Multivariate Time-Series Data Anomaly Detection via Dimension Independence and Reconstruction

Time-series data anomaly detection has received widespread attention from academia and industry, with crucial applications in service monitor, IoT surveillance, and network security. Existing anomaly detection methods for multivariate time-series data (MTS) typically (1) focus solely on low-accuracy anomalies, neglecting explicit errors, which limits the applicability; (2) treat data dimensionality as a fixed parameter during training and inference, which limits the scalability of trained model to data with varying dimensions. To address these problems, we propose a Transformer framework for MTS anomaly detection based on dimension independence and reconstruction. The framework comprises three components: (1) time-series data preprocessing, (2) MTS reconstruction model, and (3) anomaly detection. Preprocessing step resolves explicit errors and prepares data. MTS reconstruction model built on dimension independence and Transformer encoder, reconstructs each MTS dimension independently. Anomaly detection step identifies low-accuracy anomaly and outputs the final anomaly detection results. Comparative experiments demonstrate that the proposed framework outperforms all baseline methods on two real-world datasets and achieves second place on the remaining two datasets. Furthermore, preprocessing and scalability experiments yield highly satisfactory results, validating the completeness and scalability of the framework.