- 微博 QQ QQ空间 贴吧
- 视频嵌入链接 文档嵌入链接
Application and Challenges of Streaming Analytics and Machine Learning on Multi-Variate Time
Manufacturing Industries create value by transforming raw materials into products. Many of these processes are automated with sensors and control systems which, along with manufacturing execution systems (MES), generate large volumes of data. The key objective of smart manufacturing is to harness these data to enable data driven, predictive operations to optimize the process for higher throughput, quality and energy efficiency. The complex nature of manufacturing systems poses unique challenges in building a robust data pipeline. Data needs to be collected from multiple sources at different granularities.
It must be prepared and enriched to make it ML ready and for processing of 1000s of deployed ML models at the edge or cloud. To make machine learning useful, it must also be blended with complex event processing (CEP). The talk will describe the challenges of multivariate time-series data in Smart Manufacturing context, our approaches to dealing with these challenges, and our learnings.
In this talk, we share some of the challenges faced in building a streaming analytics & ML pipeline. More specifically, we discuss handling time series data with different granularities and arrival order. Data spikes are also not uncommon and can pose a serious challenge to the operating SLAs of such a system. We present in detail our streaming data pipeline, which includes production deployments of ML models and CEP on edge and cloud. Using Spark, Kafka and the ecosystem around it, our team has created a platform capable of monitoring thousands of manufacturing equipment assets with millions of data points, in near real time.
1 .WIFI SSID:Spark+AISummit | Password: UnifiedDataAnalytics
2 .Application and challenges of streaming analytics and machine learning on multi-variate time series data for smart manufacturing Pranav Prakash, Quartic.ai #UnifiedDataAnalytics #SparkAISummit
3 .Pranav Prakash • Co-Founder, VP Engineering at Quartic.ai • Ex- LinkedIn SlideShare • Passionate about – A.I., Computer Vision, 3D Printing – Music, Caffeine 3
4 . Downtime What A cool startup solving some real- Reduction use case of a critical asset in Pharma you’ll life use cases world •And a “secret” to solve such problems learn in next 40 Challenges in Industrial Stream Spark specific stuff that we learned Processing mins 4
5 .We enable Industry 4.0 • AI powered smart manufacturing platform • Processing Billions of sensor data every day • Work with top Pharma companies on multiple use cases • Team of 22 techies including Engineers & Data Scientists + 4 Domain Veterans #UnifiedDataAnalytics #SparkAISummit 5
7 . We started by • Increased uptime of sterilization autoclave by 7 days building • Increased yield of protein from fermentation process solutions for • Incubated egg harvester – increase uptime during critical flu season pharmaceutical • Cold-chain monitoring for pharma refrigeration – reduced downtime and waste manufacturing • Predictive health monitoring of air handlers for clean rooms in pharma And created a • Enable continuous validation of biologic production DIY platform process • Medical Device Assembly – reduce recalls caused by poor quality.
8 .Case study – an Intelligent Asset Health Monitoring system for an Industrial Autoclave • Mission - Improve the reliability of a complex asset. • Details - 13 different modes (cycles) • Runs 24/7 • Critical Asset
9 . • Capture process, condition data • Establish baseline and measure Equipment deviations Reliability • • Forecast the future Classify errors early • “Advisory Mode” AI
10 .SCADA = Supervisory Control and Data Acquisition PLC = Programable Logic Controller
11 . • Data – Speed: 10ms – 2 hours System – Volume: Couple 1,000s sensors per asset. 10,000s of asset per Design enterprise – Data Type: String, Numeric, Params Boolean, Array – Timeseries, Discrete
12 . • Deployment – Edge (80%) System • Hardware Limit Design • Many cloud-only solutions won’t work Params • High Uptime, Low Response Time – Cloud (20%)
13 . • Use Cases System – Automatic Model Param Tuning, Model Training Design – 1000s of ML Models Deployment – Complex Event Processing (CEP) Params – Statistical & Analytical Processing • Rule Recommendation • Near Real Time Stream Processing
14 . • ML – Multiple granularities – Late Data Arrival Challenges – Model Deployment on a heterogenous data stream – Flash Flood of Data
15 .Multiple Granularities - Both belong to same “Asset” - Target Feature – C/D or A/B Poll Frequency = 1s Poll Frequency = 5s TS Sensor A Sensor B TS Sensor C Sensor D 12:03:01.198 12:03:01.230 12:03:02.283 12:03:06.233 12:03:03.316 12:03:11.316 12:03:04.572 12:03:16.520 12:03:05.283 12:03:21.283 12:03:06.342 15
16 . Multiple • Approximation (Roundoff) • Aggregation Granularities • Filling - Forward or Backward or Average
17 .Late Data Arrival https://spark.apache.org/docs/latest/structured-streaming-programming-guide.html#handling-late-data-and-watermarking 17
18 . • Watermarking Late Data – Homogenous stream: One watermark per Stream Arrival – Heterogenous stream: multiple watermark per “Usage Condition”
19 .- Watermarking time automatically and dynamically chosen - Data later than threshold is discarded
21 . Model Deployment
22 . • Backpressure enabled Flash Flood • Allows Ingestion rate to be chosen dynamically and automatically of Data • PID Controller 22
23 . Complex Event Processing • Insights – PySpark + yahoo/graphkit • Rules – Scala Spark + drools
24 . • Industrial IoT is different • Context = Process Data + Condition Data Summing • Techniques for processing heterogenous stream it up
25 .We’re hiring firstname.lastname@example.org 2 5
26 .DON’T FORGET TO RATE AND REVIEW THE SESSIONS SEARCH SPARK + AI SUMMIT