Manufacturing Analytics at Scale

G1/PJ-DM | 7/17/2015 | © 2015 Robert Bosch LLC and affiliates. All rights reserved.
Manufacturing analytics at scale
1
Soundar Srinivasan
Bosch Data Mining Services and Solutions, Palo Alto, CA
Jeff Thompson, Ruobing Chen, Juergen Heit, Dirk Slama

Bosch@Data Science Summit, 2015
2
Outline
 Bosch overview
 Core business sectors
 World class manufacturing
 Data mining at Bosch
 Successful applications in manufacturing
 Unique challenges encountered
 Need for further research

3
2014 key figures
 One of the world’s largest suppliers of
automotive technology
Industrial
Technology
Energy and
Building
Technology
Bosch Group  48,9 billion euros in sales
 R&D investment: 4.9 billion euros
 360,000 associates as per April 1.15*
Mobility
Solutions
 Leading in drive and control technology,
packaging, and process technology
 Leading manufacturer of security
technology
 Global market leader of energie-efficent
heating products and hot-water solutions
Consumer
Goods
 Leading supplier of power tools
and accessories
 Leading supplier of household appliances
68%
share
of sales
*including BSH Hausgeräte GmbH (formerly BSH Bosch und Siemens Hausgeräte GmbH)
and Robert Bosch Automotive Steering GmbH (formerly ZF Lenksysteme GmbH).
32%
share
of sales

4
Four business sectors: A diverse, and rich field for data science applications
Mobility
Solutions
Industrial
Technology
Energy and
Building
Technology
Consumer
Goods

5
 Freely
programmable
control units
 PLC and PC based
control units
 Field bus (ethernet-
based)
 Flexible production
systems
 Digital data storage
 Usage of internet
standards
 Integrated IP-
connection
 Identifiable and
communicating
objects
 Mobile operation
 Scalable systems
(cloud as
storage, ..)
 Self-optimising
systems
 Internet-of-things
Advanced manufacturing: The next industrial revolution
Industry 1.0
2. industrial
revolution
3. industrial
revolution
4. industrial
revolution
Industry 2.0 Industry 3.0
 Mech. control (cam
disc, cam)
 Energy: water /
steam power
 Punch cards as
program memory
 Conveyer belts
 Master shafts
 Energy: electrical
1. industrial
revolution
Mechanisation
Electrification
Digitalization
Connectivity and Traceability
The transformation of Industry 3.0 to Industry 4.0 (advanced manufacturing) occurs gradually
Industry 4.0

6
Two perspectives of Bosch in advanced manufacturing
Technology and solution supplier
for OEMs and end users
LEAD PROVIDER
System manufacturer view /
production resource view
LEAD OPERATOR
Product manufacturer view /
product view
First mover in the realisation of integrated
concepts with equipment providers
Big Data
Business
processesDecentralised
intelligence
Machine
models
Software
Added value networks
Connection
Production
models

7
As of 12/2014
200+
Manufacturing
facilities
1000s
of assembly
lines
Billions
Of parts
manufactured
each year
Bosch manufacturing

8
Manufacturing use cases
Test and Calibration Time Reduction
 Prediction of test results
 Prediction of calibration parameters
Quality Improvement
 Descriptive analytics for root-cause analysis
 Early prediction from process parameters
 Self-optimizing assembly line
Warranty Cost Reduction
Prediction of field failures from
 Test and process data
 Cross-value stream analysis
Yield Improvement
 Benchmark analysis across lines and plants
 Pin-point possible root causes for
performance bottlenecks (OEE, cycle time)
Predictive Maintenance
 Identify top failure causes
 Predict component failures to avoid
unscheduled machine down-times

9
Business Objective:
Reduce test and calibration time in the
production of mobile hydraulic pumps
Impact
Example: Test time reduction
35% reduction in test and calibration time
via accurate prediction of calibration and test
results

