Machine Learning for the Sensored IoT
2 likes1,018 views
The document discusses the role of data science and machine learning in the context of the Internet of Things (IoT), focusing on sensor-generated data. It highlights the importance of signal processing, feature engineering, validation strategies, and the need for scalable tools to handle the vast amounts of data referred to as brontobytes. Key takeaways emphasize treating observations independently and acknowledging the uncertainty inherent in the data.
Machine Learning for the Sensored IoT
- 1. H2O.ai Machine Intelligence Machine Learning for the Sensored Internet of Things Hank Roark hank@h2o.ai @hankroark 1
- 2. H2O.ai Machine Intelligence Who am I? ▪ Data Scientist & Hacker @ H2O.ai ▪ Lecturer in Systems Thinking, University of Illinois at Urbana-Champaign ▪ John Deere, Research, Software Product Development, High Tech Ventures ▪ Lots of time dealing with data off of machines, equipment, satellites, radar, hand sampled, and on. ▪ Geospatial and temporal / time series data almost all from sensors. ▪ Previously at startups and consulting (Red Sky Interactive, Nuforia, NetExplorer, Perot Systems, a few of my own) ▪ Systems Design & Management MIT ▪ Physics Georgia Tech
- 3. H2O.ai Machine Intelligence IoT Data Comes From Lots of Places, Much of it from Sensors
- 4. H2O.ai Machine Intelligence The data is going to be huge, so get ready
- 5. H2O.ai Machine Intelligence Wow, how big is a brontobyte?
- 6. H2O.ai Machine Intelligence This much data will require a fast OODA loop Much of these models will then be used in control systems Image courtesy http://www.telecom-cloud.net/wp-content/uploads/2015/05/Screen-Shot-2015-05-27-at-3.51.47-PM.png
- 7. H2O.ai Machine Intelligence Machine Prognostics Use Case Sensor data of turbofan remaining useful life prediction Jupyter notebook @ http://bit.ly/1OmdBg7 Many more tips and tricks
- 8. H2O.ai Machine Intelligence Key take aways for modeling the sensored IoT • Some sort of signal processing is usually helpful, but can introduce bias • Smoothers, filters, frequency domain, interpolation, LOWESS, ... , aka feature engineering or post-processing • Knowing a little about the physics of the system will be helpful here • Validation strategy is important • Easy to memorize due to autocorrelation • Sometimes the simplest things work • Treat each observation independently; Use time, location, as data elements • Uncertainty is the name of the game • Methods that will report out probabilities are often required (not shown here) • The data can be big, get ready, it'll be a great ride • Scalable tools like H2O will help you model the coming brontobytes of data