Abstract: Load characteristics refer to the special identification of certain statistical laws during the operation of load equipment. The database containing load characteristics is the basic basis for non-intrusive load monitoring and decomposition. For pure resistive load equipment, researchers can accurately extract the load characteristics in the corresponding state by using the characteristics of the power variation during the switching process of its operating states. However, for non-pure resistive devices, the load feature extraction has the following two problems. Problem 1: The power variation is not significant during the switching of the operating states, which results in difficulty in accurately locating the switching point of the state. Problem 2: The load equipment has an operating state in which the power varies slowly, resulting in the consequence that the load characteristics in the corresponding state is not unique, which cannot be manually extracted. Regarding the above problems, this paper proposed a non-intrusive load feature extraction method based on an online feature library. The operation process of the method was divided into two stages. Stage 1: Establish an online feature library based on the steady-state periodic current array during the operation of the load equipment. The sliding window function was constructed by improving Pearson similarity coefficient to calculate the similarity between the periodic current array and the online feature library when the load equipment is running, and synchronously judge the redundancy of the online feature library to achieve the load equipment status data segmentation. Stage 2: Calculate the feature matrix of the online feature library, perform K-means cluster analysis on the feature matrix, and fuse the similar online feature library to form the state feature library of the load equipment, so as to realize the extraction of the characteristics of the current array of the load equipment. The test results on private data sets and PLAID data sets show that the load feature extraction method proposed in this paper has good robustness in different power consumption scenarios.