OpenTalks.AI - Климент Горчаков, Искусственный интеллект в промышленности: предсказание поломок оборудования и оптимизация производственных процессов

Transcript

1. Industrial AI Scalable and Accurate GPU-driven AI for Predictive Maintenance, Quality Control and Process Optimization
2. Company Intro Data Monsters is an AI lab Data Monsters − We are crazy about data! 12 years in the data science and engineering market 60+ Data Scientists, 11 PhDs 60+ data engineers 100+ projects in the portfolio 6 university areas in the USA, Europe, Japan, and Australia
3. Partners and Clients Data Monsters is an implementation center for Deep Learning Predictive Maintenance Solutions Clients Parters
4. Industry 4.0 Industry 4.0 is a name given to the current trend of automation and data exchange in manufacturing technologies. It includes cyber- physical systems, the Internet of things, cloud computing and cognitive computing. Industry 4.0 is commonly referred to as the fourth industrial revolution. • Predict equipment breakdowns • Predict product quality • Optimize manufacturing processes Opportunities Of Fourth Industrial Revolution Mechanization, Water power, Stream power Mass production, assembly line, electricity Computer and Automation Cyber Physical Systems 1th 2th 3th 4th
5. Solution A digital twin is a real time digital replica of a physical device. Predictive Maintenance Supercomputer Manufacturing Predictive Maintenance Supercomputer Nvidia GPU Digital Twin Equipment Sensors Labs Cameras Applications Predictive Maintenance Quality Control Process Optimization
6. Use Cases • Oil and Gas • Aviation • Pulp and Paper • Chemical • Metal • Mining 16 Successful Cases Pulp and Paper Chemical Oil and Gas Metal and Mining
7. What is different A digital twin is a real time digital replica of a physical device. Model Performance
8. Use Cases Challenge Reduction of losses due to aircraft downtime caused by unexpected equipment malfunctions. Optimize logistics due to prediction of aircraft equipment malfunctions. Aircraft: Airbus A320. Client Subcontracting for top-10 world airlines company (major contractor was Integro Technologies). Results Reduction of losses due to downtimes of aircraft caused by malfunctions of specific equipment up to 25%. Aircraft Parts Failure Prediction Implementation Software for prediction of aircraft parts malfunctions within particular period with built-in deep learning model. The model includes an ensemble of LSTM with meta-features pretrained on other cases of faults prediction. • DFD, Events, AMOS, refilling and other data has been used; • Noisy data and incomplete information.
9. Use Cases Challenge The company wants to maximize the quality of its products, keeping the ash and moisture indicators within certain limits according to industry standards. To estimate current product quality before laboratory measurement, to increase the production of the final product by reducing rejects. Client The client is a large coal mining and processing enterprise. Results • Less than 5% of MAPE (mean absolute percentage error) on prediction. • Up to 4% of product quality improvement on simulations. Coal Enrichment Data Values are automatically collected from ~400 sensors. Product quality is measured every 2 hours. There are over 30 adjustable parameters. Implementation Quality prediction part Provide prediction of coal quality before it was measured in laboratory. Digital twin was used • To create synthetic data to enrich train dataset; • As a basic unsupervised initialization of the final prediction model. Optimization part Provide near real-time recommendations to engineers of control values changes to improve product quality. Digital twin was used • As an environment RL agent model that recommends control values changes to engineers; • Reward function was taken from prediction model.
10. Use Cases Client Client is a large pipe manufacturer. Challenge The client experiences up to 20% downtime due to unexpected breakdowns on different production stages. One of the most painful stage is pipes welding where the client has up to 9% of downtime affects the whole production line and also affects the pipes quality. Results • Prediction of 60% of all failures with 72% precision; • 81% precision on localization of root cause (i.e. - type of breakdown); • Rough estimation of annual economy because of decreasing of downtime more than $25mln. Welded Large Diameter Pipes Production For Oil And Gas Data More than 500 sensors from electronic, welding and pipes transport systems. Implementation The product predicts in advance potential failures of welding equipments and defines the root cause of potential and realized failures to localize and accelerate maintenance works.
