Joint training of a predictor network and a generative adversarial network for time series forecasting: A case study of bearing prognostics
Joint training of a predictor network and a generative adversarial network for time series forecasting: A case study of bearing prognostics
复制标题
DOI：
10.1016/j.eswa.2022.117415
复制
发表时间：
2022-05-17
影响因子：
8.5
通讯作者：
Zimmerman, Andrew Todd
中科院分区：
文献类型：
--
作者：
Lu, Hao;Barzegar, Vahid;Zimmerman, Andrew Todd
轴承运行到故障数据的缺乏一直是开发和实际实施鲁棒轴承性能模型的挑战之一。本文提出了一种新的基于生成对抗网络（GAN）的规则语言预测方法。我们提出了一种新的联合训练策略，将轴承健康预测器的训练过程集成到GAN架构中。GAN使用可用的时间序列退化数据来生成合成退化数据，以增强预测器的学习和预测性能，从而提高RUL预测精度。我们通过两个例子证明了所提出的方法的实用性和性能。预测多项式时间序列的第一个数字玩具案例研究表明，建议联合训练的健康预测器（HP-JT）方法产生较小的一步和多步预测误差比传统的健康预测器（HP）。在第二个案例研究中，我们设计了一个交叉验证研究，利用开源轴承数据集来评估模型在RUL预测中的性能。与HP相比，该方法在五重交叉验证研究中将轴承RUL预测平均误差降低了29.4%。我们进一步将该模型与标准的数据增强技术进行比较，例如添加噪声和使用变分自动编码器（VAE）。实例分析结果表明，该方法能够生成反映实际数据分布的时间序列。
The lack of bearing run-to-failure data has been one of the challenges in developing and practically implementing robust bearing prognostics models. This paper proposes a new Generative Adversarial Network (GAN) based prognostics method for RUL prediction. We propose a novel joint training strategy to integrate the training process of a bearing health predictor within the GAN architecture. GAN uses available time series degradation data to generate synthetic degradation data that enhances the predictor's learning and forecast performance, thus improving the RUL prediction accuracy. We demonstrate the utility and performance of the proposed method through two examples. The first numerical toy case study of forecasting polynomial-like time series shows that the proposed Jointly Trained Health Predictor (HP-JT) method produces smaller one- and multi-stepahead prediction errors than a traditional health predictor (HP). In the second case study, we design a crossvalidation study utilizing an open-source bearing dataset to evaluate the model's performance in RUL prediction. Compared to HP, the proposed method decreases the bearing RUL prediction average error by 29.4% in a five-fold cross-validation study. We further compare the model with standard data augmentation techniques such as adding noise and using a variational autoencoder (VAE). The results from the case studies show that the proposed method can generate time series representing the real-data distribution.