Demystifying Cognitive Approaches to Predictive Maintenance: Part1 DataRPM webinar
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Demystifying Cognitive Approaches to Predictive Maintenance: Part1
DataRPM webinar
DataRPM webinar
Demystifying Cognitive Approaches to Predictive Maintenance: Part1
Anita RajAditya Murukutla
How to solve the current gaps in machine data analysis
How you can build a scalable and repeatable approach to harness the trillions of data points generated by sensors
How cognitive approaches to data science can be a game-changer for Industry 4.0
Use Case: Enabling cognitive predictive maintenance for a Fortune 10 Industrial Manufacturer
1
2
3
4
Anita RajPrincipal Growth Hacker | DataRPM
Aditya MurukutlaSenior Product Manager | DataRPM
@aditya_narayana [email protected]
HOST
AGENDA
SPEAKER
Welcome!Welcome!This webinar is best appreciated if you are associated with:
You want to do Predictive Maintenance on Industrial AssetsYou deal with large
amounts of Sensor Data
Industrial asset-centric organizations Industrial asset-intensive organizations
The Gap in the Current State of IIoT Data Analysis
Low compute power restricts the size of data to be analyzed
Scalability
Data scientist has to build new models for every new use-case
Repeatability
Models cannot be used in production
systems
Operationalize
KEY DATA SCIENCE CHALLENGES Static, Manual Modeling Fails for the IIOT
© 2017 DataRPM – Proprietary and Confidential 6
Only 20% of Asset Failuresare Common & Predictable
80% of Asset Failures are Random & specific to an Individual Assets
Modeling today takes:
Samples of Data fromSamples of Sensors fromSamples of Assets
Builds Generalized Models & Extrapolates to the entireAsset Population
Leading to Poor Results
Need Individual Predictive Models for every Asset
to get at that Unknown 80%
KEY DATA SCIENCE CHALLENGES For PdM For IIoT
© 2017 DataRPM – Proprietary and Confidential 7
Massive Unlabeled Datain the IIoT
Minimal Labeled Training Data
Thus Supervised-Learning not possible
Use Unsupervised-Learningof Unlabeled Data first to
Generate Labeled Data for Supervised-Learning
Predictionsare only useful ahead of time
Recommendationsonly possible with
timely Predictions
KEY DATA SCIENCE CHALLENGES For PdM For IIoT
© 2017 DataRPM – Proprietary and Confidential 8
Need Adaptive & Online Model Generation for Continuous Learning
Static Models becomes Obsolete really fast
Results in Poor Predictions in actual Operating Environments
Failures are Rare Events& caused byDynamic Changes
What is a Scalable Approach?
DistributedArchitecture
In-Memory Algorithms
Expansion Horizontally
Open Source
Key Takeaway: The technologies and methodologies used will grow to the needs of the projects long-term.
Large distributed Compute platform
Run Machine Learning in Memory to
optimize speed on a distributed Infrastructure
Architected so we can add new nodes on
the fly
All technologies
used are open source
What is a Repeatable Approach?
RepeatableData Product
DataFrame
MLRecipes
API &Interfaces
Ingestion
Consumption
Modeling
Storage
Customization possible with ability to easily do smaller tweaks in interface since all code is open and recipe parameters are modifiableClient has ability to plugin their own algorithms (as long as scala/spark code) into the Recipes
How can you Operationalize ?
PrescriptiveRecommendations
Influencing Factors PredictionSegmentation
Divides the population into groups for analysis
Finds factors in groups that are similar and/or different for an outcome
Builds a model that can classify new data points based on similar factors
Clustering Distributions Rules
Association Rules Frequent Patterns Correlation
Classification Regression
Recommendations for specific segments based on predictions & influencing factors
Collaborative Filtering Content Filtering
What DataRPM offers
Consumption
APIs
Micro Apps Framework
Security
DataManagement
Data Sync
Data Lake
Metadata
Machine Learningand Analytics
Spark Engine
Workflow Builder
Data Science Recipes
Meta Learning
Natural Language
Visualization
IIoT Sensors Data
Enterprise Asset Management Systems
RDBMSData sources
Hadoop
Insights App
Discovery App
Admin App
PdM Apps
F
Use Case|Predictive Maintenance for Industrial IOT
MANUAL ANALYSIS
CHALLENGE2 sensorsat-a-time only
15 minutes of sensor data merely a tiny Data Sample for Model
Predictions of Future Failures & Recommend
Actions
USE CASE
Identify the indicators of failures
for washing machines
Each sensor records multiple
data points in millisec range
75 Unique Sensor
Recordings
0111011100101010110110
DATA OVERLOAD
6+ Months for Team of Data Scientists to process
Resulted in poor prediction with lots of false positives
COGNITIVE DATA SCIENCE
SOLUTION
ALLsensorsin parallel
Monthsof sensor dataused to train Data Models accurately
< 2 DaysHighly Accurate Prediction Model
Automated building of thousands of models in parallel to deliver the optimal model
OTHER PdM EXAMPLES
Predict failure of robotic and other machinery in assembly lines and identify internal and external factors that cause failures
Predict failures of set top boxes and identify the internal and ext ernal causes of failures and recommend actions
FORTUNE10MANUFACTURER
Use Case|Predictive Maintenance for Industrial IOT
© 2017 DataRPM – Proprietary and Confidential 15
IF YOU’RE INTERESTED IN LEARNING MORE:
THANK YOU
Related Documents
