Currently, energy management within buildings is essential to mitigate climate change. To this end, buildings are increasingly equipped with sensors to assist the building manager. Yet, the heterogeneity and the large amount of generated data make this task quite difficult. The SANDMAN multi-agent system, described in this paper, aims to assist in the automatic detection, in constrained time, of several types of anomalies using raw and heterogeneous data. SANDMAN features a semisupervised learning by considering some feedback from an expert in the field. The results show that SANDMAN detects different types of anomalies, is resilient to noise and is scalable.