This is an attempt to develop anomaly detection in multivariate time-series of using multi-task learning. This work is done as a Master Thesis.
Abstract:
This thesis examines the effectiveness of using multi-task learning to develop a multivariate time-series anomaly detection model. There have been many studies on time-series anomaly detection. However, recent studies use either a reconstruction based model or a forecasting model. A reconstruction based model relies on the reconstruction probability, whereas a forecasting model uses prediction error to identify anomalies. Anomalies are either samples with low reconstruction probability or with high prediction error, relative to a predefined threshold. Either way, both models learn only from a single task. Therefore, this thesis attempts to combine existing models using multi-task learning. In addition to that, most recent studies use unsupervised learning due to the limited labeled datasets and it is also used in this thesis. In order to evaluate the model, the proposed model is tested on three datasets (i.e. two public aerospace datasets and a server machine dataset) and compared with three baselines (i.e. two reconstruction based models and one forecasting model). The results show that the proposed model outperforms all the baselines in terms of F1-score. In particular, the proposed model improves F1-score by 30.43%. Overall, the proposed model tops all the baselines which are single-task learning models.
Baseline:
Datasets used in the experiments:
Limitation:
此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。