Energy efficiency is becoming a challenging issue due to dramatically increasing demands. Sensing disaggregation and identification of individual electrical appliances from consumption measurements is the key to achieve the energy awareness and efficiency in smart buildings. Less intelligent approaches do not sufficiently take into account a variety of heterogeneous information sources and context knowledge to establish a continuous learning and optimizing ability. In this paper, we propose a lightweight smart home appliance behavior learning and detection system, and elaborate its data mining and behavior learning algorithms. The main contribution of this paper are twofold: first, we propose the concept of appliance fingerprint and propose a variety of appliance-based and context-based fingerprints. Second, we propose a multi-source fingerprint-weighting KNN (FWKNN) classification algorithm and present a boosting framework for continuous online learning and detection. We implement the system architecture and demonstrate a prototype based on IBM Bluemix PaaS cloud platform. Experimental result and analysis prove that FWKNN outperforms other benchmark method in detection accuracy.