A machine learning algorithm has been developed to solve the problem of predicting a vibrational state in order to improve the turbine rotor assembly processes using its digital twin. The digital twin of the rotor includes a parametric 3D model specially created in the CAD module of the NX program and a design project in the ANSYS system in which the working conditions of the rotor are simulated. The parameters of vibration acceleration and the reaction force of the rotor supports at critical speeds depending on geometric errors were calculated. To reduce the complexity of the calculations, neural network architectures were chosen to predict the parameters of the vibrational state depending on the geometric errors of the rotors. The novelty of the work lies in the creation and use of the original numerical model of balancing, taking into account the rotor manufacturing tolerances.
Traditionally, the blades of gas turbine engines are assembled manually using special techniques (sorting by mass or by several geometric parameters). The use of manual assembly leads to the need to perform reassembly of units and is characterized by non-optimal operating parameters achieved. The data obtained from the scanners when monitor-ing the geometry of the blades can be used to improve the quality of the assembly processes of compressors and engine turbines. In particular, parts can be picked based on key parameters. The results of the work will automate the receipt of such parameters. The paper presents a mathematical model for performing a computer calculation of the flow area of turbine nozzles. The model is used to work with objects of *.stl format when processing data after scanning. With the use of the developed model, the calculations will not depend on the qualifications of the controller, and the proposed approach will not require a unique adjustment and expensive repair and storage of standards for all standard sizes of nozzles. The calculation was carried out in three sections, which increases the accuracy of the calculations while reducing the complexity and cost of measurement. An increase in the number of sections practically does not increase the complexity of calculations, so there are no significant restrictions on the performance of virtual measurements in four or more sections. The model is implemented in the mathematical package MATLAB. The model was tested by the example of processing the STL model of the sector of the turbine nozzle blades.
The article is aimed at solving the problem of aerospace parts identification. A neural network model for part identification was developed. The proposed model consists of three modules: object detection using the YOLO3 model, preprocessing of the selected fragment, and classification of the processed fragment using the VGG19 model. A distinctive feature of the developed model is the use of STL objects for training the VGG19 neural network. To increase the reliability of the classification for each object we used photos made from three angles. The developed model has been tested on the parts of the rotor of a small gas turbine engine. The test was conducted on 100 test cases including 300 photos of parts. To train the neural network, 13,200 images were simulated using STL models. The loss function (categorical cross-entropy) for the training sample was 0.0004, and the classification accuracy was 100%. The accuracy of identification of real photos using the developed model was 97%.
The article considers the solution to an automation problem for assembly processes associated with determining a type of the part delivered for assembling. The parts delivered for assembling are preliminarily measured with an optical scanner. In order to solve the problem for determining the part type, convolution neural network organization was matched with the data classification. Stl-patterns for three turbine rotor parts were taken, the training and test samples were simulated, several convolution neural networks were trained and the optimal parameters ensuring 100% classification accuracy were selected.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.