Elsevier, Pervasive and Mobile Computing
We present a framework to relate variables as they occur in a modern building management system (BMS) that processes data from building-installed sensor and actuators. Our group mining framework extracts a unified event time series as changes in building management variables, derives prepositional variable association rules, and extracts hierarchical variable groups from the derived rules. Variable changes typically occur by either occupants interacting with the building or as response to outdoor environmental changes. As a user enters a building, a sequence of sensor activations will create a specific temporal event pattern that is mapped into a variable hierarchy by our framework. Similarly, as outdoor lighting changes, a variable hierarchy appears that relates variables to the change. To extract variable groups, we introduce a novel hierarchical transitive clustering (HTC) algorithm that constructs a rooted variable tree and then clusters the tree to represent variable group relations. HTC is parameter-free and works unsupervised. We evaluated the group mining framework in living-lab data recorded in different office environments during 14 months. As typical for BMS operation, variables in our dataset represent measurements and control states of building-installed devices and processed context information. HTC showed a correctness of over 0.91 and an average variable coverage of 75%, this improving variable coverage by 40 p.p. compared to previous work. We successfully detected alternative hierarchies and show how variables relate across office rooms.