ImConDM

Abstract

Multivariate Time Series (MTS) anomaly detection is critical for various applications. Yet, most existing methods rely on forecasting or reconstruction-based frameworks, which suffer from inherent uncertainties in handling complex MTS data. The recently emerging imputation-based anomaly detection leverages partial observations to reconstruct the masked values, thereby enhancing reconstruction stability and detection performance. However, in existing imputation-based methods, training on the masked data may lead to the omission on the global spatiotemporal distributional features of the MTS. Additionally, multiple masking strategies must be employed to ensure all data points have the opportunity to be reconstructed, resulting in repeated training and significant computational consumption. To address these limitations, this paper proposes ImConDM, a novel anomaly detection framework integrating imputation with conditional diffusion models. It trains a diffusion model on the complete training data and leverages the controllable sampling capabilities of diffusion models to perform conditional imputation guided by observed values, enabling stable and comprehensive MTS feature learning without retraining. Extensive experiments on six real-world MTS datasets demonstrate that ImConDM significantly outperforms eleven state-of-the-art baselines. Notably, compared with the second-best model, it achieves a 20.5% improvement in detection accuracy on high-dimensional datasets and reduces training time by at least 58%. The executable codes of the experiments with datasets and our algorithms are available at https://github.com/Labman-Wangjunjie/ImConDM.