An annotated dataset of Building Management System periodic alarms

This article presents alarm events data collected from 22 Building Management Systems (BMS) from 9 countries (mainly European) and 7 market segments (Commercial, Education, Entertainment, Government, Healthcare, Residential, Transport) over a period of 1 month. 321 alarm event timeseries were collected. Each series corresponds to a single (unknown) user-defined alarm rule switching from "inactive" (reset) to "active" (alarm) state and back. Such a rule usually consists in a simple upper or lower threshold applied to an observed variable such as a room temperature sensor, a valve pressure, etc. - possibly with hysteresis, delays, or Boolean logic combining several thresholds. Only change of state events are recorded in the dataset, together with the new state after transition. Timestamps are provided raw without any uniformization resampling, preserving the original time resolution of the series. With a volume of 110,457 alarm events ranging from 11 to 13,032 events per series, this dataset illustrates the diversity of alarm events patterns in buildings, and challenges associated with non-uniform sampling. Researchers may wish to leverage it to benchmark the execution time and performance of pattern mining, machine learning and AI approaches. Example tasks include modelling and mining of single series (periodicity detection, temporal clustering, frequent pattern mining…) or multiple series (clustering, frequent pattern mining…). Approaches may leverage the complete binary (alarm/reset) state timeseries or prefer to filter the reset events to focus solely on activation events sequences. To encourage comparable reproducible results for learning tasks, the dataset is split into predefined train and test files containing 237 and 84 series respectively. The training data corresponds to May 2024 while the test data is from Jan-Feb 2026. The test data originates from one building that is strictly distinct from training data. For this reason, researchers may wish to use the test dataset for validation purposes and interpret associated results as hints of generalization capabilities of the candidate approach under test. The names of the series are fully anonymized, with names spanning from A0 to A236 and from A1000 to A1083. Finally, each timeseries was presented to a panel of subject matter experts to annotate it as periodic or non-periodic according to a consensus of votes. Two kinds of periodicity labels were considered: "is repeat periodic", representing a perception that the time duration between consecutive alarm events is regular, and "is daily periodic", representing a perception that there are typical daily patterns that occur frequently in the series. These two labels may be used to evaluate periodicity detection algorithms, whether supervised (classification task) or unsupervised (clustering task). Both labels may be combined with a logical OR to create a broader "is periodic" label.