Aiming at the low recognition rate of non intrusive load monitoring for steel plant load action, a non-intrusive load detection decomposition algorithm of steel mill based on event perception is studied in this paper. The algorithm firstly divides the events into two types: state change events caused by load switching behavior and mode change events caused by load working process. The state change event is detected by feature difference, and the intersection number of the obtained power curve and the average power line is introduced as a new feature to detect the mode change event. Finally, load identification is realized by analyzing events. In this paper, the effectiveness of the algorithm is verified by the actual obtained electricity consumption data of the steel mill. The algorithm improves the accuracy of detection, reduces the rate of missed detection, has the capacity of load identification, and has a better effect on the restoration of production details.