Improve operations and decision-making by using real-time data insights and interactive analytics to accelerate IoT data use throughout your organization.
Discover the webinar here: https://bit.ly/38sMcrP
3. This session
3
Time series characteristics
Customer cases and scenarios
Azure Time Series Insights & Data Explorer
Azure Machine Learning
Conclusions
6. 4 components of Time Series
7
Seasonality
Trend
Cyclicity
Irregularity
7. Multiple time series
| Not all time series have easy to detect components
| Multiple related time series can vary differently over time
| Combination of parameters at a given time can indicate state
| Time windows can result in much more relevant findings
8. Examples
9
| Stock prices
| Weather reports
| Electricity demand
| Revenue numbers
| Temperature readings
| Number of passengers
| Criminality numbers
10. Some use cases
11
Improved outcomes and
increased revenue
Industrial IoT &
Supply Chain Optimization
Predictive & preventive
maintenance
Delivery optimization
Real-time anomaly detection
Energy planning & trading
Sensor stream data
Inventory data
Production data
Transport & Retail data
Tuning parameters
Manufacturing
Improved consumer
engagement with machine
learning
Data-driven stock,
inventory, ordering
Demand-elasticity
Predict inventory positions &
distribution
Right product, promotion,
at right time
Shopping history
Online activity
Demand plans
Forecasts
Sales history
Retail
Enhanced customer experience
with machine learning
Risk, fraud, threat
detection
Predictive analytics & targeted
advertising
Card monitoring & fraud
detection
Decision simulations & forecasting
Transaction data
Market data
Purchasing History
Clickstream data
Financial Services
14. Communication & runtime
15
PLCs,
Databases,
Message Buses,
SCADA Systems,
MES Systems,
ERP Systems
Processing
IoT Hub & DPS
Data integration
IoT
Edge
Publisher
Storage
Twin
File upload
Telemetry
Device twin
Commands
Methods
MQTT
AMQP
HTTPS
MQTT
Lifecycle
Provisioning
Actions
Hot path analytics
Cold path analytics
Long term storage
Applications
Digital twin
Relations
DevOps
Monitoring
Security
Infrastructure
Reference architecture
Environ-
ment
Stream
Analytics
Azure ML
Cognitive
Event Grid
Functions
Time Series
Insights
Azure SQL
Database
Blob
Storage
Data Factory
Blob Storage
Cosmos Db Data Lake Synapse
Databricks Azure ML Data explorer
ASA Azure ML Time Series I.
Logic Apps
Functions
Devops
App Service
Power BI
Data Share
Power Platform
App Service
Tenants
16. Predictive maintenance data set
17
| Public dataset (Nasa Turbo fan)
| Damage propagation for aircraft engine
| Run-to-failure simulation
| Aircraft gas turbines
| Dataset contains time series (cycles) for all
measurements of 100 different engines
https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/#turbofan
20. 21
Time Series data
Azure offers two services to ingest, process, store and query highly
contextualized, time-series-optimized IoT-scale data:
Azure Time Series Insights & Azure Data Explorer
21. Azure Time Series & Data Explorer
22
Azure Time Series Insights
| Built on top of ADX
| Very easy to set up
| Perfect for exploratory and
visualization purposes
| Query possibilities through the API
Azure Data Explorer
| Foundational service for many other
Azure services
| Extremely powerful
| No exploration portal
| Queries through KQL
| Fully customizable
23. Scenario: Prevent outage of engines
24
Job to be done
What are you trying to achieve?
Business Impact
Benefits
How will it used in the processes
What actions are linked to decisions
Data fuel
What data is available?
Are data streams available?
Is the training data labeled?
Definition of success
Predict Evaluate Trust
What do we want to predict?
Classification / estimated value
What if the model is wrong?
What accuracy do we expect?
Evaluation period
When do we trust the model?
What is needed to call this a success?
Risks
What risks do we see for the project?
Feedbackloop
Possibilities to improve & retrain
Future scenarios
Related scenarios & applications
Designed for: demo purposes
Designed by: Sam Vanhoutte
Date: July 31, 2020
Predict time to failure of engines
Stream:
100 engines, 24 sensor values
20.631 labeled records
People involved
Stakeholders, users, decision makers
Users: Operators
Time for maintenance
Classify for warning
Regression of ttf
False alerts are
better than missed
anomalies
Accuracy > 90%
Model can be used
to alert people who
can double check
When outage of production
decreases
When false alerts are not
happening a lot
Avoid downtime
Increase reliability
Impact of different engines?
