140

# GABC2018: Big data, small world: AzureML@Geberit by Christian Hidber

Our machine learning solution stems from roof drainage systems. It warrants that large buildings like stadiums or shopping malls do not collapse during heavy rainfalls. While the hydraulics form a system with given pipe diameters can readily be simulated, only heuristic algorithms are known to figure out the diameters. A brute force approach is not feasible, since a large system ends up with trillions of possibilities.

April 21, 2018

## Transcript

Hidber

4. ### The challenges 23.04.2018 AzureML@Geberit manage the water level on the

roof control & steer the water flow find the right dimensions save & reliable

8. ### What if…. 23.04.2018 AzureML@Geberit • Collapsing pipes • Collapsing roofs

• Clogged pipes • Façade damages
9. ### The problem of the «right» dimensions 4/23/2018 AzureML@Geberit • Finding

the «right» diameter is difficult • Validating is «easy» (hydraulic simulation) • Only heuristic algorithms are known => Calculation as an ML-based assistant
10. ### Planning process (conventional) 76% (Q1/2015) 1 2 X X 24%

(Q1/2015) 23.04.2018 3
11. ### Planning process (towards ML) 23.04.2018 Features (input) Labels (output) Set

diameters on neuralgic pipes Predict diameters for neuralgic pipes
12. ### Planning process (towards ML) 23.04.2018 Features (input) Labels (output) Set

diameters on neuralgic pipes Predict diameters for neuralgic pipes
13. ### Planning process (ML based) 93% (Q3/2017) 23.04.2018 Set diameters on

neuralgic pipes Predict diameters for neuralgic pipes 1 2

15. ### Implementation in Azure ML (demo) • Team • Data •

Algorithms • Training • Deployment 23.04.2018
16. ### Challenge 1: Data distillation 23.04.2018 Problem Training takes too long

Solution Train on distilled data Learning More is not always better total: 2’500’000 calculations 600’000 150’000 35’000 difficult real different • The «difficult» ones are sufficent • CPU load
17. ### Challenge 2: Big data – small world 4/23/2018 Problem error

rate too high «Image" licensed according to CC BY-SA Approach: more data, more training
18. ### Challenge 2: Big data – small world (continued) 23.04.2018 Problem

Error rate too high Solution Change the problem Learning Look for a different problem formulation predicting diameters predicting direction +–= 60% 72%
19. ### Challenge 3: model improvement (perfect algorithms) 23.04.2018 Problem Would better

algorithms / params help ? Accuracy 72 %
20. ### Challenge 3: model improvement (perfect algorithms) 23.04.2018 Problem Would better

algorithms / params help ? Solution Ceiling analysis Simulated accuracy 100 % +1% "Dieses Foto" von Unbekannter Autor ist lizenziert gemäß CC BY Learning Small improvement only by retraining with better algorithms / params
21. ### Solution Architecture 23.04.2018 AzureML@Geberit Azure Worker Role Blob Storage Queue

ML Web Role 7000+ Clients, 40+ Countries, 17 languages
22. ### Summary • Azure ML based solution increases success rate to

93% • Ceiling analysis safes a lot of time • Azure ML easy, but problem difficult • Team-structure shaped the approach • Calculation “as a service” 23.04.2018 AzureML@Geberit About Geberit The globally operating Geberit Group is a European leader in the field of sanitary products. Geberit operates as an integrated group with a very strong local presence in most European countries, providing unique added value when it comes to sanitary technology and bathroom ceramics. The production network encompasses more than 30 production facilities, of which six are located overseas. The Group is headquartered in Rapperswil-Jona, Switzerland. With around 12,000 employees in around 50 countries, Geberit generated net sales of CHF 2.8 billion in 2016. The Geberit shares are listed on the SIX Swiss Exchange and since 2012, have been included in the SMI (Swiss Market Index).