Downloaded 18 times
![제 23회 보아즈(BOAZ) 빅데이터 컨퍼런스 - [MBOAX] : ABSA를 활용한 소비자 반응 분석 기반 운영 효율화 대시보드 설계](https://cdn.slidesharecdn.com/ss_thumbnails/3-1boaz23rdconferencemboax-260203102709-9d519923-thumbnail.jpg?width=640&height=640&fit=bounds)
Mtc strategy-briefing-houston-pd m-05212018-3
- 1. Predictive Maintenance using AzureAI Azure Cloud AI MTC Strategy Briefing for Oil & Gas, May 21, 2018, Houston, TX Speakers: Ram Krishnan, Naveen Vig Microsoft Cloud AI Solutions, Redmond, WA Topics: - Predictive Maintenance – key concepts - Solution Template Demo - http://github.com/azure/AI-PredictiveMaintenance - Contact: azureaipdm@microsoft.com
- 2. Types of Maintenance Corrective– repair or replace parts as and when they fail Preventive – inspect, repair or replace by schedule Predictive – conditionally inspect, repair, replace just in time Prescriptive/autonomic + Condition-driven preventive maintenance, improves utilization & efficiency + Reduces costs from unscheduled maintenance, labor - Data readiness is a prerequisite. New, model driven processes. + Parts are used completely (no wasted life). Minimal data readiness – well established processes. - Costs from unscheduled downtime, labor, and presents high risk with catastrophic failures + Minimizes unscheduled & catastrophic failures. Moderate data readiness – established processes. - Costs from scheduled downtime, under-utilization of parts, and (more frequent) labor
- 3. Predictive Maintenance + + + + + + + + + + + + Reduce operational riskof mission critical equipment Increase asset utilization and performance Predictive Maintenance Benefits Enable just-in-time maintenance Reduce labor costs for maintenance Reduce inventory costs by smart reordering Improve people productivity Schedule timely and effective inspections Speed up root cause analysis Streamline compliance efforts Plan capex based on remaining lifetime Retain customer trust & satisfaction Maintain competitive edge
- 4. • Predict amachine failure in the near future (next N time units) • Estimate remaining useful life of a machine • Predict failure along with the type of failure • Predict the type of maintenance action • Identify main causes of failure of an asset • Other variations (Example: “when should this machine be serviced?”) Predictive Maintenance Problems
- 5. Qualifying a problemfor PdM Qualification Details The problem has to be predictive in nature • A clear target or an outcome to predict • A clear path of action to handle failures when they happen The problem should have operational history with both good & bad outcomes. • Actions taken to mitigate bad outcomes • Error reports, maintenance logs showing degradation • Repair and replacement logs Relevant and sufficient data of high quality to support the use case. • Data should be relevant to the component • Data should have sufficient bad outcomes, and a large number of good outcomes – for learning to be proportional • Data should be of high quality Domain experts with clear understanding of the problem and/or internal processes. • Awareness of the internal processes and practices • Authority to make changes to help collect the right data
- 6. Data requirements forPdM Nature of Data Example Telemetry The operating conditions of a machine. For example, data collected from sensors. Maintenance records • Machine maintenance history detailing component replacement • Regular maintenance activities with the date of replacement. Error logs Log of non-critical errors. These may indicate an impending component failure. Failure records The failure history of a machine, or component within the machine. Machine metadata Features differentiating each machine. For example, age and model.
