Abstract: Due to the complex operating conditions of turbofan engines,it is difficult to extract the degraded features of high-dimensional and multi-parameter monitoring time series data,which affects the prediction performance of the model.To break the predicament,a remaining useful life prediction model based on bidirectional gated variational encoding regression network was proposed.In regard to the model,the Bidirectional Gated Recurrent Unit(BiGRU)network was introduced to the Encoder of Variational Autoencoder(VAE)to capture the hidden features thoroughly for time series in the multidimensional degraded data.Then,the regression network was added to decoder of the VAE,which took extremely advantage of the features in the latent space of the VAE to train the regression network.Furthermore,Kullback-Leibler(KL)divergence and regression error were combined in the loss function with the purpose of improving the remaining useful life prediction accuracy.To verify the efficiency of the proposed prediction model,a comparison was made with other prediction models on the open turbofan engine data set.The experimental results indicated that the proposed model had better predictive accuracy.

Key words: remaining useful life prediction, variational autoencoder, bidirectional gated recurrent unit network, regression network, turbofan engine

摘要： 针对涡扇发动机运行工况复杂,难以提取高维度、多参数监测数据的退化时序特征,从而影响模型预测性能的问题,提出一种基于双向门控变分编码回归网络的剩余使用寿命预测模型。首先在变分编码器(VAE)网络的编码端引入双向门控循环单元网络(BiGRU),充分挖掘多维度退化数据中的隐藏时序特征;其次重构变分编码器模型的解码器为回归网络,利用变分编码器潜在空间中的退化特征训练回归网络,并在损失函数中联合KL散度和回归误差来提高剩余使用寿命预测精度。为验证所提预测模型的高效性,在公开涡扇发动机数据集上与其他预测模型进行对比,验证了所提模型具有更优的预测精度。

关键词: 剩余寿命预测, 变分编码器, 双向门控循环单元网络, 回归网络, 涡扇发动机