MAET: A Generalizable Masked Autoencoding Framework for Anomaly Detection in Time-Series Data

With the rapid development of the Internet of Things (IoT), an increasing number of Cyber-Physical Systems (CPSs) are being created, and the application of artificial intelligence technology to these systems has become a widely discussed topic. Anomaly detection for the sensor data and time-series generated within these systems to ensure efficiency and safety remains a challenging problem. Due to the lack of sufficient anomaly data in real-world scenarios, most current approaches focus on modeling normality and learning the latent patterns of normal data in order to predict or reconstruct it. These include forecasting models that use time windows to predict future moments, or reconstruction models that reconstruct current time windows to obtain normal time-series. In this paper, we introduce a masked autoencoder for time-series anomaly detection in CPSs (MAET), which features a unique autoencoder structure (with LSTM as the base network) and a random masking mechanism. A latent representation of the unmasked time-series fragment is obtained by feeding it to the encoder, which is then sequentially combined with the masked part and fed to the decoder to obtain the reconstructed time-series. MAET represents an improvement over traditional reconstruction methods. Our experimental results show that MAET can reconstruct normal data with smaller errors and achieve better reconstruction results.