industrial analytics - bhge s2... · 9 • a live up-to date digital representation of an asset,...
TRANSCRIPT
Confidential. Not to be copied, distributed, or reproduced without prior approval. © 2017 Baker Hughes, a GE company, LLC - All rights reserved.
September 29, 2017
Industrial Analytics: Finding a Needle in an Ocean of Data
Arun SubramaniyanVP Data Science & Analytics, BHGE Digital
Physical streams create large data streams
Confidential. Not to be copied, distributed, or reproduced without prior approval.
September 29, 2017 3
Industrial data volume, velocity is already high… and will increase
Seismic Data100 GB/survey
Drilling Data0.3 GB/well/day
Wireline Data5 GB/well/day
Pipeline Inspection 1.5 TB / 600 km
Process Data6 GB/plant/day
Ultrasound: Tubes1.2 TB/ 8 hrs
ERP Sys. Predix
• ~10-100x more volume• ~100-1000x more velocity
Industrial data requires a fundamentally different approach
Confidential. Not to be copied, distributed, or reproduced without prior approval.
Converting Data to Actionable insights … Not Easy
Lack of Infrastructure
• Connectivity to remote assets
• Ingestion & management of large volumes of data
• Scaling was costly
Historical Practices
• Fragmented, discrete (and sometimes very good) software solutions.
• Highly localized deployment
• Time consuming & costly to move data around
A very limited amount of data generated is captured.
Leads to data & workflow silos.
Analytical Dearth
• Relied heavily on domain experts & brute force methods.
• Asset & data growth outpaces the growth of domain experts. Aging workforce.
Limited amount of insights.
Only 2-3% of data collected is proactively analyzed in Oil & Gas
BHGE is uniquely positioned to address these challenges
Confidential. Not to be copied, distributed, or reproduced without prior approval.
September 29, 2017 5
Confidential. Not to be copied, distributed, or reproduced without prior approval.
Azure AWS GE Private Cloud
Data Fabric
Rapid Query Engine Analytics Engine
• Ingest
• Federated Access• Intelligent Cache
• Curate • Semantic Model • Deep Search
• Explore: Visualize, Build • Learn: Optimize, Recommend• Refine: Ingest code, productize• Connect: Orchestrate & Run
Full Stream Applications
• Light Weight • Cloud First • Edge Deployable
Our technology is a differentiator
PLM SystemsERP SystemsOT SystemsHistorians3rd Party SWMachines
3rd Party Analytics
A new approach for Analytics
6
Traditional Analytics (BI / Point Solutions) BHGE Analytics
Build
DeployUpdate
Idle
Schema definitions
Manual
Custom
Cumbersome
Large data warehouses
Relational databases
Scale
Deploy & maintain
Distributed queries across data silos
Data Fabric
Schema definitions
Augmented modeling & auto-updating
Automated
Guided
SeamlessElastic, distributed
computing
Distributed queries
Automated schema
Elastic computing
Rapid augmented modeling
Seamless
BHGE Analytics Framework delivers a step change for the industry
7
Finding a Needle in an Ocean
Data Science
Domain Knowledge
Software
“Needle in a haystack” Traditional Solutions
BHGE’s play zone
8
Building Enterprise-Grade Industrial Analytics
Approach
Probabilistic Learning System
Machine Learning
Deep Domain Models
A collection of
templates for broad industry
outcomes
Blueprints
Collection of kernels, technique& models for specific outcomes
Templates
Codifies relation between inputs & outputs
Models
Solution methods
Techniques
Building block for models
Kernels
Digital Twins
9
• A live up-to date digital representation of an asset, system or process• Used to predict performance
Physical Asset Digital Twin
Real TimeOperational Data
Maintenance History
Operational History
Fleet Aggregate Data
FMEA
CAD Model
FEA Model
Control Response
Hybrid Models
Physics Based
Probabilistic
Machine Learning + AI
+
+
.
.
✓Continuously Tuned✓Scalable✓Adaptable
BHGE Enables Analytics @ At All Scales
10
Raw Data Machine Learning / AI Hybrid Analytics
ObservationsBillions - Trillions
Rich DataHundreds - Thousands
Business DecisionsTens - Hundreds
Pro
cess
System of Systemse.g. Oilfield
System of Assets
e.g. WellAsset
e.g. Artificial Lift
Hybrid Approach
Inte
gra
ted
S
olu
tio
ns
Probabilistic Learning System
Machine Learning
Deep Domain Models
BHGE’s Unique Hybrid Approach
11
Traditional Analytics BHGE Analytics
PHYSICS TOOLS
MACHINE LEARNING/AI
DATA
PROBABILISTIC LEARNING
HYBRID MODELING
CONTINUOSLY UPDATE MODELS
DATA
ISOLATED DECISIONSReduce flow
Increase operating temperature
Adjust power use
ISOLATED INSIGHTSProbability of failureProduction forecastImage classification
The Physics Advantage: Sparse Data
12
Requires only 1 observation to predict precisely
𝑣
𝜃
Requires 10,000+ observations to predict approximately
𝑥 = 𝑣 cos 𝜃 𝑡
𝑦 = −1
2𝑔𝑡2 + 𝑣𝑠𝑖𝑛 𝜃 𝑡
Data only approach Hybrid Physics approach
The Hybrid Advantage: Overcoming Sparse & Uncertain Data
13
Input Data
Data Analytics
Pre
dic
ted
Le
ngt
h
Actual Length
Data Analytics Only
Hybrid Model: Data + Probabilistic + Physics Models
Data + Estimation Physics of Failure Model
Pre
dic
ted
Le
ngt
h
Actual Length
• Accurate Prediction• 0 False Negatives
• Poor Prediction• No correlation
14
Building a Nonlinear Probabilistic Model with 1 Observation
𝜂
Stress Stress intensityMetal temperature Damage
Digital Twins…Forecast Events with Accuracy
15
Physics Based Models + Machine Learning = Ability to Predict with High Accuracy
Statistical models Physics-based models Machine learning
Heavy Duty Gas Turbine cracking model
• Sparse events typical in industrial settings
• Gas Turbines don’t crack
everyday
• Poor correlation with statistical models
• No ability to forecast
✓ Physics based models capture variation with better accuracy
✓ Reduces false positives and false negatives & computes uncertainty
• Machine Learning is used to estimate missing data. Blue = Collected data Red / Green = Estimated data.
✓Data + Estimated Data + Physics Model →Prediction
Data is not always complete
Time
Time
Time
Da
ma
ge
Da
ma
ge
We invent smarter ways to bring energy to the world.