Multivariate time series anomaly detection using diffusion models based on fusion and inference

Multivariate time series anomaly detection is crucial for ensuring reliability in critical applications such as Industrial Internet of Things (IIoT) equipment monitoring and cloud computing platform failure early warning. However, existing methods often struggle when dealing with complex time series characterized by high dimensionality, strong inter-dependencies, and significant variability, particularly in regions with concentrated anomalies. To address these challenges, we propose FIDiff, a diffusion model based on fusion and inference. The framework designs a diffusion model tailored for masked data, which incorporates a time series filling denoiser to enhance the modeling of complex temporal dependencies and mitigate interference from anomalies and noise. Furthermore, a unique multi-step diffusion-based fusion and inference strategy is introduced, which integrates intermediate outputs from models with different diffusion steps and employs a dynamic thresholding mechanism for anomaly scoring, thereby significantly improving detection accuracy and stability. Extensive experiments on six public datasets demonstrate that FIDiff achieves an average improvement of 3.7% in F1-score compared with the best-performing model among 13 state-of-the-art baselines, with consistent performance gains across all datasets, validating its effectiveness and practical value. The code for this project is available at https://github.com/secMTS/FIDiff.git .