Monitoring complex systems requires multi-dimensional signal data collected by sensors, and the signal could be applied to perform various downstream tasks, such as structure state labeling, damage detection, and anomaly detection. However, determining optimal sensor locations within these systems poses significant challenges and requires specialized knowledge. Previous research has extensively investigated various sensor placement methods, including heuristic approaches, optimization techniques, statistical methods, and machine learning techniques. While these methods offer valuable insights, they often face limitations such as oversimplified assumptions, high computational demands, and inadequate scalability. Additionally, they do not simultaneously consider the complex structure of systems and the graph level signals at the same time within their approaches. In this paper, we propose a K-means Sampling approach for sensor selection. Our K-means Sampling method treats sensor placement as a graph vertex sampling problem, combining both signal and graph-level information with the criteria of signal reconstruction. In addition, we introduce a Binary Mask Learning (BML) Model that integrates sensor selection with reconstruction and anomaly detection tasks. During these tasks, signals from both low and high frequencies, sourced from simulations and real-world data, are investigated to examine the quality of our selected sensor. In summary, our framework for sensor sampling incorporates concepts from Graph Signal Processing (GSP), Vertex Sampling, Graph Signal Reconstruction, and Anomaly Detection.