering University of PLA, Shijiazhuang 050003, China
2 School of Economics and Management, Hebei University of Science and Technology, Shijiazhuang 050018, China
Show less
文章历史+
Received
Published
Revised
2026-02-10
2026-05-18
2026-04-21
Online Date
2026-08-06
PDF (7970K)
摘要
在装备维保工程实践中,监测数据常存在高缺失率且工况覆盖不全,在导致退化轨迹关键阶段信息丢失的同时,也因训练数据工况单一而削弱了剩余寿命(remaining useful life, RUL)预测的跨工况泛化能力,从而制约预测可靠性。为解决上述问题,提出一种退化扩散插补(degradation diffusion imputation, DDI)方法,用于轴承性能退化数据的概率重建。该方法采用“观测掩码约束”与“采样过程观测替换”的双条件引导机制,确保逆向去噪过程严格遵循已知观测点;同时引入频域一致性损失,使生成序列的频谱幅值与真实序列相近,从而有效保留退化信号的单调趋势与关键频带特征。在XJTU-SY轴承数据集上,针对10%~90%的随机缺失与块缺失场景,DDI的补全精度(以平均绝对误差(mean absolute error, MAE)、均方根误差(root mean squared error, rMSE)为指标)整体优于线性插值、K近邻(K-nearest neighbors, KNN)、双向循环插补网络(bidirectional recurrent imputation for time series, BRITS)、自注意力插补网络(self-attention-based imputation for time series, SAITS)以及当前最先进(state-of-the-art, SOTA)的条件评分扩散插补(conditional score-based diffusion imputation, CSDI)方法。消融实验进一步验证了上述双重条件机制的有效性。此外,利用DDI生成的数据对训练集进行增强后,多种下游剩余寿命预测模型如随机森林(random forest, RF)、梯度提升树(gradient boosting, GB)、多层感知机(multilayer perceptron, MLP)的RMSE平均降低约20%。综上,研究成果为不完整监测数据补全提供了一种可量化不确定性的数据层解决方案。
Abstract
In equipment maintenance engineering practice, monitoring data often suffer from high missing rates and incomplete operating-condition coverage, which lead to loss of critical degradation phase information and, due to the narrow range of training conditions, weaken the cross-condition generalization capability of RUL (remaining useful life) prediction, thereby compromising the reliability of performance degradation prognostics. To address these issues, DDI (degradation diffusion imputation) method for probabilistic reconstruction of bearing performance degradation data was proposed. A dual conditioning mechanism consisting of “observation mask constraint” and “observation replacement during sampling” was employed to ensure that the reverse denoising process strictly follows the known observations. Meanwhile, a frequency-domain consistency loss was introduced to enforce spectral amplitude similarity between the generated and real sequences, thereby effectively preserving the monotonic degradation trend and key frequency-band features. On the XJTU-SY bearing dataset, under missing rates from 10% to 90% for both random missing and block missing scenarios, the imputation accuracy measured by MAE(mean absolute error) and RMSE(root mean squared error) of DDI was generally superior to that of linear interpolation, KNN(K-nearest neighbors), BRITS(bidirectional recurrent imputation for time series), SAITS(self-attention-based imputation for time series), and the current SOTA(state-of-the-art), CSDI(conditional score-based diffusion imputation) method. Ablation studies further validate the effectiveness of the above dual conditioning mechanism. Moreover, after augmenting the training set with data generated by DDI, the RMSE of multiple downstream RUL prediction models e.g., RF(random forest), GB(gradient boosting), MLP(multilayer perceptron) was reduced by about 20% on average. In summary, this results provide a data-layer solution with quantifiable uncertainty for incomplete monitoring data imputation.
LIU Qing, DI Yan-qiang, CUI Hao-hao, FENG Shao-chong, WANG Tao.
Probabilistic Reconstruction of Bearing Performance Degradation Data Using Conditional Diffusion Models[J].
