direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Es gibt keine deutsche Übersetzung dieser Webseite.

BA: Anomaly Detection for IoT-based Water Network Monitoring

Berlin and other cities around the globe operate hundreds of thousands of kilometers of water networks to provide reliable public services. Most of them were built over the course of decades while the corresponding urban areas have been rapidly growing and changing in the meantime. Additionally, climate change increases the likelihood of extreme weather events which exert significant pressure on infrastructure. Realtime information about the state of the system can be used to maintain and control the network more efficiently at a fraction of the cost of extending the physical infrastructure. However, there are very few commercial monitoring solutions for this domain and the existing ones are quite expensive for large scale deployment. Therefore, the goal of the WaterGridSense 4.0 [1] project is to develop a low-cost scalable IoT sensor platform for water network monitoring based on modern communication standards (LoRaWAN, IEEE 802.15.4, etc.) and scalable data processing (Kafka, Flink, Cassandra). The architecture consists of different kinds of sensor nodes that transmit measurements via smart city infrastructure to an analytics backend developed at CIT. One of the uses cases of such a platform is to use water level measurements from street inlets to estimate the degree of filter clogging for predictive maintenance. Another use case is to combine rain meter data with stationary water level measurements from across the city to predict and warn about imminent combined sewer overflows (CSO). Finally, dedicated measurement campaigns using floating sensors are used to measure water quality and detect illegal inflows and pipe damages.

In the current phase of the project, work at CIT focuses on the algorithmic aspects of the analytics engine. We already conceived and deployed a scalable and fault-tolerant distributed compute cluster based on kafka, zookeeper, flink and cassandra. The next goal is to identify, implement and evaluate several state of the art methods for anomaly detection and time series prediction that are suitable for the use cases at hand. Currently considered approaches include clustering, identity function learning, Autoencoders, spectral methods, and Dynamic Bayesian Networks. Special consideration should be given to the various domain specific challenges, such as missing or erroneous sensor readings, heterogenous measurement units, and spatiotemporal correlations in the data. For model training and testing we have a small real-world dataset provided by one of our partners as well as a custom simulation for generating synthetic data based on a simple model of the target environment.

Concrete theses in this area may focus on the following topics: Anomaly detection, failure prediction, (unsupervised) machine learning, and time series analysis. In general, a thesis should entail identifying a suitable method, implementing the corresponding flink module, and experimentally evaluating the module using one of our commodity clusters.

If this sounds interesting to you, please send an email to felix.lorenz@tu-berlin.de [2] with a little bit of background information on yourself, so we can quickly identify a fitting thesis topic together.


Felix Lorenz
+49 30 314-27998
Raum TEL 1211
E-Mail-Anfrage [3]
------ Links: ------

Zusatzinformationen / Extras


Schnellnavigation zur Seite über Nummerneingabe

Copyright TU Berlin 2008