胡艳艳, 白雅婷. 基于贝叶斯图注意力Transformer的航空发动机剩余使用寿命概率预测[J]. 工程科学学报. DOI: 10.13374/j.issn2095-9389.2024.05.30.003
引用本文: 胡艳艳, 白雅婷. 基于贝叶斯图注意力Transformer的航空发动机剩余使用寿命概率预测[J]. 工程科学学报. DOI: 10.13374/j.issn2095-9389.2024.05.30.003
Probabilistic prediction of aero engine remaining useful life based on Bayesian graph attention Transformer[J]. Chinese Journal of Engineering. DOI: 10.13374/j.issn2095-9389.2024.05.30.003
Citation: Probabilistic prediction of aero engine remaining useful life based on Bayesian graph attention Transformer[J]. Chinese Journal of Engineering. DOI: 10.13374/j.issn2095-9389.2024.05.30.003

基于贝叶斯图注意力Transformer的航空发动机剩余使用寿命概率预测

Probabilistic prediction of aero engine remaining useful life based on Bayesian graph attention Transformer

  • 摘要: 航空发动机作为飞机的心脏,其健康状态对飞机的安全飞行至关重要。深度学习强大的数据挖掘能力,为通过海量历史数据预测航空发动机的剩余使用寿命提供了新方法。然而,传统基于深度学习的方法大都关注于挖掘数据在时间上的关联,而忽略了多个传感器监测数据之间复杂的非欧式空间关系。此外,少有研究考虑数据或者预测过程本身具有的不确定性,缺乏对预测结果可靠性的评估。为解决上述问题,本文提出了一种基于贝叶斯网络和图注意力Transformer的航空发动机剩余使用寿命概率预测方法。将图注意力机制融入Transformer的时间多头注意力模块,结合图注意力网络在空间特征提取上的优势和Transformer模型在时间特征提取的优势,实现数据特征时空关系的联合提取。同时,利用改进的贝叶斯网络度量预测不确定性,在得到剩余使用寿命预测点值的同时给出相应的置信区间。最后,通过在公开航空发动机数据集上的实验,证明了所提模型的有效性和先进性。

     

    Abstract: As the heart of the aircraft, the health of the aircraft engine is very important to the safe flight of the aircraft. The powerful data mining capability of deep learning provides a new method for predicting the remaining useful life of aero engine through massive historical data. However, most of the traditional methods based on deep learning focus on mining the correlation of data in time, and ignore the complex non-Euclidian spatial relationship between the monitoring data of multiple sensors. In addition, few studies consider the uncertainty of the data or the prediction process itself, and the reliability of the prediction results is lacking. To solve the above problems, a probabilistic prediction method of aero engine remaining service life based on Bayesian network and graph attention Transformer is proposed in this paper. The graph attention mechanism is integrated into Transformer's time-multi-attention module, and combined with the advantages of graph attention network in spatial feature extraction and Transformer model in temporal feature extraction, the joint extraction of spatial-temporal relationship of data features can be realized. At the same time, the improved Bayesian network is used to measure the prediction uncertainty, and the corresponding confidence interval is given when the remaining service life prediction point is obtained. Finally, experiments on open aero-engine data sets prove the effectiveness and advancement of the proposed model.

     

/

返回文章
返回