Template for MATLAB EXPO 2019 · What Type of Question? 10 For each (trip, day, serial #, customer,...
45
© 2018 The MathWorks, Inc. Machine Learning Proven Applications and New Features SHAYONI DUTTA Senior Application Engineer
Transcript of Template for MATLAB EXPO 2019 · What Type of Question? 10 For each (trip, day, serial #, customer,...
© 2018 The MathWorks, Inc.
Machine LearningProven Applications and New Features
SHAYONI DUTTA
Senior Application Engineer
How to Get Started with Machine Learning?
2
Machine Learning Success Stories
Kinesis Health Technologies
Predicting a patient’s fall risk with
machine learning.
3
Machine Learning+
X4
Machine Learning+
Industry Knowledge Application Knowledge
Your Own Expertise
5
Examples of Successful Machine Learning Applications
Fleet Data Analytics
Energy Forecasting
Manufacturing Analytics
6
New Capabilities
▪ MATLAB apps
▪ AutoML
▪ Signal Processing with
Machine Learning
▪ C/C++ Code Generation
Examples of Successful Machine Learning Applications
Fleet Data Analytics
Energy Forecasting
Manufacturing Analytics
7
Fleet Data Analytics
8
Design Decisions
Test Plans
0.15
0.0
7
Temperature
0 50 100 150 200 250 300 350
Temperature (oF)
0
2
4
6
8
10
12
14
16
Cou
nt
What Level of Data?
Equipment
Signals
Messages
Trip/Session
Time – Value pairs
9
What Type of Question?
10
For each (trip, day, serial #, customer,
etc) in the fleet data set, calculate
some Key Performance Indicator
(KPI*) given parameters XYZ".
Across All (data) in the fleet data set,
calculate descriptive statistics of
specific variables (min, max, median,
count, etc.) to summarize and
visualize (histograms).
Question Type “Across All”“For Each”
Scale to Large Collections of Data with Datastore
11
Available Datastores
General datastore
spreadsheetDatastore
tabularTextDatastore
fileDatastore
Database databaseDatastore
Image imageDatastore
denoisingImageDatastore
randomPatchExtractionDatastore
pixelLabelDatastore
augmentedImageDatastore
Audio audioDatastore
Predictive
Maintenance
fileEnsembleDatastore
simulationEnsembleDatastore
Simulink SimulationDatastore
Automotive mdfDatastore
Custom subclass matlab.io.Datastore
Transformed transform an existing datastore
Performing “Across All” Calculations with Tall
12
▪ Visualizations
▪ Data preprocessing
▪ Machine Learning
Exploring Fleet Data with Unsupervised Learning
13
Unsupervised Learning for Operational Mode Clustering
14
“Cold Storage” “Hot Storage”
Data
Historic data:
• Batch processing
• Large data on cluster
• Explore long term trends
• Build models
Streaming data:
• Near real-time
• Test and implement model
for new data
• Stream processing
Vehicle data, driver
profiles
Deploying Fleet Analytics
15
Fleet Analytics Streaming Architecture
16
Fleet Analytics in Practice: Volkswagen Data Lab
Develop technology building block for tailoring
car features and services to individual
▪ Driver and Fleet Safety
▪ Driver Coaching
▪ Driver-Specific Insurance
Data sources
▪ Logged CAN bus data and travel record
Results
▪ Proof-of-concept model for “telematic fingerprint”
▪ Basis for the “pay-as-you-drive” concept
Source: “Connected Car – Fahrererkennung mit MATLAB“
Julia Fumbarev, Volkswagen Data Lab
MATLAB EXPO Germany, June 27, 2017, Munich Germany
17
Machine Learning + X
18
Fleet Analytics
Equipment Expertise
Design Specs
Operating Modes
Operating Conditions
Machine Learning
Statistical Analysis
Unsupervised Learning
Energy Forecasting
Electrical Grid Expertise
Seasonality
Weather Effects
Generator Characteristics
