Authors:
Alexander Smirnov
1
;
Nikolay Shilov
2
;
Andrew Ponomarev
1
;
Thilo Streichert
3
;
Silvia Gramling
3
and
Thomas Streich
3
Affiliations:
1
SPIIRAS, 14th Line, 39, St. Petersburg, Russia
;
2
ITMO University, Kronverksky pr., 49, St. Petersburg, Russia
;
3
Festo SE & Co. KG, Ruiter Str., 82, Esslingen, Germany
Keyword(s):
Artificial Intelligence, 3D Simulation Data, Mechanical Stress Evaluation, Geometric Features, Resnet18, VoxNet.
Abstract:
When a new mechanical part is designed its configuration has to be tested for durability in different usage conditions (‘stress evaluation’). Before real test samples are produced, the model is checked analytically via 3D Finite Element Simulation. Even though the simulation produces good results, in certain conditions these could be unreliable. As a result, validation of simulation results is currently a task for experts. However, this task is time-consuming and significantly depends on experts’ competence. To reduce the manual checking effort and avoid possible mistakes, machine learning methods are proposed to perform automatic pre-sorting. The paper compares several approaches to solve the problem: (i) machine learning approach, relying on geometric feature engineering, (ii) 2D convolutional neural networks, and (iii) 3D convolutional neural networks. The results show that usage of neural networks can successfully classify the samples of the given training set.