Deep Machine Learning on Marine Sensor Data

25. April 2017, 14:00 , 15:30

Veranstalter:  Stefan Oehmcke, Universität Oldenburg
Ort:  A06 4-418

Abstract:
Measuring and collecting data of ocean processes is a challenging task. Marine sensors can often only observe one specific location, but to achieve an in depth insight, it is preferable to operate sensors at as many locations as possible. The collected time series data are plagued by noise or shifts due to events such as storm surges, algae blooms, or irregular maintenance. Furthermore, repairs are expensive and often not immediately possi-ble, which is why sensors often malfunction for longer periods of time.


My thesis is focused on applying and extending machine learning algorithms to a fused dataset of the Time Series Station Spiekeroog and a project about “Biodiversity – Ecosystem Functioning across marine and terrestrial ecosystems” (BEFmate). First, I present a framework to preprocess this dataset including shift correction, outlier detection, and missing data imputation. For the multivariate imputation task I combine k-nearest neighbors, standard interpolation methods, and dynamic time warping with a penalty for longer in-tervals of missing data to deliver a new robust method. In the second part, I devise a deep learning architecture with bidirectional recurrent neural networks to compensate for the loss of permanently failed sensors. This network represents a virtual sensor that predicts the output of the lost sensor based on the preprocessed meas-urements of functioning sensors nearby. For this architecture I extend a time dimensionality reduction method to train the recurrent network on the most vital information, while reducing the runtime. Later, I introduce convolutional layers that are usually employed in the vision domain. To achieve a better understanding of the virtual sensor’s output quality I add model uncertainty predictions that also incorporate the known input quality. All methods are empirically evaluated and if applicable compared to state-of-the-art machine learning methods on a set of benchmark problems.


Betreuer: Prof. Dr. Oliver Kramer