Science Technology and Engineering, 2026, 26(14): 5978-5991 DOI:10.12404/j.issn.1671-1815.2600873
滚动轴承作为旋转机械的关键承载部件,其性能退化程度直接关联设备故障风险,进而影响其运行安全与可靠性。通过振动传感器实时监测轴承状态所获取的振动信号,构成了反映其健康状态的间接性能退化数据,蕴含从早期故障萌生至最终失效的全生命周期信息,是实施预测与健康管理(prognostics and health management, PHM)决策的核心依据[1-3]。这类数据具有鲜明的时序特点:承载设备的明确退化趋势[4];受个体差异、复杂工况与随机扰动影响,其演化过程呈现非线性的趋势、时变尺度及非高斯(如重尾分布)的随机波动等不确定特性[5]。在实际工程场景中,尽管每个数据点都承载着关键退化信息,具有重要价值(全寿命数据获取难度高、成本大[6]),实际采集中却常出现缺失,因传感器故障、传输中断或采样不连续,数据普遍存在缺失[7-8],研究表明其平均缺失率可达10%,极端情况下甚至超过30%[9],呈现出高度的非完备性与稀疏性;尤其当缺失由离散点状演化为连续段缺失时,关键退化信息会更易遭到破坏[10]。关键退化信息的损失,严重制约高精度剩余寿命(remaining useful life,RUL)预测模型的构建与发展[11-12],对后续故障诊断与寿命预测性能产生根本性影响。因此,系统性地处理这种不完整时序监测数据、实现高精度补全,已成为该领域的关键挑战与前沿课题[1,13]。
针对GAN的不足,近年来,扩散模型作为一种基于分数匹配原理的新型生成范式,在理论基础上避免了像GAN一样陷入对抗博弈的复杂动态平衡,而是把复杂生成任务简化为多个“降噪”问题,从而避免了模式崩溃[17],又凭借其训练过程稳定、生成样本多样性好等优势,在图像、语音等领域迅速崛起,如今正逐步向时序数据领域拓展[18]。近年已有综述系统梳理了不完整时序数据补全的深度学习方法演进过程,并进行了问题分类[1],同时亦有最新综述对工业时序插补领域的技术现状与挑战进行了集中归纳[19]。在工程应用层面,面向工业过程监测数据的扩散式插补[20]、面向卫星遥测等复杂缺失模式的条件扩散补全[21],以及面向RUL估计的扩散概率模型数据增强[22]等研究,为缺失条件下的时序重建与任务协同提供了可参照路径。特别是条件扩散模型,能够有效利用观测值作为先验信息,引导生成过程;其中面向通用时序的概率插补框架(如条件评分扩散建模范式[23])与上述工业、遥测及RUL相关实践[20-22]相互补充,整体展现出成为当前性能最优(state of the art,SOTA)方法的潜力。此外,这一“基于条件引导的生成思想”在处理不完整信息、实现高保真复原方面亦得到跨领域讨论和应用[24]。
(3)基线方法(选取5类代表性方法对比)。①传统插值:线性插值(linear interpolation, LI)、K最近邻插值(K-nearest neighbors, KNN);②深度学习-循环模型:双向循环插补网络(bidirectional recurrent imputation for time series,BRITS)[33];③深度学习-注意力模型:自注意力插补网络(self-attention-based imputation for time series,SAITS)[34];④生成模型-SOTA:条件评分扩散插补模型(conditional score-based diffusion imputation,CSDI)[23],作为主对照。
针对轴承性能退化数据缺失及工况覆盖有限制约预测与健康管理(prognostics and health management, PHM)可靠性的问题[2],在不完整时序数据利用与补全研究日益受到关注的背景下[1],所提出的退化扩散插补方法,实现概率重建(含数据补全与可控生成)。采用“观测掩码约束+采样过程观测替换”的双条件引导,使逆向去噪过程严格服从已知观测;训练阶段引入频域一致性损失Lfreq,以约束幅频结构,并配合趋势-周期建模保留退化轨迹中的趋势与关键频带信息。推理阶段可通过多次独立采样给出95%置信区间(confidence interval, CI),实现不确定性量化(uncertainty quantification, UQ)。通过系统性实验与分析,在XJTU-SY公开轴承加速寿命数据[30]上,在统一协议下与条件评分扩散CSDI[23]及线性插值(LI)、K近邻(KNN)、BRITS[33]、SAITS[34]等基线完成主对比,主要研究结果如下。
ZhouLi, WuYangyang, MiaoXiaoye. A review of deep learning-based imputation methods for incomplete time series data[J]. Journal of Computer Research and Development, 2025, 62(9): 2233-2258.
