2D-Variation convolution-based generative adversarial network for unsupervised time series anomaly detection: a MSTL enhanced data preprocessing approach

Abstract

Time series anomaly detection (TSAD) is a critical task in various research fields such as quantitative trading, cyber attack detection, and semiconductor outlier detection. As a binary classification task, the performance of TSAD is significantly influenced by the data imbalance problem, where the datasets heavily skew towards the normal class due to the extreme scarcity of abnormal data. Furthermore, the limited availability of anomaly data makes it challenging to perform manual labeling, which leads to the development of unsupervised anomaly detection approaches. In this paper, we propose a novel generative adversarial network (GAN) with Multiple-Seasonal-Trend decomposition using Loess (MSTL) data preprocessing algorithm for unsupervised anomaly detection on time series data. With the MSTL data preprocessing algorithm, the network architecture is simplified, thereby alleviating computational burden. A 2D-variation convolution-based method is integrated into the GAN to enhance feature extraction and generalization capabilities. To avoid the model collapse problem caused by data deficiency, multiple generators are employed, and a joint loss function is designed to improve the robustness of the training process. Experiments on several benchmark datasets from various domains demonstrate the efficacy and superiority of our approach compared to existing competitive approaches.