Slides from hybriData's presentation on using deep learning and computer vision to improve corrosion risk assessments at the Emerging Flow Assurance Technology Forum and Workshop organized by Saudi Aramco on Oct. 7th - 9th, 2019 at Dhahran, KSA
2. ● Challenges and Benefits
● Corrosion Types
● Computer Vision and Deep Learning
● Results
● Conclusion
Agenda
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
3. ● Limited amount of data
● Quality of data (image) samples
● Inference Response time
● Model training
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
Challenges
4. ● Systematic, unbiased and more objective risk assessment
● Reduced time to implement facility corrosion risk
management strategy
● Reduced costs - Inspection and Execute once
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
Benefits
5. Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
NACE Corrosion Types
● Group 1: Readily identifiable by ordinary visual examination
● Group 2: May require supplementary means of examination
● Group 3: Verification is usually required by microscopy
6. Group 1
● Uniform corrosion
● Pitting corrosion
● Galvanic corrosion
Corrosion Types Considered
Group 2
● Intergranular corrosion
Group 3
● Stress Corrosion Cracking(SCC)
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
7. ● Object Detection and Localization
● Instance Segmentation
Computer Vision
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
8. Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
Object Detection and Localization
Stress Corrosion Cracking: 90%
Uniform Corrosion: 72%Uniform Corrosion: 78%
● Detect multiple corrosion types on a pipeline by their locations
9. Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
Instance Segmentation
Stress Corrosion Cracking: 98%
Uniform Corrosion: 75%Uniform Corrosion: 84%
● Identify multiple corrosion types present on the pipeline at their pixel level
10. ● Backbone Network: Resnet
● Object Detection and Localization: Faster R-CNN
● Instance Segmentation: Mask R-CNN
Deep Learning
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
11. Faster R-CNN
● Unified network for object detection
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
12. Mask R-CNN
● Takes output from Faster-RCNN and generates masks for the regions
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
13. Results: Object Detection and Localization
Original Image Result
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
15. Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
Results: Object Detection and Localization
Original Image Result
17. ● Improvement in model accuracy - Advances in feature
extraction and compute infrastructure
● Limited data - Generate data with GAN
● Distributed deep learning - Amplified by distributed file
system
Conclusion
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA
18. Thank you!
Emerging Flow Assurance Technology Forum and Workshop, Oct. 7th - 9th, 2019, Dhahran, KSA