Classification of 3D designs of mechanical parts using machine learning techniques

Machine learning techniques and data processing applications have been developing rapidly in recent years and offer solutions to many problems in different fields. Efficiency has increased in all sectors using software-supported solutions, and human errors and delays have been prevented. This study aims to classify standard mechanical parts used in the production and design industry using deep learning algorithms. In addition, the performances of the models used in the study will be compared. A two-stage experiment was established within the scope of the study. The stages of the experiment were independent from each other and the results were evaluated independently. In the first stage, a ready-made data set consisting of digital designs of mechanical parts of bolts, pins, nuts and washers was used. This data set consists of 7616 samples in total, including an equal number of 224x224x3 images from each piece. In the second stage, in addition to the same design visuals, real photographs of mechanical parts were added and a new data set consisting of 1600 samples in total was created. Then, in both steps of the experiment, AlexNet, VGG, EfficientNet, ResNet and Xception classification models were preferred to find a solution to the problem of recognition of parts, and the results were recorded. Then, in both steps, AlexNet, VGG, EfficientNet, ResNet and Xception models were preferred to list the broken ones and the results were generated. Initial replacement models offered average accuracy rates of 0.25, 0.25, 0.60, 0.69 and 0.75, respectively. Secondary renewal was successful with an average accuracy of 0.27, 0.25, 0.65, 0.85 and 0.96, respectively. The results of the study are stated in the comparison performances of the models.