Time Series Anomaly Detection Based on Asymmetric Autoencoder and Gaussian Mixture Model

Time series data are widely used in domains such as finance, industry, transportation, and healthcare. Anomaly detection in multivariate time series is a challenging unsupervised machine learning task due to the high dimensionality, noise, and lack of labeled data. Building upon the Deep Autoencoding Gaussian Mixture Model (DAGMM), this paper proposes an asymmetric network architecture featuring a fully connected encoder and a convolutional decoder. This design offers two key advantages: the simple decoder helps to filter out noise in the encoding process, and the convolutional structure of the decoder preserves contextual dependencies in the temporal data. Furthermore, we introduce a novel dataset construction framework that integrates traditional unsupervised learning methods to enhance data selection and preprocessing. Experimental results demonstrate that the proposed method achieves an F1-score of 94.4 on benchmark time series anomaly detection datasets, showing significant performance improvement.