In order to more reliably improve the non-intrusive load decomposition ability of residents, the whole identification process is divided into two processes: load event identification and state identification. A MK change point detection algorithm based on dual sliding windows and a weighted dual-parameter hidden Markov model are respectively proposed for load event recognition and state recognition. In the state recognition, the appearance probability of each internal state of each device in each time period is brought into the algorithm as a weight. Through data verification and comparative experiments, it is shown that the NILM algorithm mentioned can more effectively identify the residents’ load.