11. Use Cases Challenge Optimize maintenance schedule of submersible equipments in oil wells. Equipment includes pumps from many manufacturers: REDA, Schlumberger, EZLine, etc Client Top-10 oil/gas company in the world Results Best model out of seventeen companies. 15% growth of submersible equipment lifetime. Predictive Maintenance For Oil Wells Data Data from sensors from more than 6000 oil wells, pumps, internal components. Data includes failures events Data sampling varied from 1 second to 2 hour including asynchronous data sources. Implementation TensorFlow, LSTM neural networks, modbus protocol, SCADA, PID controllers.
12. Use Cases Challenge Predict tights and drags events using sensors data from well drilling. Such an events lead to both short-term delays of drilling team (about 4-8 hours) and non-recoverable in field’s conditions equipment breakdowns resulting in great increase of project’s cost. Client Top-20 oil/gas service company in Europe Results 30% decrease of drilling downtime because of tights and drags of drilling bit. Predictive Maintenance Of Drilling Oil Wells Data Data is collected during drilling process through different sources and include such measurements as drilling block position above rig floor, rate of penetration, rotation per minutes, torque etc. 5 second sampling of sensors data. Different sets of sensors for different oil wells with common sensors alongside all of them
13. Use Cases Challenge Reduce costs due to unexpected failures of plant equipment. Client Joint R&D with Kaspersky Lab for large chemical production company Results The solution can predict 81% failures with 92% precision within an hour before a failures. Predictive Maintenance For Plant Equipment Implementation Software for prediction of equipment faults with built-in deep learning model. The model includes an ensemble of LSTMs with meta-features. The software connects directly to control loop of the plant through modbus protocol, and inference prediction results to SCADA. Solution features: • Works with different types of sensors and control loops (PID-controllers such as Schneider M580); • Asymmetrical information models (partially damaged or inactive sensors); • Faults reasons detection and classification.
14. Technology We use the most advanced algorithms for the solution: • RCGAN • Reinforcement Learning • Anomaly Detection • LSTM • Supervised and semi- supervised learning • Data Augmentation • Attention Methods • Clustering • Optimization • Denoising Inference Process Dockers Preprocess Data normalization, denoising, cleaning, measurements frequency normalization, delayed signals handling Predict Anomalies Predict Sensor Influence Classify Anomalies Real-time dashboard 3d Party Systems Nvidia GPU Sensor, Labs and Cameras data flow Nvidia Docker Store data Docker Docker API API API
15. Predictions • Real-time dashboard • Light-weight Dashboard for an engineer Prediction Failure Reasons
16. System Health • System performance monitoring • Forecast horizon optimization • Gains per month Dashboard for an engineer Model is stable Optimal forecast horizon is 40 mins
17. Statistics • Failures • Downtime • Prevented failures • Gain • Failure reasons Dashboard for an engineer
18. Technology - Inference Depending on the environment and requirements, we recommend the right hardware. Nvidia DGX 1/2 Nvidia Tesla P100/V100 Nvidia Jetson TX1/TX2 Remote Sites, On premise* <1k sensors or 1 video stream prediction - 1/sec Low volume data* 1k-5k sensors, up to 2 video stream prediction 1/sec High volume data* >5k sensors (or many equipments), >2 video streams Prediction 1/sec * Recommendations are subject to change depending on client requirement Inference Process Hardware
19. Deployment Just 10 to 15 weeks to get a fully productional system. Define POC Deploy Definition 2 - 3 days Prepare Historical Data Prepare Digital Twin and Models Present the Demo in a Cloud Performance Proof 2 weeks Integrate with client IT ecosystem (1-4 weeks) Test solution in real production with engineers (8 weeks) Deployment 7 - 12 weeks Manage Solution and quality Add more equipment, etc Production Fill questionnaire Discuss answers Production Deployment Roadmap
20. Questions? Kliment VP of Sales +1.650.924.9282 kliment@datamonsters.com