Non-intrusive load monitoring(NILM) can identify the appliance categories, using time, work duration and the work states through data collected from a smart meter. With NILM technique,people can conduct the power supply and power consumption scientifically and reasonably so as to partially or completely solve the problem of imbalance between power supply and demand. The low-frequency non-invasive load state identification algorithm can identify the load using low-frequency signal without additional hardware facilities, so it has broad development prospects. However, it is difficult to improve the accuracy of low-frequency load identification because there is little information available. In this paper, a low-frequency non-invasive load state identification algorithm combining electrical switch state classification LSTM algorithm and maximum likelihood probability algorithm based on recursive reasoning is proposed. The load identification is divided into two steps, which reduces the overall difficulty and improves the accuracy of state identification. The experimental results on Ampsd2 data set show that the proposed algorithm can achieve better results than other NILM algorithms.