基于多模态Transformer的机电作动器剩余寿命预测  

Prognosticating Remaining Useful Life of Electro-Mechanical Actuators Using a Multi-mode Transformer Model

作　　者：陈子涵 CHEN Zihan(Institute of Remote Sensing Satellite,China Academy of Space Technology,Beijing 100094,China)

机构地区：[1]中国空间技术研究院遥感卫星总体部,北京100094

出　　处：《兵工学报》2023年第10期2920-2931,共12页Acta Armamentarii

摘　　要：机电作动器在航空航天装备中扮演着重要角色。针对机电作动器剩余寿命预测问题,提出一种基于多模态Transformer模型的机电作动器寿命预测方法。该方法直接使用多通道传感器数据作为输入,综合考虑多模态数据信息,并且不需要人工特征提取等预处理步骤。多模态Transformer模型利用多头自注意力机制从不同的表示子空间中自适应学习全局特征,能够避免传统深度学习方法难以学习全局特征的缺点。利用多模态Transformer的编码器部分并行提取多模态传感器时间序列中不同传感器的特征,并实时直接预测剩余使用寿命。采用由编码器和解码器组成的完整多模态Transformer模型预测机电作动器的关键性能参数,可同时更直观地预测关键寿命参数的退化过程。使用机电作动器全寿命试验数据验证该方法用于寿命预测的有效性。试验结果表明,所提方法能够准确地直接预测剩余寿命,同时预测关键性能参数的寿命退化过程。Electro-mechanical actuators play an important role in next-generation spacecraft.To address the challenge of predicting the remaining useful life of electro-mechanical actuators,a failure prognostic algorithm based on a multi-mode Transformer model is proposed.Multi-channel sensor data are directly used as inputs to the Transformer model without feature extraction as a pre-processing step.The multi-mode Transformer uses multi-head attention to adaptively learn global features from various representation subspaces.The Encoder of the Transformer is utilized to extract features from different sensors across the time series in parallel and predict the remaining useful life directly.Simultaneously,the full Transformer,composed of an Encoder and a Decoder,is used to prognosticate key performance parameters of electro-mechanical actuators.A benchmark dataset is used to validate the effectiveness of the proposed model for electro-mechanical actuator failure prognostication.Experimental results reveal its advantage in accurate prediction and failure prognosis.

关 键 词：机电作动器 寿命预测 多模态数据 注意力机制 

分 类 号：V474[航空宇航科学与技术—飞行器设计]