engleski

Exempler Based Synthetic Dataset Generation From Magnetic Testing Data For The Purpose Of Training An AI Model For Crack Detection

Magnetic particle testing under UV light is one of the most sensitive NDT methods, widely used in the field of mechanical engineering with applications from construction to aerospace. In general, most of the NDT methods rely solely on inspection engineer's experience and knowledge to detect surface defects. Norm's dictat that these defects should be called linear or nonlinear. These linear or nonlinear defects can later be classified into several other categories, which can be used to our advantage while developing the system. As of writing this paper, there are many stories of neural networks accomplishing previously unthought off tasks as in the case of GAN's from 2017 or Convolutional neural networks from 2012. Also, there are currently companies like Cornis who use AI for services for maintenance, primarily wind turbine blade inspection. With all of this in mind a new concept of magnetic particle testing is proposed. There are many industrial applications that uses robots to automate the process, same applies for magnetic testing for example Tracked inspection robot Versatrax 100 MicroMag or Wall- Climbing Robots with Permanent-Magnet Contact Devices. From now on we will call this external device just a robot and try to generalize as applications and environments can vary. This type of human in the loop system would only work with adequate initial artificial intelligence training set. For starters let’s assume we collected some data as shown below. These types of defects are rare in the field of MT. Data collected from the field is not open source, and more then often, linear defects which are more dangerous occur once every ten measurements. This poses a question. How do we make a dataset if we don’t have the data? This is where computer graphics comes in. System shown to the right can output any number of images determined by the user. While typically only the green channel is used to detect the defect because it has the highest pixel values because of its florescence from lumogen, other channels can be used all the same. Results from one iteration are shown below. These images are stored in a dataset with correct labels taken from the norm classification, and defect bounding box location. This can be further used for the training of AI model which will be the topic of future work.

non destructive testing, in-paint exempler, synthetic datasets

nije evidentirano