面向可迁移跨域自适应学习的航空发动机剩余寿命预测方法

Transferable cross⁃domain adaptive learning for aircraft engine remaining useful life prediction

航空发动机剩余寿命(RUL)预测任务中数据集标签较少且工况多变,导致传感器时间序列之间存在明显分布差异,限制了RUL预测方法的泛化能力。跨域学习的提出为该任务提供了一种可行的解决方案。传统跨域学习通过域自适应方法最小化源域和目标域特征之间的分布差异,得到跨域对齐特征,实现跨域知识迁移。但随着航空发动机的退化,前后时间步的语义信息也发生变化,导致原先对齐特征的局部语义不匹配,影响模型性能。针对该问题,提出方法基于可迁移对抗方法对跨域RUL预测方法展开研究,通过优化局部域鉴别器输出的概率熵,使得对齐特征在局部上难以区分。利用模型在RUL预测过程中的目标互信息进行语义约束,得到同时具有局部可迁移性和目标语义重要性的域不变特征,提升模型的泛化能力。在CMAPSS航空发动机数据集上进行的实验表明,该方法在RMSE和SCORE两个指标上均超过现有的其他跨域自适应方法,证实了其有效性。

In the task of predicting the remaining useful life(RUL)of aviation engines,the scarcity of labeled data and the variability of operating conditions result in significant distribution differences among sensor time series,hampering the generaliza⁃tion ability of RUL prediction methods.Cross⁃domain learning offers a feasible solution to this challenge.Traditional cross⁃domain learning minimizes the distribution discrepancy between the source and target domains to obtain aligned features,facilitating cross⁃domain knowledge transfer.However,as aviation engines degrade,the semantic information between consecutive time steps changes,causing local semantic mismatches in the previously aligned features,which adversely affects model performance.To ad⁃dress this issue,this paper explores a transfer adversarial approach for cross⁃domain RUL prediction,optimizing the probability en⁃tropy of local domain discriminator outputs to make aligned features indistinguishable at the local level.The method utilize the tar⁃get mutual information during RUL prediction to impose semantic constraints,resulting in domain⁃invariant features with both local transferability and target semantic importance,thus enhancing the model’s generalization ability.Experimental results on the CMAPSS aviation engine dataset demonstrate the effectiveness of this approach,outperforming existing cross⁃domain adaptation methods in terms of RMSE and SCORE metrics.