Finding the right
ttf threshold
Side / side human validation Integrate alerts with
servicedesk system
Deploy to the edge
24. MLOps process (generalized)
25
Analyze Signals for Retraining
Register Model
Model Registry
Model Telemetry
Validate &
Deploy
Collect
Feedback
ML Pipeline
Publish training pipeline
Submit Code
for review
Experiment
Interactively
Data
Scientist
ML Engineer
Batch
predictions
Real time
predictions
User-facing
application
Train Model
26. 3. Stream Analytics: in the cloud & on the edge
27
Presentation &
Action
Storage &
Batch Analysis
Stream
Analytics
Event Queuing
& Stream
Ingestion
Event
production
IoT Hubs
Applications
Archiving for long
term storage/
batch analytics
Real-time dashboard
Stream
Analytics
Automation to
kick-off workflows
Machine Learning
Reference Data
Event Hubs
Blobs
Devices &
Gateways PowerBI
27. Takeaways
28
| Ingest data into Time Series
Insights
| Enable Data exploration, querying
and visualization
| Extend to Machine Learning, Data
Science and Front End
applications
| Out of the box integration with
Data Lake, Power BI, etc
Azure offers
plenty options for
Time Series
processing
28. Reference case
29
Getting Started
| Request your workshop
| 2 flavors
| IoT
| Data / AI
| Outcomes
| Business case definition & strategy
| Requirements
| Azure capabilities
| Architecture
| First proof-of-concept
Seasonality: variations that repeat over periode (shorter periods)
Trend : long term variation
Cyclical effect: fluctuations around trend (economic / political circumstances)
Irregularity / Residual (random variations, without pattern – external influences)
Seasonality: variations that repeat over periode (shorter periods)
Trend : long term variation
Cyclical effect: fluctuations around trend (economic / political circumstances)
Irregularity / Residual (random variations, without pattern – external influences)
Duco Ventilation & Sun Control wanted to lay the groundwork for AI with a first step in IoT. Enabling a more accurate view of its residential ventilation systems’ performance and stakeholders’ experience, Duco saw IoT as the key towards optimizing its products through data-driven processes.
Duco needed a solution with multi-location data capture, centralized system monitoring, as well as device and data management – all while providing an enhanced experience for various stakeholder (end user, R&D, partners, field services, …)
ACR: 30k/year
Engie has over 500 renewable energy production sites, including wind turbines and solar panels, collecting billions of messages every day – and counting. They needed a secure, scalable IoT solution to maximize real-time control, minimize lost time due technical issues and intelligent energy production.
Pain: High maintenance costs, manage energy streams
Solution: Build entire IoT platform to co capture and process data coming out of the SCADA using IoT Edge.
ROI: Capture more data in less time with better traceability, and scalable solution. Moreover they can balancing its energy production portfolio based on market data, when it’s most optimal to produce energy.
Now making the next steip and bringing in machine learning algorythems and Digital Twin
ACR: 250K/Year
Time Series ID : iothub-connection-device-id
Job to be done:
describe the scenario
why the customer needs this
Business impact:
What are the benefits for the organisation
How will the model and solution be used in the entire process of the organisation
Which actions and consequences depend on the outcomes of the model
Data fuel:
Define which data is already available
Will the data grow and new data be fed into the system?
Do we have labeled data (for supervised learning) or is the data unlabeled
Definition of success:
What is needed to call the project a success.
Describe adoption blockers that need to be tackled, dependencies in the organisation
Predict:
What do we want to predict
Describe the case for the model
Indicate the type of prediction (classification, regression (values), clustering, sentiment analysis, etc)
Evaluate:
Please reflect on the impact in case the model has wrong predictions (False Negatives & False Positives)
Should the model focus on overall accuracy (get as much as possible guesses right), or do we have to decrease the amount of False Negatives/Positives for example ?
Trust:
How long does the model needs to be evaluated and used before it’s considered approved and trusted?
Which dependencies do we have on the rest of the processes in order to gain trust
Execution:
Define where the model should be executed (in the cloud, on the edge, in a device, wherever)
Feedback loop:
How can the model be monitored and improved, once it’s operational?
Who will be monitoring the model and how will feedback be collected?
Future scenarios:
Related solutions, applications or scenarios that can be made possible