Reconstruction-Based Methods for Multivariate Time Series Anomaly Detection : A Review and Taxonomy

Reconstruction-based methods have become a central paradigm for
unsupervised multivariate time series anomaly detection (MTSAD),
especially in the context of cyber-physical and industrial control
systems. Their ability to learn normal patterns and detect deviations
without supervision makes them highly suitable for safety-critical
environments. In this work, we conduct a structured review of recent
deep learning models for MTSAD based on reconstruction principles.
We propose a four-axis taxonomy that captures key design dimensions: (i) backbone architecture (e.g., AE, VAE, GAN, Transformer);
(ii) temporal/frequency modeling; (iii) hybridization strategies; and
(iv) integration of attention mechanisms. We analyze representative models including MCA-VAE, TFMAE, ALGAN, DiffGAN, and
CAE-Transformer, discussing their loss functions, anomaly scoring
strategies, interpretability, and thresholding mechanisms. We also
highlight critical challenges such as overfitting to anomalies, fixed
thresholds under distribution shift, and the difficulty of modeling
both temporal and inter-variable dependencies. This survey aims
to clarify terminology, compare architectural families, and identify
trends and research gaps. We provide a structured comparative table
and discuss the implications for real-world deployment.