Topology in the furnace: TDA as a diagnostics tool for process control systems
Event details
| Date | 19.05.2016 |
| Hour | 14:15 › 15:30 |
| Speaker |
Dr. Mikael Vejdemo-Johansson, KTH Royal Institute of Technology, Stockholm Bio: I am a postdoc specializing in applied algebraic topology. My research is into topological data analysis, partially of data sets from robotics. I have previously been a postdoc in the applied topology group at Stanford, and in the computer algebra group at St Andrews. |
| Location |
CO 011
|
| Category | Conferences - Seminars |
Steel smelting is a high-volume, high-throughput industry, where the smallest performance gains translate into large dividends. Model construction to predict conditions inside the furnace is a centrally important part of process control. Machine learning and statistical methods have been shown to improve on purely metallurgical models, but in either case, the failure modes of the model are poorly understood, and tools for analyzing them not well developed.
We work in collaboration with Outukumppu Stainless with their electric-arc scrap furnace, to analyze and improve their temperature prediction models. Temperature prediction in particular is an important model to improve: reference measurements can be done by inserting probes, but these are costly and if measuring too early, more probes will be needed — measuring too late risks overheating the steel and spoiling the entire batch.
Based on work and ideas from Anthony Bak and Ayasdi, and in collaboration with Ayasdi, we are studying the use of the Mapper algorithm to construct intrinsic models of the fibres (preimages) of failed predictions. These models help classify different modes of failure for the models, and direct attention for improvement or for learning compensation transforms to improve precision of temperature detection.
In this talk, I will describe the approach we take for modeling and classifying failure modes, and give some examples from our ongoing study of the steel smelting data. The talk will assume no previous knowledge of topological data analysis, and will explain Mapper completely and accessibly.
We work in collaboration with Outukumppu Stainless with their electric-arc scrap furnace, to analyze and improve their temperature prediction models. Temperature prediction in particular is an important model to improve: reference measurements can be done by inserting probes, but these are costly and if measuring too early, more probes will be needed — measuring too late risks overheating the steel and spoiling the entire batch.
Based on work and ideas from Anthony Bak and Ayasdi, and in collaboration with Ayasdi, we are studying the use of the Mapper algorithm to construct intrinsic models of the fibres (preimages) of failed predictions. These models help classify different modes of failure for the models, and direct attention for improvement or for learning compensation transforms to improve precision of temperature detection.
In this talk, I will describe the approach we take for modeling and classifying failure modes, and give some examples from our ongoing study of the steel smelting data. The talk will assume no previous knowledge of topological data analysis, and will explain Mapper completely and accessibly.
Practical information
- Informed public
- Free
Organizer
- Kathryn Hess