<p>Multivariate time series anomaly detection is crucial for maintaining the safe and stable operation of industrial equipment. Most existing reconstruction-based methods focus on identifying data patterns by reconstructing the entire time window. However, these methods often generate data from a lower-dimensional space, leading to essential feature compression and information loss in the original high-dimensional space. Furthermore, the noise or anomalies existing in complex monitoring data significantly impact the generalization capability of models. Several generative or memory-based methods have been proposed to mitigate this issue through implicit or explicit noise suppression structures. However, the former encounters challenges in training convergence, while the latter needs sacrificing some reconstruction performance. In this paper, we propose a self-supervised imputed reconstruction method for multivariate time series anomaly detection based on diffusion models (IRDM). First, IRDM incorporates the reconstructed masking strategy to divide data into known and unknown areas. Subsequently, IRDM constructs imputed-reconstruction diffusion models and trains them via self-supervised conditional generation. This process imputes the unknown area based on the known data and obtains a reconstructed output closely aligning with the original data distribution. Finally, a negative exponential noise schedule is designed, which constrains the noise variance size in each step of the denoising process. This allows the model to learn more original data information, thereby mitigating the negative impacts of noise or anomalies. Extensive experiments conducted on 5 representative cyber-physical system datasets demonstrate that the proposed algorithm outperforms 17 typical baselines.</p>