Abstract: Compressed sensing (CS) has become the main way to solve the problem of massive monitoring data in smart grid due to its simple compression and applicability to monitoring environment. However, its application research in non-intrusive load monitoring (NILM) has not been carried out. In order to meet the needs of time-space intensive acquisition or high-frequency information acquisition in traditional NILM, a method of NILM based on CS (NILM -CS) is explored in depth for the first time. First, the types of original load signals and their eigenvalues are analyzed while the applicability of CS in NILM is proved. Then, three frameworks of NILM-CS and their implementation processes are proposed based on scene recognition, mathematical optimization model and event detection, respectively. On this basis, aiming at the key problems to be solved in three frameworks, the specific processes of the applicable methods for eigenvalues extraction, event detection, mathematical optimization model, and three elements of CS are proposed. Experiments show that three NILM-CS frameworks and solutions to their key issues proposed in this paper are feasible and adaptable. It can ensure that the load recognition accuracy is close to 90%, the load decomposition accuracy is above 92%, and the signal-to-noise ratio is greater than 70 dB in NILM-CS, fully complying with the standards of NILM.