RECOGNITION OF DEFECTS ON THE METAL SURFACE USING MACHINE LEARNING

Abstract

Due to the increase in product quality requirements in the metallurgical and machine building industries, it is necessary to introduce modern technologies for automatic quality control. Surface defects of metal products (cracks, scratches and inclusions) directly affect the reliability and durability of products. Traditional methods of visual and optical control require significant time and labor costs, are subject to the influence of the human factor and do not always provide sufficient accuracy. Within the framework of the study, a review of modern publications was conducted, which consider approaches to automatic defect classification, as well as discuss the possibilities and limitations of neural network architectures. The analysis of the sources made it possible to identify development trends in the field under consideration and justify the choice of the model architecture. An approach to the detection of defects in images of metal surfaces using convolutional neural networks is proposed. The architecture of the model has been developed, which includes three convolutional layers and fully connected neurons optimized using the ReLU activation function, the Dropout layer and the Softmax output layer. To train the model, we used an open dataset containing 1800 black and white images with six different types of defects. The classification accuracy was 95.83 %, and the value of the loss function was 0.0862. When tested on a test sample, the model correctly recognized 70 out of 72 images. The conducted research confirms the effectiveness of neural networks in the task of detecting visual defects. The presented model can be used in automated quality control systems and additionally adapted to various industrial conditions. In the future, optimization of the model architecture is planned to increase noise tolerance and data variability.