10
Example: Test time reduction
Problem:
Bottleneck Test Benches
Approach:
1) Identify candidate tests for removal
2) Identify test ‘groups’ run in parallel
3) Use feature selection methods (group
Lasso) to identify least important test
measurements.
4) Remove least important test
measurements (saving test time)
5) Train a predictive model to predict test
outcome from remaining measurements.
Layout of the assembly line

11
Scalable Group Lasso
min
𝛽
1
𝑛
𝑖=1
𝑛
log 1 + 𝑒𝑥𝑝 −𝑦𝑖 𝛽0 +
𝑔=1
𝐺
𝑋𝑖,𝑔 𝛽𝑔 + 𝜆
𝑔=1
𝐺
𝑤𝑔 𝛽𝑔
• We used Limited-memory BFGS (L-BFGS) with Block Coordinate Descent (BCD)
to solve the optimization problem.
• L-BFGS is used to obtain a quadratic approximation of the logistic regression.
• BCD is used to solve the resulting sub-problem, i.e., a quadratic problem with
group penalty.

12 G1/PJ-DM | 7/17/2015 | © 2015 Robert Bosch LLC and affiliates. All rights reserved.
Scalable Group Lasso
 Three parts of the approach can be distributed
 Gradient computation of the logistic function
 Storage of L-BFGS history
 BCD sub-problem solver: minimize each block simultaneously
 When to and why distribute?
 Distributing the gradient computation is beneficial when sample size is large
 Distributing the storage of L-BFGS history is beneficial when there are a lot of features
Chen et al., (NIPS 2014) show that this distributed version is advantageous over the
original only when the number of feature is larger than 10Mil.
 Distributing BCD is beneficial only when the number of groups is large

13
Analytics and production environment
 Device Management
 Device Abstraction
 Event Management
 Software Provisioning
 Identity Management
Production Env.Analytics Environment
Hadoop
MongoDB
DB Connectors
Custom Scripts
SAS
IBM SPSS
Statistica
Alpine
KNIME
Revolution R
RapidMiner
Extraction, Trans-
formation, Loading
Aggregate
Data
Historic Training Data
Analytics,
Machine Learning
Descriptive Analysis
Predictive Model
Extraction,
Transformation
Predictive Model
Prognosis,
Decision (-Support)
Sales
Data
Production
Data
Warranty
Data
Device
Data

Challenges in predicting defects in manufacturing
14
 Large, but distributed data
 E.g. One product variant in one plant
 ~15 million units, 29 data sources, 17
TB data, 22 billion measurements
 High dimensional
 100s-1000s typical
 Schema- and dictionary-migration over time
 Near real-time and resource-constrained
deployment

Other data science challenges in manufacturing
 Data is short-term stationary
 Time and feature correlation
 Label noise
 Very low (but costly) incidence rates
 0-few ppm typical
 Unequal costs of false alarms and
false negatives
 High accuracy and quality
requirements

Need for expanding research in manufacturing
 IEEE Big Data for Advanced Manufacturing Workshop
2015 IEEE International Conference on Big Data
Oct 29 – Nov 01 2015 @Santa Clara, USA
http://ieeebdam15.stanford.edu/

Bosch@Data Science Summit, 2015 Backup
17
Advanced manufacturing
App Store/Digital Services (2)
Connected
Products (1)
3D
Printing
Next-Gen.
Robots
Intelligent
Powertools
Top floor
Shopfloor
End-2-End
Digital
Engineering
Sales/Marketing
& Business Models
Product
Customization (5)
Product Usage Data (3)
Batch-Size One (7)
Work Environment
IoT-Enabled
Manufacturing
CPS
De-Coupling,
Product Memory
Servitization (4)
(9)
(6)
(8)
IoT Service
Implementation
Embedded | Cloud (10)
IoT Service
Operation
Adaptive
Logistics
Aftermarket
Services
Remote Monitoring,
Predictive Maint. (12)
Source:
www.enterprise-iot.org

Manufacturing Analytics at Scale

More Related Content

What's hot

Viewers also liked

Similar to Manufacturing Analytics at Scale

More from Turi, Inc.

Recently uploaded

Manufacturing Analytics at Scale

Editor's Notes