How We Learned to Predict Failures

Hello, Habr! This is Olga Peshina again, an expert on the development of new technologies at Severstal-Infocom JSC. While pumping over the telemetry of our giant metallurgical plant, we want to operate the data obtained from the units not only in real time (β€œsomething has broken, needs to be repaired”), but also to build a model for predictive analytics of equipment failures (β€œsoon there will be a problem, it is necessary take action in advance ”).





I will tell you how we got our hands on the ball trying to implement predictive repairs, what we had planned and what we didn’t, and why.





It sounds magical - always know at what moment each roll or bearing will fail, schedule maintenance according to condition, minimize overservice and stock of spare parts, go to zero on emergency stops.





Any new technology moves through a cycle of hype: excessive expectations give way to deep disappointment, after which comes the realization of real opportunities and limitations. In 2017-2020, we went through this path too. 





The hype cycle.  Find Bitcoin
The hype cycle. Find Bitcoin

As we saw it: the trigger of innovation

2014 , (), , . , , – .





500 . 12 000, 2025 100 000. , , 50 .





Vibration and RPM sensor on the engine
Roll vibration sensor

2017 Data Lake , . SAP ERP ; , .





: , . .





, . , : , . , , .





Roll kits

, , , , .





, . , 24/7 : . . , – - – .





2017-2018 . , , , , , .





: , , , , , . , , .





, – , , . , : , , , .





. , , ? , , . - , , ? , , . – , . , . ?





, , . – -2000. , . – .





Stan-2000.  Stop cannot be repaired
-2000.

. -2000 1800 . – (). 





( ) . – .





Finishing stands

. , , , . 





, , . : MVP . , , , . 





- , - . : , - . 2018 . , … .





. : , , – .





, . .





-

- -, Data Lake. .





. : – , . – – . – , . . , - .





: – , . , 50%. , : , – . Β« Β».





, – . - : . , , , , , … , , - !





: , . – . , . . , . , : Β« , , , … , Β».





, (, , ) . : .





, … , . , .





, , , . . , .





Data from thousands of sensors flow to the monitoring center.
.

:

, . , 50%, 3 5-6 1,5 . - – , , .





. - Gerdau , SmartSignal, General Electric, .





650 , . – . , SmartSignal , : . , .





– , . , , : , , , . . , - .





Β« , – Β». . . , .





, , , . , – , , . : – , , , .





SmartSignal - , , . . . , – , .





SmartSignal , . .





General Electric , : , , . . , . .





, , Β« Β». , - , , … - , .





, , . : , , . 





, , , . .





, . , , . , .





SmartSignal (). 





UNRS.  It is she who monthly generates 50GB of data about her condition
. 50

MVP . : , . 





. , SmartSignal 5-7 , . 20 , . . 





– . – , – .





Β« Β», - 80% . – . – . .





, . , .





: GE . . , – . - .





, - , - , - . , , . : , ?








All Articles