PatchAD: Patch-based MLP-Mixer for Time Series Anomaly Detection

• DOI: 10.48550/arxiv.2401.09793
• Publication Date: 2024-01-01
• Authors: Zhijie Zhong; Zhiwen Yu; Yiyuan Yang; Weizheng Wang; Kaixiang Yang
• Keywords: Anomaly (physics); Code (set theory)

Anomaly detection stands as a crucial aspect of time series analysis, aiming to identify abnormal events in samples. The central challenge this task lies effectively learning the representations normal and patterns label-lacking scenario. Previous research mostly relied on reconstruction-based approaches, restricting representational abilities models. In addition, most current deep learning-based methods are not lightweight enough, which prompts us design more efficient framework for anomaly detection. study, we introduce PatchAD, novel multi-scale patch-based MLP-Mixer architecture that leverages contrastive extraction Specifically, PatchAD is composed four distinct MLP Mixers, exclusively utilizing high efficiency architecture. Additionally, also innovatively crafted dual project constraint module mitigate potential model degradation. Comprehensive experiments demonstrate achieves state-of-the-art results across multiple real-world multivariate datasets. Our code publicly available https://github.com/EmorZz1G/PatchAD