MPPG: Pluggable Multi-Periodic Pattern-Guided Approach for Multivariate Time Series Anomaly Detection

In recent years, reconstruction-based deep models have been widely applied in time series anomaly detection. However, these methods often face challenges in extracting sufficient features for complex time series, leading most models to also reconstruct anomalies well. Moreover these methods obtain patterns at a single scale and do not take into account the multi-periodicity of the time series. To address this issue, we propose Multi-Periodic Pattern Guidance (MPPG), an approach that utilizes multi-periodic patterns to guide model reconstruction. MPPG is a pluggable method applicable to various reconstruction models based on the Encoder-Decoder Architecture. First, through an unsupervised loss function, it ensures that features within the same period are closely together, while features from different periods are distinctly separated. The process facilitates the extraction of pattern information unique to each period. Then it integrates this pattern information into the feature space of any reconstruction model to improve the reconstruction accuracy and anomaly detection capabilities. We evaluated MPPG on six real-world datasets and achieved good experimental results.