Probabilistic Graphs for Sensor Data-driven Modelling of Cyber-Physical-Systems - overview


Abstract

With the increasing availability of sensors, machine learning has the potential to provide more a efficient, robust and scalable approach to modelling of physical systems for the purposes of prediction or anomaly detection.


A solution based on probabilistic graphs is being developed, at IBM Research Ireland, where custom non-linear, localised models of the joint density of subset of system variables, for example based on neural networks, can be combined to model arbitrarily large and complex systems.


The graphical model allows to naturally embed domain knowledge in the form of variables dependency structure or local quantitative relationships. Real-world data (smart grid and smart building projects among others) will be available for evaluation and demonstration.

 

Required skills:

  1. Strong foundations and working experience on machine learning methods for at least one among neural networks, Gaussian graphical models, non-linear latent variable modelling.
    2. Excellent programming skills in Python and working knowledge of machine learning programming frameworks for deep learning (e.g. Tensorflow) .
    3. Interest in working with real data.