Damage detection in laser nozzle images with convolutional neural networks methods

Today, laser technologies are one of the advanced production methods for cutting, punching and shaping metal materials. Thanks to laser technologies, unprocessed metal plates called sheet metal can be easily processed according to production needs. Thanks to laser cutting methods that work with different gas types such as carbon dioxide, oxygen and nitrogen, sheet metals processed with heat treatments show more precise and better cutting results by minimizing deformation. Laser machines vary in size according to their sheet metal processing capacity. They can process unprocessed sheet metals of meters in length and width in a short time in line with cutting programs prepared according to production needs and to obtain workpieces. For this reason, it produces both faster and more precise results than conventional processing methods. Due to the precision processing capacity of laser technologies and the modernity of the technology, the accuracy and precision of the materials used in laser cutting machines directly affect the precision of the workpieces to be produced. One of the factors directly affecting the cutting quality of the workpieces is that the laser beam, as it leaves the cutting head, must be correctly focused and the laser beam must not touch any surface other than the sheet metal. The closest element of the laser cutting heads to the sheet metal is the nozzle. The nozzle, which is the last element of the laser cutting heads, plays an important role in ensuring the contact of the laser beam with the material, the contact of the focal point of the laser beam and the auxiliary gas to the material in a fully centered manner, and the removal of slag and smoke, which are molten materials splashed on the sheet metal during cutting, from the laser. During the laser cutting process, it directly affects the quality of the workpieces after production. The nozzle material, which is generally made of brass or copper material, can be damaged and lose its function due to many different reasons such as splashing of slag, which is the molten material during cutting, incorrect determination of cutting parameters that directly affect the cutting process, errors made in the cutting program prepared for the workpieces, hitting or contacting the cutting head with sheet metal or guards, inexperienced or untrained operators using laser machines. Due to damage or malfunctions in the nozzle material, the cutting quality decreases or the cutting cannot be continued. With deep learning and artificial neural network (ANN) technologies, which are sub-branches of artificial intelligence, the decision-making mechanism needed in processes is now left to software. The data obtained for the problems encountered are taught to the software and they are expected to make decisions by utilizing the learned situation when faced with the same problem. In this thesis, AlexNet, DenseNet, ResNet, VGG and EfficientNet models, which are convolutional neural network models developed with deep learning methods, are trained to develop a system that detects whether the nozzle material, which directly affects the cutting quality in laser cutting machines, can continue the cutting process. The nozzle images used in the model training were obtained with the help of a camera mounted on the laser cutting machine and each image was labeled as damaged and undamaged under expert control. After each model is trained, an ensemble learning method is developed that learns from the predictions made on the data set and determines the final result.