Anomaly detection on dynamic graphs is crucial for monitoring the security of industrial systems. The challenge in identifying anomalies in time-varying data arises from complex and flexible structures, compounded by the absence of labelling in the data. In particular, the representation of graph patterns and capturing the evolving nature of graphs become challenging due to time varying nature, i.e., dynamic graphs. Contrastive learning in graph-related contexts has gained considerable traction recently, primarily attributed to its label independence and the robustness in representations. In order to address the limitations in dynamic graph representation and anomaly detection, we propose a novel Contrastive learning-based dynamic graph anomaly detection framework (ConDGAD) to improve the time series data representation learning and prediction through dynamic graphs. This enables detection of multivariate time series anomalies at specific time window of measurement levels. ConDGAD first converts the multivariate time series data into dynamic graphs. Then multiple graph augmentations are performed and a novel contrastive learning process is applied on the dynamic graphs. This enable to train a model that can effectively capture the graph dynamics and perform accurate prediction, which subsequently is used for anomaly detection. Evaluation performed on widely used time series datasets, including SWaT and WADI, reveal that the ConDGAD has achieved improved recall and F1 scores for anomaly detection over the state-of-the-art methods. Ablation studies reveal the significance of the our proposed multi-augmented constrastive learning process in achieving the improved performance for anomaly detection on time series data via dynamic graphs.