改进型变分自编码器及其在航天器故障检测中的应用

An Improved Variational Autoencoder and Its Applications on Fault Detection of Aerospace Craft

故障检测技术是实现航天器智能故障诊断和健康管理的基础。针对航天器故障检测领域存在的测试数据量大、故障标签稀缺问题以及对实时性的要求,提出了一种基于改进型变分自编码器(VAE)的无监督故障检测算法。所提算法首先利用长短期记忆(LSTM)单元替换VAE中传统的神经元,提取航天器测试数据间的时间依赖性和相关性等特征。然后提出了一种重构概率计算方法。该网络在只含有正常数据的训练集上进行训练,学习特征,并以较高的概率对训练集上的数据进行重构。当测试集中输入数据的重构概率小于设定的阈值时,则判定对应的输入数据为故障数据,从而实现故障检测。实验表明,所提方法可行,能够有效地对故障进行检测。

Fault detection technology is the prerequisite for realizing intelligent fault diagnosis and health management of spacecraft.In view of the large amount of test data,the scarcity of fault labels and the real-time requirements in the field of spacecraft fault detection,an improved unsupervised fault detection model based on Variational Autoencoder(VAE)is proposed.The algorithm uses Long Short-Term Memory(LSTM)recurrent neural networks to extract features such as time dependence and correlation between test data,and uses VAE to learn the correlation features of multiple test data.The network is trained on the training set containing only normal data,and reconstructs the data on the training set with a high probability.When the reconstruction probability of the input data in the test dataset is lower than the pre-set threshold,the corresponding input data is considered fault data.Experiments show that the proposed method is feasible and can effectively detect faults.