基于生成式对抗网络的多维时间序列补插研究
Multivariate Time Series Imputation Based on Generative Adversarial Network

随着传感器和物联网的广泛应用，大量的多维时间序列被收集。然而，由于传感器损坏、环境变化和机器故障等不同原因，在多维时间序列中存在着许多缺失值，这些缺失值给多维时间序列的下游应用及分析带来了进一步挑战。为此，本文提出了一种基于生成式对抗网络的多维时间序列缺失值补插算法。具体来说，我们使用自编码器作为生成式对抗网络的生成器，循环神经网络作为生成式对抗网络的判别器。利用生成式对抗网络强大的生成能力对多维时间序列数据中的缺失值进行修复。此外，在自编码器的结构中引入注意力机制，使得自编码器在进行缺失值修复时，不但能够考虑到其他维度对该缺失值的影响，还可以直接为重要信息分配更大的权重比例，使得自编码器在修复缺失值时能够更加关注这些重要信息，从而使得修复的缺失值更加准确。通过在PhysioNet数据集上的实验证明，本文提出的方法在多维时间序列缺失值补插方面具有优越的性能。
With the wide application of sensors and IoT, a large number of multidimensional time series are collected. However, there are many missing values in the multidimensional time series due to dif-ferent reasons such as sensor damage, environmental changes and machine failures, and these missing values bring further challenges to the downstream application and analysis of multidimensional time series. To this end, in this paper, we propose a generative adversarial network-based missing value interpolation algorithm for multidimensional time series. Specifically, we use a self-encoder as the generator of the generative adversarial network and a recurrent neural network as the discriminator of the generative adversarial network. The missing values in the multidimensional time series data are repaired by using the powerful generative power of the generative adversarial network. In addition, the attention mechanism is introduced into the structure of the self-encoder, so that the self-encoder can not only consider the influence of other dimensions on the missing values, but also directly assign a larger proportion of weight to the important information, so that the self-encoder can pay more attention to the important information when repairing the missing values, thus making the repaired missing values more accurate. Experiments on the PhysioNet dataset demonstrate the superior performance of the proposed method in interpolating missing values in multidimensional time series.