### 基于深度卷积生成对抗网络的缺失数据生成方法及其在剩余寿命预测中的应用

1. 火箭军工程大学 导弹工程学院, 西安 710025
• 收稿日期:2021-04-23 修回日期:2021-06-29 出版日期:2022-08-15 发布日期:2021-06-29
• 通讯作者: 司小胜,E-mail:sxs09@mails.tsinghua.edu.cn E-mail:sxs09@mails.tsinghua.edu.cn
• 基金资助:
国家自然科学基金(61833016,62073336,61922089)

### Missing data generation method and its application in remaining useful life prediction based on deep convolutional generative adversarial network

ZHANG Shengfei, LI Tianmei, HU Changhua, DU Dangbo, SI Xiaosheng

1. College of Missile Engineering, Rocket Force University of Engineering, Xi'an 710025, China
• Received:2021-04-23 Revised:2021-06-29 Online:2022-08-15 Published:2021-06-29
• Supported by:
National Natural Science Foundation of China (61833016,62073336,61922089)

Abstract: Data-driven Remaining Useful Life (RUL) prediction methods are designed to determine the remaining useful time of the equipment through analyzing and modeling the equipment operation monitoring data; hence, data quality is directly related to prediction accuracy. The development of sensor technology and popularization of the Internet of Things induce the exponential growth of the data scale; however, massive data often face the data missing problem, imposing critical impact on the accuracy of modeling and the reliability of decision-making. Aiming at the problem of RUL prediction with missing data, this paper proposes a generation method for missing data based on the Deep Convolutional Generative Adversarial Network (DCGAN). This method uses a deep convolutional network to learn the true distribution, fills in the missing data based on the powerful learning ability of the Generative Adversarial Network (GAN), and improves the network defects of unstable training and too free modeling in combination with the idea of the Kolmogorov-Smirnov (K-S) non-parametric test. A Bidirectional Long Short-Term Memory (Bi-LSTM) network is further adopted to establish a device degradation trend prediction model, and the RUL prediction is realized by prediction of the time for the degradation to exceed the failure threshold. The proposed method is validated based on the lithium battery degradation data. The RUL prediction results with the complete data, the generated data, as well as the missing data, respectively, are compared to further demonstrate the superiority of this method in solving the missing data problem. The robustness of the RUL prediction towards missing data can also be improved.