Machine Learning Applications

The Hollandse Brug and the InfraWatch project

Keeping vital infrastructure safe for use is challenge that is increasingly being met by the application of machine learning and datamining techniques.  An example is the Hollandse Brug of which the integrity is now being monitored and evaluated with a myriad of sensors.

The Hollandse Brug is a Dutch highway bridge along the A6 highway, which connects the Dutch capital Amsterdam to the large commuter city Almere, which is located in Flevoland, a large piece of land that was reclaimed from the sea.
The bridge is over 40 years old, and after inspection in 2007, it was decide to perform large scale renovations, and to start monitoring the integrity of the bridge extensively, during the maintenance work and in the period afterwards. For this aim, the bridge was fitted with a large collection of sensors that continuously measure the traffic load on the bridge, and the infrastructures response to that. Since the fitting of the monitoring system in 2008, data is being gathered with a steady rate of around 11 Gb per day.

The data gathering was initiated by the team of Bas Obladen at Strukton B.V., who also designed and implemented the monitoring system on the bridge. Although large amounts of data were being collected, representing a detailed and constantly growing history of bridge performance, no integrated analysis was being done. In other words, a rich source of information concerning the relation between traffic load and bridge response was not being exploited to its fullest. In 2009, a consortium of civil engineers (TU Delft), computer scientists (LIACS, Leiden University) and domain experts (Strukton), along with interested third parties, was formed with the aim of leveraging the potential knowledge contained in the ever growing stream of data. With considerable funding from the Dutch STW agency, the InfraWatch project (see was started.

The current sensor network on the bridge contains over 145 sensors of various types, measuring different aspects of the bridge at a rate of up to 100 times per second. The majority of the sensors concern so-called strain gauges, which monitor the strain on specific parts of the structure. Whenever a large vehicle passes in the vicinity of the sensor, this appears as a wide bump in the signal. Additionally, there are vibration sensors, which measure shaking of the bridge, as well as a weather station and a video camera, to observe the traffic that causes the different levels of strain and vibration. The main goals of InfraWatch are to manage the incoming flood of data (not a small challenge), to model the different phenomena clearly present in the sensor-data, and to assess the health of the bridge based on gradual or sudden changes in the behavior of the bridge.

At the Leiden Institute for Advanced Computer Science (LIACS), machine learning techniques are being used to analyze the data, and model the relationship between traffic and bridge response. The bridge data is an interesting challenge, because knowledge about the location of the sensors and the response of reinforced concrete as per the design of the bridge can be involved. Furthermore, the effects of a single vehicle are distributed over time (4 seconds or more), and multiple vehicles may show interesting interactions in the data. Although existing algorithms for the analysis of such time-series may go a long way, new and domain-specific algorithms will probably be developed during this project.