Enhancing Multivariate Time Series Anomaly Detection with 2D Spatial Representations and Channel Attention

Anomaly detection in multivariate time series is crucial for safeguarding the reliability of industrial, financial, and cybersecurity systems. However, conventional deep learning approaches often struggle with noise sensitivity and limited ability to capture complex intra- and inter-channel dynamics. To address these challenges, we propose a vision-inspired framework that transforms multivariate time series into multi-channel two-dimensional representations, allowing convolutional neural networks (CNNs) to extract local temporal patterns and spatial dependencies across variables. This representation enhances noise robustness and facilitates the modeling of periodic and evolving patterns often overlooked by 1D methods. To further improve performance, we incorporate Instance Normalization (IN), which preserves instance- and channel-specific variations critical for detecting subtle anomalies, and a Squeeze-and-Excitation (SE) block for adaptive channel weighting. Experiments on five benchmark datasets demonstrate that our approach consistently outperforms state-of-the-art methods, particularly under noisy conditions, highlighting its effectiveness and practicality for real-world anomaly detection tasks.