What algorithms are used to detect leaks?

Core Machine Learning Approach

At the heart of the system is a machine learning engine that profiles water usage. This profiling involves:

• Historical Data Analysis: The algorithms analyze past water consumption data, distinguishing between typical usage patterns on weekdays and weekends. This segmentation allows for more accurate baseline establishment.
• Learning Consumption Events: The learning algorithm identifies individual water consumption events by analyzing both the time and volume of water used within specific periods throughout the day. This approach recognizes that water usage often follows habitual cycles.
• Predictive Consumption: By understanding these historical patterns, the algorithms can extrapolate and predict future consumption, establishing an expected range of water usage.
• Dynamic Threshold Setting: Based on the maximum consumption events observed in the past for specific time periods, the algorithm automatically sets "alarm thresholds." These thresholds define the upper limits of what is considered normal water usage.

Identifying Deviations and Triggering Alerts

Any significant deviation from these established normal usage patterns indicates a potential leak. The system utilizes two levels of thresholds:

• Alert Threshold (Minor Alert): Exceeding this threshold triggers a notification to the user.
• Shut-off Threshold (Major Alert): If consumption continues to exceed this higher threshold, another notification is sent, and the system can automatically shut off the water supply to prevent further damage.

Specialized Drip Detection

Beyond general consumption analysis, a specific "drip detection algorithm" is employed. This algorithm focuses on identifying an unusual frequency of tiny flow spikes (flows less than 0.5 litres per minute) within a defined time window, indicating a persistent drip or small leak. More information can be found here.

Adaptation and User Control

The system's intelligence lies in its ability to adapt to changing usage patterns and provide users with control:

• Dynamic Learning Mode: This mode allows the system to set detailed thresholds tailored to a building's unique consumption profile. By segmenting data by weekdays, weekends, and time, the algorithm accounts for predictable daily and weekly routines.
• User-Initiated Learning: Users or building staff can manually run the learning algorithm when needed, such as after initial installation, when a new tenant moves in, or after a leak has been fixed. This updates the system's profile based on current usage.
• Manual Threshold Overrides: Users can view and manually adjust or override the learned thresholds within the software, allowing for adjustments based on anticipated changes in usage.
• Holiday Alarms: For periods of abnormal usage, such as vacations (lower use) or filling a swimming pool (higher use), holiday alarms can be set to temporarily override normal settings within a chosen time range.
• Continuous Relearning (On-the-Fly): The system can also learn "on-the-fly." If a user confirms that water usage during an alert was normal, this feedback tells the software not to shut off the water supply and helps refine future predictions. This can also be done by tagging alerts in the app.

In essence, these leak detection systems use machine learning on historical, time-segmented data to establish and dynamically adjust consumption thresholds. This adaptive profiling, combined with user-driven learning and manual overrides, provides a robust and flexible approach to identifying potential water leaks.