![[Un] Supervised Learning
Supervised
• Need to train
• Text (previously classified)
for train is required.
Unsupervised
• Needless to train
• Text is not required.
Test data
Training data
Drawing the line
(Classification)
ICEFA08@Budapest, Hungary](https://image.slidesharecdn.com/icefa08-180712171414/85/Application-of-deep-learning-for-automatic-classification-of-fracture-surface-s-SEM-image-7-320.jpg)
Application of deep learning for automatic classification of fracture surface’s SEM image.
- 1. Application of deep learning for automatic classification of fracture surface’s SEM image. Kenta Yamagiwa (kenta@yamagiwa.org) Ph.D., Senior Researcher Mechanics and System Safety Group, National Institute for Occupational Safety and Health, Japan ICEFA08@Budapest, Hungary
- 2. Problems in Japan • Number of people – who are interested in failure analysis – can understand the fracture surface is decreasing. • Qualitative analysis – Low quality of fractography – Wrong results lead wrong countermeasures. • Unstable operation. ICEFA08@Budapest, Hungary
- 3. Supporting system of fractography Current status Human based Future status Machine based ICEFA08@Budapest, Hungary ! Needless to work. More beer and life elevated.
- 4. SEM (Scanning Electron Microscopy) of fracture surface Ductile (Dimple) Brittle (Facet) Fatigue (Striation) Feature of fracture surface depends on fracture mechanics. They are quite different as you can see. Is it possible to classify automatically? ICEFA08@Budapest, Hungary
- 5. Classification of Images Before classification After classification DOGS CATS ICEFA08@Budapest, Hungary
- 6. Deep Learning • Huge number of pictures – Ex. More than 1 million. • Deep layer of neural network • Supervised learning / Unsupervised learning • Needless to consider characteristic values/vectors like roughness parameters for the classification. https://www.xenonstack.com/ ICEFA08@Budapest, Hungary
- 7. [Un] Supervised Learning Supervised • Need to train • Text (previously classified) for train is required. Unsupervised • Needless to train • Text is not required. Test data Training data Drawing the line (Classification) ICEFA08@Budapest, Hungary
- 8. Convolution Neural Network (CNN) . . . . . . . . . . . . Convolution process (filters) ICEFA08@Budapest, Hungary 2 4 0 5 8 3 6 1 7 9 image number
- 9. Convolution Neural Network (CNN) Fatigue Brittle Ductile . . . . . . . . . . . . Convolution process (filters) ICEFA08@Budapest, Hungary Prob. = 0.2 Prob. = 0.1 Prob. = 0.7
- 10. Conditions • Library: Tensorflow (https://www.tensorflow.org) by Google • Machine: Mac mini (3GHz, Intel core i7, Memory 8G) – Too poor… • Method: CNN / Supervised Learning – Hyper parameters: Learning rate=10-5, Image size = 32 • Number of classification: 3 (ductile, brittle, fatigue) • Number of training images: 3132 (with augmentation) – ductile: 972 – brittle: 572 – fatigue: 1588 • Number of test images: 214 – Ductile: 14 – Brittle: 150 – Fatigue: 50 ICEFA08@Budapest, Hungary
- 11. Calculation process 1. Training process by using training images 2. Obtaining Training Score 3. Testing process by using test images 4. Obtaining Testing Score 5. Return to the Training. – Filter values in CNN is adjusted to reduce Error values. Training Score Err val. (Cross Entropy) Testing Score CNN CNN CNN CNN CNN CNN D D F F B D Training Process CNN CNN CNN D D B Testing Process Changing filter values in CNN ICEFA08@Budapest, Hungary Next Step
- 12. 0.5 0.6 0.7 0.8 0.9 1 1 10 100 1000 Score Number of training Training Testing ICEFA08@Budapest, Hungary Result Perfect score about 2k steps. Finally, testing score is more than 0.9.
- 13. Result (correct classification) Blue: Prob. of Fatigue Green: Prob. of Brittle Purple: Prob. of Ductile The network judged the pic as fatigue from the beginning. Correct fracture mechanics Fracture mechanics detected by the network ICEFA08@Budapest, Hungary
- 14. Result (correct classification) Blue: Prob. of Fatigue Green: Prob. of Brittle Purple: Prob. of Ductile Fatigue is negative from the beginning. At first, the network judged the pic as brittle. As the learning progressed, the network judged the pic as ductile. ICEFA08@Budapest, Hungary
- 15. Result (misclassification) Blue: Prob. of Fatigue Green: Prob. of Brittle Purple: Prob. of Ductile Brittle is negative. Dithering over whether fatigue or ductile. Different result ICEFA08@Budapest, Hungary
- 16. Confusion matrix Step = 1 Ductile Brittle Fatigue Trans-granular Ductile 0 28 0 0 Brittle 0 150 0 0 Fatigue 0 56 0 0 Transgranular 0 14 0 0 Step = 500 Ductile Brittle Fatigue Trans-granular Ductile 5 21 0 0 Brittle 1 149 0 0 Fatigue 0 0 56 0 Transgranular 0 0 0 14 Label of classification result CorrectLabel ICEFA08@Budapest, Hungary
- 17. Next step • Appending other fracture mechanics. Creep and SCC(Inter/Trans granular). • Application to the surfaces which are difficult to estimate the load (cat iron/aluminum) • Application to non-metallic materials (carbon composite). • Automation for taking SEM pictures. Fatigue fracture surface of cast iron (FCD450) ICEFA08@Budapest, Hungary
- 18. Conclusion • Deep learning for automatic classification of fracture surface’s SEM image is well worked. • It is enabled to support the beginner engineer to make decision for the detection of fracture mechanics. • More images and more deep network, more calculation resource are required. ICEFA08@Budapest, Hungary 0.5 0.6 0.7 0.8 0.9 1 1 10 100 1000 Score Number of training Training Testing
- 19. Acknowledgement • ELIONIX (SEM manufacturer) • Kiguchi Technics (Specimen) • TDU students (Fatigue test) • Thank you for paying your attention. ICEFA08@Budapest, Hungary