Machine learning in the physical world by Kip Larson from AWS IoT
- 1. AI NEXTCon Seattle ‘18 1/17-20th | Seattle #ainextcon http://aisea18.xnextcon.com
- 2. Machine Learning in the Physical World Kip Larson Principal Product Manager AWS IoT Analytics
- 3. • Analytics in IoT • Introducing AWS IoT Analytics • AWS IoT Analytics Components • ML Across IoT Fleets • LSTM for IoT What to expect from this session
- 4. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Devices Sense & Act Cloud Storage & Compute Intelligence Insights & Logic → Action Three pillars of IoT
- 5. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Why is AWS IoT Analytics important? Industrial automation Improved product design Optimized business processes Improved user experience
- 6. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Analytics within AWS IoT Devices Cloud Greengrass Rules Engine AWS Amazon Redshift a:FreeRTOS Amazon Kinesis Collect Preprocess & Enrich Store Analyze Visualize Corp Apps Corp Data Center Enterprise Applications & Non-IoT Data Contextual Data Intelligence Device shadow Message Broker Device Gateway Device Registry AWS IoT AnalyticsAWS IoT Core
- 8. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AWS IoT Analytics IoT Analytics is a fully managed IoT analytics service that collects, preprocesses, enriches, stores, analyzes, and visualizes IoT device data at scale. From raw sensor data to sophisticated IoT analytics
- 9. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AWS IoT Analytics requirements High volume of data Clean data Contextual information
- 10. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Challenges with AWS IoT data and analytics High volume Multiple sources Noisy and no standard format Incomplete and no contextual information
- 11. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. What customers are asking for Business relevant reporting Preprocessing unstructured, noisy data Data collection from multiple sources Time series data storageAdvanced analytics and machine learning
- 12. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. From AWS IoT data to sophisticated analytics Thing Registry Log InkLevel Queue SQS Ink Level CloudWatch Log Job History Tables DynamoDB deviceID/usage/ inklevels Forward Ink Levels Rule Low Ink Test Queue SQS Log Ink Remaining Lambda Worker Alarm State Worker Lambda Send Mobile App Alert deviceID/events/ consumables Ink Replace Rule Device Enricher Worker Lambda deviceID/events/jobs Send Ink Alarm Queue SQS Notifier Lambda Low Ink Alarm Queue SQS Log Job Queue SQS Log Job Lambda Worker Job Ink Process Queue SQS Job Calculation Lambda Worker Job Alarm Threshold Lambda Worker Print Job Raw Data Storage Lambda Logging 2 2 21 3 3 4 4 Ink Level Remaining CloudWatch Alarm 5 5 8 8 4 Sales Cloud SFDC Customer, Ink Order Table & Promotions Table 6 7 7 5 6 Low Ink Test Lambda Worker Check Ink Threshold 9 9 deviceID/inbound/ inkrefill SFDC Ingest Firehouse 10 11 11 9 12 19 14 14 18 Print Job Error 13 13 Device Enricher Worker Lambda Thing Registry 16 14 15 15 14 16 16 17 17 17 20 20 21 21 deviceID/*/* Backup Rule Job History & Ink Level APIs Enrich Job Queue SQS Enrich Job Lambda Worker Job Alarm Queue SQS Job Entry Completion Lambda 19
- 13. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AWS IoT Analytics overview Create IoT data Collect Preprocess & enrich Store Analyze & visualize
- 14. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AWS IoT Analytics components DatasetsPipelines Collect Preprocess & enrich Store Analyze & visualize, machine learning Channels Datastores Notebooks & Amazon QuickSight QueryCreate IoT data Operational and business applications
- 15. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. IoT Analytics: Collect via channels • Entry point to AWS IoT Analytics • Native integration with AWS IoT Core • Authoritative store of raw data from devices • Built-in micro-batching partitioning by date • Supports both binary and JSON data MQTT Amazon Kinesis
- 16. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. IoT Analytics: Preprocess & enrich via pipelines • AWS IoT specific device data preparation pipeline • Consumes raw data from a Channel and sends processed data to a datastore • Set of activities for filtering/transforming/ contextualizing messages • Works best with JSON data
- 17. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AWS IoT Analytics: Pipeline activities • Consume from a channel • Send to a datastore • Invoke AWS Lambda • Add/remove/select attributes • Apply regex • Filter messages • Validate messages • Evaluate math expression • Enrich from MQTT topic • Enrich from AWS IoT device registry • Enrich from AWS IoT device shadow
- 18. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AWS IoT Analytics: Store via datastores • Store optimized for analytical & time series queries against processed data • Is not a database, but an abstraction on top of several database technologies • Works on semistructured JSON data • Partitioned by time
- 19. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AWS IoT Analytics: Analyze via datasets • Result of analysis against data store • Run on schedule or ad hoc • Conceptually similar to materialized view • Accessible from the console/API/Jupyter Notebooks/Amazon QuickSight
- 20. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AWS IoT Analytics: Powerful analytical notebooks • Jupyter-based machine learning notebooks (using Amazon SageMaker) • Integration with existing AWS IoT Analytics datasets • Built-in templates for predictive maintenance and other IoT-specific use cases
- 21. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AWS IoT Analytics: Visualize via Amazon QuickSight • Native integration with Amazon QuickSight • Makes it easy to build visualizations, perform ad-hoc analysis, and get business insights.
- 22. ML Across IoT Fleets
- 23. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. IoT Fleet Problem Space Fleets come in two classes: Consumer and Industrial Consumer Homogeneous, non-interactive fleets Industrial Heterogeneous, interacting fleets
- 24. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ML Mission Types in an IoT Context • Missions have different levels of granularity across fleet types • Business loss functions vary • Some missions are fleet type-specific Segmentation “What do they have in common?” Survival Analysis “How much longer till it breaks?” Classifiers “Is it broken?” Forecasting “How much will it produce?” Optimization “Can we improve the way it works?”
- 25. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ML IoT Model Development • Learn across the fleet, deploy to the device • Model may be deployed in the cloud or on the device • Opportunity to retrain may be limited, opportunity to update covariates may be limited • Almost always time series data • Often multivariate
- 26. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Obvious Approaches Don’t Always Work For business process forecasting (e.g. customer demand) we often deploy seasonalized time series models, such as SARIMAX 𝑦"# = 𝜇 + 𝜑( 𝑦")( + 𝜑* 𝑦")* + ⋯ 𝜑, 𝑦"), + 𝜃( 𝜀")( + 𝜃* 𝜀")* + ⋯ 𝜃, 𝜀"), + Φ( 𝑦")0 + Θ( 𝜀")0 + 𝜉( 𝑥" + 𝜀" IoT data often creates problems for these models: Seasonal component may exhibit epicycles Data usually has gaps Seasonal component may be evolving Relevance of regressors may be conditionally-relevant
- 27. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. LSTM-based approaches • Long Short-Term Memory networks show promise in forecasting time series device telemetry • Seasonal epicycles accommodated more easily than SARIMA-style lag operations (needs data and hyper- parameter optimization) • Forget-gates useful for noise management • Fewer cold-start problems • Gap tolerant • Areas of exploration: companion devices as inputs
- 28. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. After Model Training, now what? Train in SageMaker or Lambda GreenGrass ML on the edge Automatically trigger in IoT Analytics Action in the cloud
- 29. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Get started with AWS IoT Analytics aws.amazon.com/iot-analytics
- 30. Thank you! Interested in joining our team? Message us at [email protected]