MLP-Mixer based Masked Autoencoders are Effective, Explainable and Robust for Time Series Anomaly Detection

Time series anomaly detection remains one of the most active research areas in data mining due to its wide range of real-world applications. In recent years, numerous deep learning-based methods have been proposed for this task. However, deep learning-based methods fail to detect subsequence anomalies with long durations, lack explainability, and are vulnerable to training set contamination. This paper addresses these issues by proposing a novel deep learning framework for effective, explainable, and robust time series anomaly detection. Our framework, MMA, incorporates the MLP-Mixer backbone with a Masked Autoencoder-based anomaly detection approach to allow for a significantly larger input window size (10 to 20 times larger than the input window sizes of current models). This larger input window enables our model to detect challenging subsequence anomalies. Meanwhile, a contrast learning module is proposed to aid in detecting subtle anomalies that fail to be identified by residual errors. Furthermore, a dynamic anomaly filtering method is introduced to mitigate the impact of subsequence anomalies on the reconstruction of surrounding normal regions to reduce false alarms. Extensive experiments on univariate and multivariate time series datasets demonstrate that our proposed framework significantly outperforms state-of-the-art methods across rigorous evaluation metrics. Additionally, MMA has a strong ability to reconstruct potential normal patterns in anomalous regions, providing high levels of explainability. Moreover, MMA demonstrates high robustness to various types of training set pollution.