- 7. INGEST TRAIN –TEST DEPLOY PUBLISHSTAGE PREPARE PdM – Analytics pipeline Feature Engg Model Training Model Testing Data Prep Data Staging Model Deploy Data Ingest Publish results Iterative inner loop Processing outer loop
- 8. Feature Engineering inPdM – lag features • Predictive Maintenance is ‘in future time’ • Tell me what can happen in future N days/months based on your learnings from past N days/months. (points in time are aggregated over a longer horizon) • Predictors are ‘lag features’ defined by aggregating time series data • Example aggregates: counts, max, min, average, variance, cumesum, … Example: To predict ahead by W = 3 days, aggregate data for each event over last three days (rolling aggregate) Example: To predict ahead by W = 3 days, aggregate data for each event for every three days (tumbling aggregate)
- 9. Feature Engineering –aggregating & labeling data # m/c_id Timestamp LABEL 0 625 2015-01-01 12:00:00 M3 12 (1,0,1,0) 1 625 2015-01-02 00:00:00 M3 12 2 625 2015-01-02 12:00:00 M3 12 • Labels “teach” the ML algorithm how to discern between pass & fail cases. • Label defined as failure events that occur within a time window, rather than failure event that happen at a point in time. • Label the target variable based on the business problem & ML technique. Reference: John Ehrlinger, feature_engineering.ipynb – Machine Learning Sample for PdM
- 10. Modeling • Modeling algorithms •Model evaluation • Precision • Recall • Accuracy • F1 score • Cost based ROC curves
- 11. Documentation Azure AI Guidefor Predictive Maintenance (link)
- 12. Documentation Learn AI forPredictive Maintenance (learnanalytics.microsoft.com)
- 13. Azure AI Applications Frameworks MachineLearning and Deep Learning Toolkits CNTK – C#Tensorflow ML Server – R, PythonPython Libraries: Scikit, Caffe, Keras, .. Inferencing Spark, SQL, Cosmos DB Kubernetes Docker Edge Infrastructure Azure Storage, Data & Event Services Hardware (CPU, GPU, FPGS & ASIC) CPUs Services Cognitive Services AML Web Services BOT Framework Processing DSVMAML Model Experimentation AML Model Management Batch AIAML Data Wrangling AML Toolkits Computer vision Forecasting, Text Dev&DeployTools Azure AI Platform
- 14.
- 15. GENERATE Storage Blobs Sensor data from devices Sensor Data Generator IoT Hub INGESTTRAIN – TEST DEPLOY PUBLISH CONSUMESTAGE Service Bus Maintenance & Error logs PREPARE Admin Dashboard Data Science VM Storage Tables KubernetesOnline Scoring Data Predicted Output Logs & Device Metadata Input data Train-Test data New data (to be scored) Output data (with predictions) Modeling operations Predictive Maintenance Solution Template AZTK – Azure Batch Train-Test Data Feedback for retraining
- 16. GENERATE Storage Blobs Sensor data from devices Sensor Data Generator IoT Hub INGESTTRAIN – TEST DEPLOY PUBLISH CONSUMESTAGE Service Bus Maintenance & Error logs PREPARE Admin Dashboard Data Science VM Storage Tables Kubernetes Batch Scoring Data Online Scoring Data Train-Test Data Predicted Output Logs & Device Metadata Train-Test Data Input data Train-Test data New data (to be scored) Output data (with predictions) Modeling operations Predictive Maintenance Solution Template 2 Feedback for retraining
- 17. GENERATE Storage Blobs Sensor data from devices Sensor Data Generator IoT Hub INGESTTRAIN – TEST DEPLOY PUBLISH CONSUMESTAGE Service Bus Maintenance & Error logs PREPARE Admin Dashboard Data Science VM Batch AI Storage Tables Kubernetes Batch AI Batch Scoring Data Online Scoring Data Train-Test Data Predicted Output Logs & Device Metadata Train-Test Data Input data Train-Test data New data (to be scored) Output data (with predictions) Modeling operations Predictive Maintenance Solution Template 3 Feedback for retraining
- 18. Python is takingover Cloud ML & analytics (python.org) SIGS Books
- 19. MTC Strategy Briefingfor Oil & Gas, May 21, 2018, Houston, TX - Predictive Maintenance – key concepts - Solution Template Demo - Azure AI Guide for Predictive Maintenance - http://github.com/azure/AI-PredictiveMaintenance Contact: azureaipdm@microsoft.com Ram Krishnan – Senior Program Manager, Cloud AI Engineering (ramkri@microsoft.com) Naveen Vig – Principal Solutions Architect, Cloud AI Engineering (navig@microsoft.com) Summary
- 20. © 2017 MicrosoftCorporation. All rights reserved. Microsoft, Windows, and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries. The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION. 2018