|M.Sc Student||Shikler Roi|
|Subject||Fractographic Image Analysis using Computer Vision and|
Deep Learning Methods
|Department||Department of Mechanical Engineering||Supervisor||Professor Anath Fischer|
|Full Thesis text|
Many mechanisms operate under dynamic loading regime during daily use, e.g., helicopter rotors, electric saws, engine parts, etc. If a structure breaks, the producer or the operator of the machine investigates the failure cause. The insights may help the company plan ahead for critical events, by updating the maintenance policy or the operating instructions or even lead to redesigning the whole mechanism. Substantial financial and engineering resources are directed towards fractographic analysis, as a key component in uncovering the failure causes of these mechanical systems. When done successfully this process provides critical information regarding the mechanical loading history.
Today, most of the fractographic analysis is being executed manually, so failure investigation processes take a long time, are prone to human error and demand a fair amount of resources. In this study, we propose to automate one of the most time-consuming stages of the fractographic analysis process: locating fatigue striations on SEM (Scanning Electron Microscope) images. FS (Fatigue Striations) are lines appear on the fracture surface, which created by crack propagation under a cyclic loading regime. These lines appear in groups, and those groups are the most significant indicator of a fatigue failure on fracture surfaces (on the microscopic level).
The advantage of the detection algorithm developed in this study is that the FS groups texture properties does not define manually. Defining the FS groups unique features may be complicated, especially using basic computer vision tools which are not built to find abstract features.
In this study, we used DL (Deep Learning) to detect the FS groups features. DL is a branch of machine learning based on a set of algorithms that model high-level abstractions in data. Using a dataset of marked images, DL is capable of building image-processing mathematical models which define complex textures and patterns, that sometimes are impossible to identify by classic computer vision methods.
Automatic detection of FS on the fracture surface may reduce the fractographic analysis time from days to minutes. This enormous time reduction would allow the execution of the full fractographic analysis more often, a fact that could improve the safety of critical mechanisms (such as aircraft's parts) - which eventually prevent fatal accidents.