In-Network Data Analysis of spatially distributed quantities

• Applicants: Dr.-Ing. Reiner Jedermann; Dr.-Ing. Henning Paul
• Subject Area: Electronic Semiconductors, Components and Circuits, Integrated Systems, Sensor Technology, Theoretical Electrical Engineering
• Term: from 2014 to 2017
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 253282643

A typical application field for the estimation of spatially distributed physical quantities is environmental monitoring, i.e., the measurement of, for example, temperature and humidity or the concentration of chemicals in air or water.In spatial supervision tasks, measurements are usually taken by wireless sensor nodes. Forwarding of all sensed information to a central data fusion unit leads to high communication volume and shorter battery life time. Additionally, the failure of the central node causes an outage of monitoring, rendering the system unable to react to an event. Thus, it is reasonable to implement the monitoring using a network auf equal, autonomous sensor nodes, resulting in a high robustness against failure of nodes and inter-node links. For the distributed processing, the application of iterative In-Network Processing (INP) algorithms is intended.In order to keep the communication effort by iterative processing low, we peruse the approach to represent the observed quantities by a parametric model and thus yield a compressed representation of the spatially distributed quantity. In combination with INP-based distributed estimation algorithms, it is feasible to reduce communication to an exchange of estimated model parameters. The main objective of this proposal is the combination of parametric modelling with INP distributed over the sensor nodes. Existing algorithms for distributed parameter estimation will be evaluated and novel ones developed. A particular challenge is the non-convex distributed optimization resulting from the application of non-linear models. Furthermore, possible advantages of interlinking the network nodes in bi-level structures will be analyzed with regard to communication efficiency. Additionally, the effect of time variations of the observed quantities and of the network topology will be considered with the goal of developing adaptive algorithms. Another substantial part of the project is concerned with the detection and classification of measurement errors and signal deviations caused by physical events. The application of a model based approach in combination with In-Network-Processing for parameter estimation is expected to improve the accuracy of fault and event detection. Local measurement cannot only be compared with the measurement of direct neighbor nodes, but their concurrence can be tested against the predicted spatial distribution of the quantity based on information form the whole network.The developed models and algorithms will be tested on physical sensor node hardware on the basis of recorded data from real-world scenarios. It has to be verified that typical sensor nodes with limited recourses with regard to memory, CPU speed, and energy are capable of executing complex estimation algorithms. Furthermore, the tests should prove that the developed models provide a sufficiently high data compression for typical application examples in environmental monitoring.

• DFG Programme: Research Grants