Center for AI

Preventing water damage with machine learning algorithms

Water damage can become costly in short order, as happened in the Highlight Towers in Munich when a ruptured water line in the 21st floor of the IBM Watson Center was first discovered early Monday morning. Icy temperatures over the weekend led to a frozen water pipe that eventually burst. Of all places, it was the client center with the most expensive hardware in the floor below that was impacted. The fully-booked customer event facility, equipped with sensitive technology worth several million euros, could not be utilized for a period of six weeks. One thing was obvious to everyone: this type of incident could never happen again. For this reason fortiss and IBM worked together to develop a prototype of a novel sensor with machine learning algorithms.
rof. Dr. Birte Glimm
Professor Birte Glimm, fortiss research fellow: “With the combination of knowledge- and learning-based methods, we managed to develop a transparent algorithm that also increases the detection rate and accuracy.”

However, the experts at IBM were not satisfied with the capability to detect water damage early in all systems. To address this issue, fortiss joined forces with IBM to develop a prototype of a special sensor that relies on AI and machine learning to detect even the smallest leaks and thus prevent water damage in the future.

While the new sensors are only 6 centimeters, the compact dimensions in no way belie their enormous range. Only four sensors are needed to monitor an entire floor – or 1,000 m2 - of the IBM Watson Center in Munich. Built into a false floor, false ceiling or a wall, the sensors continually monitor the humidity. If this parameter rises, the algorithm reacts and sends a push notification to all relevant persons. In turn, these people can use a specially-designed dashboard to see where and to what extent humidity has been detected. That means all persons in a position of responsibility are aware of possible water damage, at any time and from anywhere, and can thus respond on short notice.


Two algorithms for maximum transparency

The sensors detect whether unusually high humidity is present at a certain point in the building, localize the problem and also indicate what kind of problem has occurred. fortiss and IBM worked together to develop two algorithms for analyzing the sensor measurement data. The results are visualized in the dashboard.

The IBM-modified sensors continually measure the humidity in the spaces being monitored. A specially-developed swarm algorithm shows whether, and if so where, the humidity levels are changing. The algorithm reacts to heavy and irregular changes in humidity. As part of the process, artificial particles move toward a specific point. The particles gather faster and in increasing numbers as the humidity level rises, thus allowing the detection and localization of leaks. The dashboard visualization makes it easy for users to understand at what point a concrete problem exists and how big it is. The more particles that collect at a specific point, the higher the humidity level. In addition, a confidence level indicates how certain the algorithm is that it has identified a leak or how serious it is. This additional information helps to better structure the notifications in the dashboard and reliably identify potentially dangerous leaks.

A further algorithm developed by fortiss merges information from various sources together: humidity sensor data, swarm algorithm coordinates, weather and building plan. Using logic rules as a basis, the algorithm not only decides if a water break has occurred. It also provides reasons for the higher humidity. In this case ruptured pipes and open windows are initially detected. The rules that tie a leak to possible causes calculate the degree of reliability of each possible cause. That means for instance that an open window is more likely the cause for higher humidity when it’s raining. To quickly process the continuously incoming data, a technique referred to as “incremental reasoning” is used and adapted so that the process can be directly executed in the embedded systems, which in this case means the humidity detection sensors.

Further visualizations provide more precise insights into the reported incidents. The affected area is marked on a floor plan. Another graphic shows the development of the humidity over a longer period of time. Additional 3D animations show how much water is leaking and where. It all adds up to a simple overview with a user-centric design in order to provide a clear and understandable view of all relevant data and as a result significantly simplify the decision-making process.


Dashboard provides a uniform overview

The dashboard features an “urgency” area that shows how quickly particular damage has to be resolved. This aspect takes into account how much humidity is present and can point to what type of water pipe is impacted provided this is the cause. Is it a pipe that operates under high pressure and feeds the sprinkler system for instance? Or is it a smaller pipe that is only leaking a few drops? The dashboard also indicates whether the higher humidity is likely tied to an open window or a leak. Using this information, the person performing the check can determine which measures have to be taken. During the week an open window is certainly not a problem as long as it is not currently raining into the room. On the weekend however, it can lead to serious damage, making it necessary to close the window. In addition, a person with the primary responsibility can be defined in the software, who in turn is shown all of the steps that have to be taken to resolve the issue.


Sensors can prevent water damage

Two-and-a-half years of intensive research led to this comprehensive solution, which is now integrated into the IBM Watson Center in Munich. And it has already paid dividends as diverse leaks were more quickly identified and resolved. The new system has thus been able to prevent serious water damage from occurring. Furthermore, the sensors brought more underlying problems with the water pipes to the attention of the building managers.

The next steps include plans to test algorithm improvements parallel to live operation in order to make the localization even more precise. If successfully implemented, decisions can be made as to how this exciting research project can be scaled. That means costly leaks in other buildings and systems could be prevented in the future. And building managers can head into the weekend somewhat more relaxed.

  Marketing & press

Your contact

Marketing & press

presse@fortiss.org