Developing Prognostics Algorithms:

Data-Based and Model-Based Approaches

Seth DeLand
May 9, 2017
MathWorks Automotive Conference

© 2017 The MathWorks, Inc.

1
What is Prognostics?

….

2
Why Prognostics?

▪ Improved operating efficiency ▪ New revenue streams ▪ Competitive differentiator

Source: GE Oil & Gas Source: Oliver Wyman 2015 MRO Survey

5
How does it work? Prognostics Algorithm Workflow

Access and Develop Predictive Integrate Analytics

Preprocess Data
Explore Data Models with Systems

Business Data Data Reduction/ Model Creation Enterprise Systems

Transformation

Sensor Data Feature Extraction Model Validation Embedded Devices

6
Challenges for Prognostics Development

▪ How long will it take to collect fault data?

▪ How expensive is it to collect?

▪ How complex is the system?

7
Sources of Data for Prognostics Development

Fleet Experiments Simulation

8
Spectrum of Approaches for Prognostics Algorithms

First Principles Modeling Data-Driven Modeling

Physical Models Kalman Filters Curve Fitting Statistical Methods

Techniques for Modeling Dynamic

Systems

9
 Examples

1. Data-based prognostics using machine learning

2. Fault injection and failure analysis using simulation

10
 Examples

1. Data-based prognostics using machine learning

2. Fault injection and failure analysis using simulation

11
Example 1: Data-based Prognostics with Machine Learning

Data provided by NASA PCoE

http://ti.arc.nasa.gov/tech/dash/pcoe/prognostic-data-repository/
12
Different Types of Learning
Type of Learning

Predicting Failures
Supervised
Learning

Develop predictive
model based on both
Machine input and output data
Learning

Unsupervised
Learning

Anomaly Detection
Discover an internal
representation from
input data only

13
Principal Components Analysis – what is it doing?

14
Different Types of Learning
Type of Learning

Predicting Failures
Supervised
Learning

Develop predictive
model based on both
Machine input and output data
Learning

Unsupervised
Learning

Anomaly Detection
Discover an internal
representation from
input data only

15
Data-based Prognostics with Machine Learning - Takeaways

▪ Use machine learning to identify outliers and build predictive models

▪ Many choices for algorithms, apps make it easy to compare options

▪ Workflow-focused tools help you fine-tune the model to your particular data

16
 Examples

1. Data-based prognostics using machine learning

2. Fault injection and failure analysis using simulation

17
Example 2: Fault injection and failure analysis using simulation

18
Run multiple parallel simulations from the parsim command

• Speed up simulations
and simplify workflow
• Simplifies large
simulation runs

19
Leverage Parallel Computing with Simulink

Reduce the total amount of time it takes to…

Run multiple independent simulations (E.g.
Parameter sweeps, Monte Carlo Analysis)

20
Fault injection and failure analysis – Takeaways

▪ Use simulation when measured data is not available

▪ Run what-if analyses to explore scenarios that are difficult to

recreate

▪ Comparing field data to simulation data can help diagnose

cause-of-failure

21
Fleet Data for Prognostics Development
Data Storage
Server

Vehicle Data

22
Deploying Prognostics Algorithms
Data Storage
Server

Vehicle Data
Server-side
prognostics

Onboard prognostics

23
Considerations for System Architecture
Data Storage
Server

“Infinite”
resources on
Connectivity server
is fragile

Some
algorithms
Limited more viable
resources than others
on ECU

24
Considerations for System Architecture
Data Storage
Server

Reasonable Compromise

• Hybrid approach – split between onboard and server-side

• Onboard: buffer, preprocess

• Server-side: classify, take action

• Recommendation: Think about modularity when designing

25
Server-side Prognostics
Data Storage
Server

RESTful
Java
.NET
Python Vehicle Data
C/C++

Package algorithms
for MATLAB
Production Server

26
Onboard Prognostics
Data Storage
Server

Vehicle Data

Generate standalone C/C++

code using Simulink Coder

27
Key Takeaways

▪ No “one-size-fits-all” approach to prognostics.

▪ Prognostics system architecture is evolving.

▪ MATLAB and Simulink provide a platform for developing prognostics

algorithms.

28
 mathworks.com/big-data
Learn More

mathworks.com/machine-learning

Example: Model-based approach

29