Time series anomaly detection is a pervasive task in cyber-physical system. Due to the intricate dynamics of multivariate time series (MTS), unsupervised anomaly detection has always been a research hotpot. Common methods are mainly based on achieving accurate reconstruction or maximizing probability estimation, but these optimization objectives may not be sufficient to distinguish abnormal from normal. In our work, we introduce a learnable uncertainty factor to connect reconstruction and probability estimation, which can reflect the uncertainties of the reconstructed results and facilitate more precise probability estimates. Meanwhile, a novel Reconstruction-Limited Probability (RLP) loss is proposed, its core is to reduce the estimated probabilities of suspicious time segments, while weakly influence the estimation of other points. Based on Transformer, we propose a modified framework TRLP to improve the effectiveness of probability estimation, which realizes reasonable uncertainty factor estimation through some elaborate modules. Extensive experiments and visual analyses are conducted on four real-world datasets, the results under different evaluation metrics demonstrate that our method outperforms existing state-of-the-art frameworks.