[3]
NguyenK T, MedjaherK, TranD T. A review of artificial intelligence methods for engineering prognostics and health management with implementation guidelines[J]. Artificial Intelligence Review, 2023, 56(4): 3659-3709.
[4]
WangD, TsuiK L, MiaoQ. Prognostics and health management: a review of vibration based bearing and gear health indicators[J]. IEEE Access, 2017, 6: 665-676.
[5]
LeiY, YangB, JiangX, et al. Applications of machine learning to machine fault diagnosis: a review and roadmap[J]. Mechanical Systems and Signal Processing, 2020, 138: 106587.
[6]
BianL, GebraeelN. Stochastic modeling and real-time prognostics for multi-component systems with degradation rate interactions[J]. IIE Transactions, 2014, 46(5): 470-482.
[7]
LiuJ, LuoP, ShenJ. Bearing fault prognosis method based on priori knowledge-enhanced particle filter[J]. Mathematical Problems in Engineering, 2022(1): 8634753.
LiXiaorui, BanXiaojuan, YuanZhaolin, et al. Time series prediction methods and applications based on deep learning in industrial scenarios[J]. Journal of Engineering Science, 2022, 44(4): 757-766.
XuXinyao, XuDe. Missing data imputation for industrial process monitoring based on GRU network[C]//Proceedings of the 2020 Chinese Automation Congress (CAC2020). Beijing: Chinese Association of Automation, 2020: 82-87.
[12]
HuangH, EvansS. FlowFill: a fast and energy-efficient streaming framework for recovering missing values in industrial sensor data[C]//Proceedings of the 2025 IEEE 41st International Conference on Data Engineering. New York: IEEE, 2025: 4262-4268.
[13]
ZhuJ, ChenN, ShenC. A new data-driven transferable remaining useful life prediction approach for bearing under different working conditions[J]. Mechanical Systems and Signal Processing, 2020, 139: 106602.
[14]
DengJ, WenG, LeiZ, et al. A digital twin of life-cycle bearing based on a defect-guided conditional diffusion model for predicting remaining life under limited degradation data[J]. ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part A: Civil Engineering, 2026, 12(1): 04025114.
[15]
EscobarC A, Morales-MenendezR. Machine learning in manufacturing: quality 4.0 and the zero defects vision[M]. Amsterdam: Elsevier, 2024.
[16]
ChenX, WangH, YuH, et al. A review on machinery health prognostics: from physics-based to data-driven approaches[J]. IEEE Transactions on Industrial Informatics, 2019, 15(11): 6024-6035.
LiZhinong, LiShuyang, LiuBao, et al. Application of infinite hidden Markov model in bearing degradation trend prediction with missing data[J]. Journal of Vibration Engineering, 2023, 36(2): 574-581.
LiuJingtao, QiuMing, LiJunxing, et al. SGAIN fused with TCN for predicting remaining useful life of rolling bearing under missing data[J]. Journal of Ordnance Equipment Engineering, 2024, 45(8): 240-247.
[21]
ArjovskyM, ChintalaS, BottouL. Wasserstein generative adversarial networks[C]//Proceedings of the 34th International Conference on Machine Learning. Sydney: PMLR, 2017: 214-223.
