04-28, 11:00–11:45 (America/Los_Angeles), St. Helens
This talk will present U-Net-style networks for discrete feature identification in one dimensional time-series data. We will present applications of this technique for identification of pipe joints in oil & gas and water pipeline inspection data, abnormal heart rhythms in EKG signals, and airport runway deflections. This lighthearted, hands-on, talk is for data science practitioners and their immediate supervisors.
This talk will cover how we adapted the U-Net architecture for feature identification in one-dimensional time-series data. We will discuss how 1D segmentation neural nets naturally give rise to point identification via the softmax function and how this can turn problems that are not quite segmentation into segmentation problems. Applications to be discussed include oil and gas pipeline inspection, ideas for thinking about non-standard applications of deep learning (problems that cannot be made to look like exactly like NLP or image processing), challenges and early failed attempts in the development process, and how this strategy applies more broadly than just to the problems we solved here. We will show examples of applications of this strategy for identifying joints in oil/gas/water pipelines from remnant magnetometry, abnormal heart rhythms in ekg signals, and airport runway deflection modalities.
The talk will be light-hearted and informative. There will be practical take-aways for most of the audience who work on homogeneous time series data and there will be some humor for everyone else.
Previous knowledge expected
Michael has been a data scientist with a focus on machine learning at INGU Solutions since January 2021. He has a PhD in Chemical Engineering from the University of Houston. His thesis work focused on protein crystal nucleation precursors and image processing techniques, and it is this expertise that he now applies at INGU.