Indoor Fire Dataset with Distributed Multi-Sensor Nodes
Description
The dataset comprises 4 fire experiments (repeated 3 times) and 3 nuisance experiments (Ethanol: repeated 3 times, Deodorant: repeated 2 times, Hairspray: repeated 1 time), with various background sequences interspersed between the conducted experiments. All exeriments were caried out in random order to reduce the influence of prehistory. It consists of a total of 305,304 rows and 16 columns, structured as a continuous multivariate time series. Each row represents the sensor measurements (CO2, CO, H2, humidity, particulate matter of different sizes, air temperature, and UV) from a unique sensor node position in the EN54 test room at a specific timestamp. The columns correspond to the sensor measurements and include additional labels: a scenario-specific label ("scenario_label"), a binary label ("anomaly_label") distinguishing between "Normal" (background) and "Anomaly" (fire or nuisance scenario), a ternary label ("ternary_label") categorizing the data as "Nuisance," "Fire," or "Background," and a progress label ("progress_label") that allows for dividing the event sequences into sub-sequences based on ongoing physical sub-processes. The dataset comprises 82.98% background data points and 17.02% anomaly data points, which can be further divided into 12.50% fire anomaly data points and 4.52% nuisance anomaly data points. The "Sensor_ID" column can be utilized to access data from different sensor node positions.
Files
Steps to reproduce
We conducted our experiments in a standard EN 54 test room. To ignite the fire materials (beech wood, commercially available tea lights, cotton material lunts, and halogen-free NHXMH-type cable), we utilized three primary ignition sources: a quartz radiant heater for the wood fire, electrical overload for the cable, a glowing spiral for the lunts, and a heating plate for the candles. Each sensor node recorded sensor measurements at a frequency of once every 10 seconds.