Recent advancements in Artificial Intelligence of Things (AIoT) technologies have brought about an increasing popularity in leveraging deep learning algorithms to detect potential failures in cyber-physical systems (CPS). Typically, an anomaly detection model is deployed to monitor the multivariate time series (MTS) generated by sensors to identify abnormal operation states. However, the contemporary unsupervised deep learning models for MTS anomaly detection are susceptible to contamination in the training dataset and are incapable of capturing the spatial-temporal correlations in MTS, result in suboptimal practical detection performance. In this paper, we propose a novel framework called Spatial-Temporal Normalizing Flow (STNF) to tackle the above problems. Our framework extends the conditional normalizing flow for MTS density estimation, aiming to achieve robust anomaly detection against training dataset pollution. Additionally, we introduce a patched Long Short-Term Memory (LSTM) module to effectively learn robust representations of long-term dependencies within MTS. Moreover, a dynamic graph construction module is devised to model the complex and evolving correlations among different dimensions of MTS. We evaluate our approach on three real-world CPS datasets and report improvements over the state-of-the-art approaches in terms of both performance and robustness.