Deep Anomaly Detection on Time Series

时间序列是无处不在的，但鉴于最近深度学习带来的巨大进步，时间序列的异常检测(AD)仍处于初级阶段。在这个项目中，我们将开发基于深度学习的在时间序列上创建现代AD的方法。为此，我们将把图像AD中的思想(如自我监督的对比学习和离群点曝光)带到时间序列上的AD。这种方法依赖于各种补充数据(如强大的数据增强方案和大量负样本的可用性)，而这些数据对于时间序列来说是出了名的困难或根本不可能获得。因此，我们将研究取代现有补充剂的创新策略。我们的方法是基于务实地学习所需的数据增加和负样本。此外，我们还将从头开始研究时间序列上的AD方法。建议的模型将基于随机场来模拟数据梯度上的分布。最后，我们将使用对比学习将化学过程的先验知识融入到我们的方法中，从而实现对化学过程数据的少量异常检测。

Time series are ubiquitous, but - in the light of the enormous recent advances enabled by deep learning - anomaly detection (AD) on time series is still in its infancy. In this project, we will develop methods creating modern AD on time series, which is based on deep learning. To this end, we will carry ideas from image AD (such as self-supervised contrastive learning and outlier exposure) over to AD on time series. Such methods rely on a variety of supplemental data (such as the availability of powerful data-augmentation schemes and massive corpora of negative samples) that is notoriously difficult or simply impossible to obtain for time series. As a result, we will study innovative strategies replacing the established supplements. Our approach is based on pragmatically learning the required data augmentations and negative samples. In addition, we will develop methods for AD on time series from the ground up. The proposed model will be based on a random field to model distributions over the gradients of the data. Lastly, we will use contrastive learning to incorporate prior knowledge on chemical processes into our methods, enabling few-shot anomaly detection on chemical process data.