Machine Learning
Time Series Modeling
Regression
Manufacturing Analytics
Manufacturing Expertise
Process Equipment
Process Variables
Performance Metrics
Machine Learning
Anomaly Detection
Regression
Classification
Examples of Successful Machine Learning Applications
Fleet Data Analytics
Energy Forecasting
Manufacturing Analytics
19
The Need for Energy Forecasts
20
Jul 01 Jul 02 Jul 03 Jul 04 Jul 05 Jul 06 Jul 07 Jul 08 Jul 09 Jul 10 Jul 11 Jul 12 Jul 13
Time 2019
0
1
2
3
4
5
6
7
8
9
10
Win
d S
pe
ed
(m
/s)
Wind
Aug 04 Aug 05 Aug 06 Aug 07 Aug 08 Aug 09 Aug 10 Aug 11
Time 2019
2000
2500
3000
3500
4000
4500
5000
5500
6000
6500
Lo
ad
(M
W)
Demand
Price
Jun 01 Jun 02 Jun 03 Jun 04 Jun 05 Jun 06 Jun 07 Jun 08 Jun 09 Jun 10 Jun 11 Jun 12
Time of day 2018
0
50
100
150
200
250
En
erg
y p
rodu
ctio
n (
kW
)
Solar
How Energy Forecasting Works
21
Electricity Demand
Weather
Electricity Prices
Combine
Historical Data
Preprocess Features
Machine
Learning
Jul 01 Jul 02 Jul 03 Jul 04 Jul 05 Jul 06 Jul 07 Jul 08 Jul 09 Jul 10 Jul 11 Jul 12 Jul 13
Time 2019
0
1
2
3
4
5
6
7
8
9
10
Win
d S
pe
ed
(m
/s)
Aug 04 Aug 05 Aug 06 Aug 07 Aug 08 Aug 09 Aug 10 Aug 11
Time 2019
2000
2500
3000
3500
4000
4500
5000
5500
6000
6500
Lo
ad
(M
W)
load
temp
wind
daymonth
24hr
1week
Building Forecast Models with Regression Techniques
22
Using Energy Forecasting Models
23
New Data
Electricity Demand
Weather
Electricity Prices
Jul 01 Jul 02 Jul 03 Jul 04 Jul 05 Jul 06 Jul 07 Jul 08 Jul 09 Jul 10 Jul 11 Jul 12 Jul 13
Time 2019
0
1
2
3
4
5
6
7
8
9
10
Win
d S
pe
ed
(m
/s)
Aug 04 Aug 05 Aug 06 Aug 07 Aug 08 Aug 09 Aug 10 Aug 11
Time 2019
2000
2500
3000
3500
4000
4500
5000
5500
6000
6500
Lo
ad
(M
W)
Combine Features Trained
Machine
Learning
Model
Forecast
load
temp
wind
daymonth
24hr
1week
Deploying Energy Forecasts
Dashboards for operators
and traders
24
API for App Developers
ChallengeMaximize margins in energy trading by predicting
available supply and peak demand
SolutionUse MATLAB to build and optimize models
that incorporate historical data, weather forecasts,
and regulatory rules
Results▪ Response time reduced by months
▪ Productivity doubled
▪ Program maintenance simplified
“Because we need to rapidly respond to shifting production
constraints and changing demands, we cannot depend on
closed or proprietary solutions. With MathWorks tools we get
more accurate results — and we have the flexibility to develop,
update, and optimize our models in response to changing
needs.”
- Angel Caballero, Gas Natural FenosaLink to user story
Portomouros hydroelectric dam.
Energy Forecasting in Practice: Naturgy Energy Group S.A.
26
Machine Learning + X
27
Fleet Analytics
Equipment Expertise
Design Specs
Operating Modes
Operating Conditions
Machine Learning
Statistical Analysis
Unsupervised Learning
Energy Forecasting
Electrical Grid Expertise
Seasonality
Weather Effects
Generator Characteristics
Machine Learning
Time Series Modeling
Regression
Manufacturing Analytics
Manufacturing Expertise
Process Equipment
Process Variables
Performance Metrics
Machine Learning
Anomaly Detection
Regression
Classification
Machine Learning apps
▪ Try out many models
▪ Compare Results
▪ Get to a reasonable model
without worrying about the
details
28
Perform
Hyperparameter
Optimization in apps
AutoML
▪ Build many machine learning models
▪ Find a good model without becoming
an expert
29
Preprocess
DataDeploy &
Integrate
Train
Model
Extract
Features
Import
Data
Wavelet
Scattering
Decision Tree?
SVM?
KNN?
Ensemble?
