profile_document

Smart sensor at the edge presentation. Technology and approach being used by Nexedi during Windelin EU R&D project.

Page Content

Windelin R & D project

Who am I?
EU funded project (2.5M EUR), 2015 - 2018
Participants: Nexedi, MariaDB and Micromega Dynamics (Woelfel group)
Goal: develop a big data solution on top of Wendelin platform for wind turbines management and smart sensor at the edge utilizing GPU and machine learning technology

Smart sensor at the edge architecture

Placed inside wind turbine, based on Nvidia TX2 board
Close to real time computation of an anomaly index and action (shutdown turbine, alarm)
No need of network connectivity or human interventions
How: Machine Learning for failure prediction and anomaly detection using GPU
ML Model built server side, same model runs inside smart sensor - i.e. developed once, used anywhere -> less code and maintenance

Data = usually a set of numbers representing a machine state (wind turbine's state)
Model = "formula" for converting with minimal losses input data to output data where
- Model created by iterating over and over (hundreds of times) over TBs of data using powerful GPU cards (server side)
- Model = very small file, quickly executable on either GPU or CPU into any embedded system
Anomaly = how well input data "fits" into model, the less if fits the higher its anomaly score is which mean "ALARM" state

Industry agnostic approach. Data source abstraction.
Generic sensor which can interface with any device in industry supporting TCP / IP protocol
Machine learning - no more black box and magic but hard-stone mathematics constantly being improved
Machine learning library agnostic approach. Use what you need: tensorflow, pytorch or scikit-learn both at the edge (sensor) or server side (backend)