A proposed method for unsupervised anomaly detection for a multivariate building dataset

Supervisors: Julien Nembrini, Denis Lalanne

Student: Roberto Gonzalo Sanchez Alban

Project status: Finished

Year: 2017

Ubiquitous devices employed in building facilities are allowing us to acquire a diverse amount of data relative to the internal systems of buildings. This is contributing to the growing awareness of the gap that exists between the desired performance of a building and its actual performance. Automated fault detection and diagnostic (AFDD) systems have been showed to be effective at detecting the root cause of performance problems. This master thesis is interested in finding motif cluster (typical patterns) and discord clusters (atypical/abnormal patterns), two types of patterns used by some AFDD approaches. Our approach attains to discover daily patterns in a multivariate fashion for a studied building dataset by using the Gaussian Hidden Markov models. The discovered motif cluster profiles define the typical performance of the building, while the discovered discord cluster profiles spot potential performance problems of the building. Three proposed models create a label data frame that summarize all the daily patterns in a table allowing the researcher to do further aggregation about the motif/discord cluster profiles. The proposed models where tested in a case study where the North-East and South-West ventilation systems of the studied building were compared. The results provide information about the pattern evolution across different seasons and years, as well as the dynamics between various variables. In addition, anomalous daily profiles were spotted as a multivariable pattern in the North-East ventilation system, that demonstrate how powerful this approach is. Finally, this approach had good feedback from the building experts and the potential of our approach motivates further research.

Document: report.pdf