[22]
SongY, Sohl-DicksteinJ, KingmaD P, et al. Score-based generative modeling through stochastic differential equations[C]//Proceedings of the 9th International Conference on Learning Representations. Virtual: OpenReview.net, 2021. DOI:10.48550/arXiv.2011.13456.
[23]
HoJ, JainA, AbbeelP. Denoising diffusion probabilistic models[J]. Advances in Neural Information Processing Systems, 2020, 33: 6840-6851.
[24]
YuanY, ZhangX, WangJ, et al. Industrial time series data imputation: a survey on methods, applications, and challenges[J]. IEEE Transactions on Industrial Informatics, 2023, 19(1): 1078-1090.
[25]
XuH, LiuZ, WangH, et al. Denoising diffusion straightforward models for energy conversion monitoring data imputation[J]. IEEE Transactions on Industrial Informatics, 2024, 20(10): 11987-11997.
PangZhaochen, LiuMing, ZhangLixian, et al. Satellite telemetry data imputation using conditional diffusion models[J]. Acta Automatica Sinica, 2025, 51(10): 2302-2312.
[28]
WangW, SongH, SiS, et al. Data augmentation based on diffusion probabilistic model for remaining useful life estimation of aero-engines[J]. Reliability Engineering & System Safety, 2024, 252: 110394.
[29]
TashiroY, SongJ, SongY, et al. CSDI: conditional score-based diffusion models for probabilistic time series imputation[J]. Advances in Neural Information Processing Systems, 2021, 34: 24804-24816.
WenWen, LiYacong, LiYang, et al. Digital restoration of murals guided by line drawings and damaged images[J]. Science Technology and Engineering, 2025, 25(36): 15622-15632.
[32]
WangC, HuangC, ZhangL, et al. Denoising diffusion implicit model combined with TransNet for rolling bearing fault diagnosis under imbalanced data[J]. Sensors, 2024, 24(24): 8009.
SunJiajia, LiChengli, ChangDexian, et al. Data augmentation method based on generative adversarial network for class imbalance in intrusion detection[J]. Science Technology and Engineering, 2022, 22(18): 7965-7971.
NieQingke, SunGuang, HaoYongpan, et al. Methods and application of multi-source heterogeneous monitoring data fusion[J]. Science Technology and Engineering, 2022, 22(13): 5348-5357.
[37]
WangS, YanL, DuS, et al. Bearing prognostics and health management based on hybrid physical mechanism and data models: a systematic review[J]. Measurement Science and Technology, 2025, 36(5): 052002.
[38]
LiM, QinJ, FanH, et al. SiMBA-augmented physics-informed neural networks for industrial remaining useful life prediction[J]. Machines, 2025, 13(6): 452.
[39]
WangB, LeiY, LiN, et al. A hybrid prognostics approach for estimating remaining useful life of rolling element bearings[J]. IEEE Transactions on Reliability, 2018, 69(1): 401-412.
[40]
LittleR J A, RubinD B. Statistical analysis with missing data[M]. 3rd ed. Hoboken: John Wiley & Sons, 2019.
[41]
SongJ, MengC, ErmonS. Denoising diffusion implicit models[C]//Proceedings of the International Conference on Learning Representations. Virtual: OpenReview.net,2021. DOI:10.48550/arXiv.2010.02502.
[42]
CaoW, WangD, ZhouJ, et al. BRITS: bidirectional recurrent imputation for time series[C]//Advances in Neural Information Processing Systems 31. Montreal: Curran Associates, 2018: 6776-6786.
[43]
DuW, CoteD, LiuY. SAITS: self-attention-based imputation for time series[J]. IEEE Transactions on Knowledge and Data Engineering, 2023, 35(10): 10143-10157.
[44]
SaxenaA, GoebelK, SimonD, et al. Damage propagation modeling for aircraft engine run-to-failure simulation[C]//Proceedings of the 2008 International Conference on Prognostics and Health Management. Denver: IEEE, 2008: 1-9.