Recognizing the states of household appliance is helpful to monitor the power consumption and model user behaviors at home. Non-Intrusive Load Monitoring (NILM) receives widespread attention as it can identify a individual appliance state using a single sensor. However, presented approaches today can not be adopted in actual home scenarios because they either ignore the energy limitation of sensors or require a complex user configuration. To solve this problem, this paper proposes IEHouse which is a Non-Intrusive Household Appliance State Recognition System. It leverages a supervised learning process over the labeled appliance data sets which can be constructed dynamically based on a small number of appliance profiles. It uses Deep Neural Network (DNN), Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) models, to identify appliance states and improves the accuracy through online learning gradually. By simulating a common household scenario, the energy consumption of sampling sensor is 5.12fcJ per week and the average accuracy of recognizing 10 mixed typical appliance states is 92.9%, which achieves better accuracy with low energy.