…?
fitcauto
Wavelet
Scattering
Hyper-
parameter
Optimization
Feature
Selection
Model
Selection
Wavelet
Scattering
Examples of Successful Machine Learning Applications
Fleet Data Analytics
Energy Forecasting
Manufacturing Analytics
31
What is Manufacturing Analytics?
Definition: Apply modeling (AI) to process and sensor data to maximize operational
performance
Key Use Cases:
1. Automate the monitoring of manufacturing process
2. Ensure product quality
3. Optimize yield of complex production processes
32
Challenges in Applying AI to Manufacturing
Lots of Data – much in “Data Historians” (SCADA, LIMS, OSISoft PI)
Reliable measurements or modeling
– Sensor failures
– Hidden variables
Use of many different tools
– Limited Predictive modeling
– Handle streaming data
– Customization
33
Uncover Hidden Variables with Process Modeling
34
pretty big →
Case Study: Anomaly Detection
35
Case Study: Anomaly Detection
2. One-class SVM1. Cluster with DBSCAN
36
Deployment
Integration with Data Historians
▪ OPC Toolbox (Database tbx via ODBC
or JDBC) connects with PI Server
▪ Access MATLAB analytics within the
PI Asset Framework
Customize Analytics
Delivery
▪ Accessing insights via
GUI critical for plant staff
and process engineers
▪ Build a custom dashboard
with App Designer
37
PI Data
Archive
OPC, ODBC,
JDBC
MATLAB Production Server
Calls
MATLAB
function
PI Asset
Framework
Write
result to
PI point
PI System
Explorer
Case Study:
Application: “Virtual sensor” for accurate
prediction of ore bin levels
Objectives:
– Reduce downtimes
– Maximize throughput
Approach:
– Sensor Fusion using Kalman filters
– Sys ID outperformed traditional models
Results: 5% prediction accuracy over 3 hr time horizon
– Reduced downtime, saving $100k for each sensor failure
– Integration with OSI Soft and Azure IoT+CI Do NOT use with customers until
user story published
38
Machine Learning + X
39
Fleet Analytics
Equipment Expertise
Design Specs
Operating Modes
Operating Conditions
Machine Learning
Statistical Analysis
Unsupervised Learning
Energy Forecasting
Electrical Grid Expertise
Seasonality
Weather Effects
Generator Characteristics
Machine Learning
Time Series Modeling
Regression
Manufacturing Analytics
Manufacturing Expertise
Process Equipment
Variables & Set Points
Parameter Impact
Machine Learning
Anomaly Detection
Regression
Multivariate Statistics
Machine Learning + Signal Processing
40
Data Preprocessing Feature Engineering
Frequency domain
Time domain
Bandwidth measurements Spectral statisticsDetrending Smoothing
Resampling Filtering
Kinesis Health Technologies
Predicting a patient’s fall risk with
machine learning.
41
From Desktop to Production
Reasons for Updates:
▪ Found a better model
▪ New data became available
▪ Business needs change
▪ …
42
C/C++
Automatic C/C++ Code Generation
1. Prediction for most Classification and Regression models
2. Update deployed models without regenerating code
– SVM, Decision Trees, Linear Models
1. Fixed-Point support
– SVM, Decision Trees, Ensemble of Trees
– Shallow Neural Network (through Simulink)
1. Integrate with Simulink models as
MATLAB Function Block
43
Integrate MATLAB
with Other Languages
Examples of Successful Machine Learning Applications
Fleet Data Analytics
Energy Forecasting
Manufacturing Analytics
44
New Capabilities
▪ MATLAB apps
▪ AutoML
▪ Signal Processing with Machine Learning
▪ C/C++ Code Generation
Machine Learning+
XFleet Data Analytics
Energy Forecasting
Manufacturing Analytics
Signal Processing
Industry Knowledge
Application Knowledge
Medical Devices Mining
AppsC/C++ Code
GenerationAutoML
45
Learn More
- Exploratory Data Analysis- Data Processing and Feature Engineering- Predictive Modeling and Machine Learning- Data Science Project
Training Courses
MATLAB Fundamentals (3 days)
MATLAB for Data Processing and
Visualization (1 day)
Processing Big Data with MATLAB (1 day)
Statistical Methods in MATLAB (2 days)
Machine Learning with MATLAB (2 days)
Signal Preprocessing and Feature
Extraction with MATLAB (1 day)
Deep Learning with MATLAB (2 days)
Accelerating and Parallelizing MATLAB
Code (2 days)
46
