Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Semi-supervised machine learning for chatter detection in turning with data augmentation
Submission Author:
Ana Julia da Silva de Oliveira , PR
Co-Authors:
Ana Julia da Silva de Oliveira, Daniel Awada Elarrat Canto, Giuliana Sardi Venter
Presenter: Ana Julia da Silva de Oliveira
doi://10.26678/ABCM.COBEM2023.COB2023-0853
Abstract
Chatter occurrence in machining can lead to component failure, production delays, and increased costs. To prevent these issues, interrupting machining before chatter occurs is essential. Hence, this project pursues to create an online tool for identifying the onset of chatter enabling an automatic interruption, aiming to reduce the consequences of chatter in turning and prevent waste and breakage. Machine learning algorithms can be used to predict and identify chatter, but they require a large amount of experimental data. Finding a suitable dataset can be challenging, and data augmentation may be necessary to increase the dataset size and robustness. In this study, the authors used a publicly available dataset consisting of 115 files with time series of accelerometer data, in which two uniaxial accelerometers were mounted on the tool holder itself, and a triaxial accelerometer was mounted on the base of the tool holder. The placement allowed for the measurement of vibration and acceleration during cutting tests. Data augmentation using windowing and noise was applied to the dataset, increasing the number of entries of data from the initial 115 to 728. The initial files were labeled using four classes based on the existence of chatter: "chatter", "without chatter", "intermediate" and "unknown". To effectively monitor and interrupt the process, it is ideal to use only two classes: "chatter" and "without chatter". Moreover, in the process of data augmentation, each time series entry was separated into four and their labels could be inconsistent due to the transient nature at the beginning of chatter, meaning that one entry could both have “without chatter” and “chatter” when separated. Therefore, a semi-supervised method for labeling through clustering was used to effectively separate the data into the desired two classes. Mathematical features both in the time and frequency domains were used for training and testing. Multiple machine learning algorithms were tested for efficiency and accuracy. GridSearchCV was used to find the optimized hyperparameters for all models. Results of accuracy as high as 98\% with 10-fold cross-validation were obtained with Random Forest. Overall, this study provides insights into the use of machine learning algorithms for identifying and preventing chatter in machining, using data augmentation and noise to increase the dataset size and quality. The results show promising potential for the use of relatively small datasets with missing labels, enabling a more efficient way of creating new monitoring tools for chatter.
Keywords
machine learning, Turning, data augmentation, machining, Chatter
