Otimização paramétrica de redes neurais artificiais aplicadas em processo de soldagem com datasets reduzidos

Abstract

Industry 4.0 has increasingly adopted digital solutions such as Machine Learning and Deep Learning algorithms to improve manufacturing processes, such as welding. Regulating welding parameters is essential to achieve high-quality weld geometries, but it can be a challenging task for operators in the field. A promising approach is to use parameter prediction techniques from specific geometries, with artificial neural networks (ANNs) being one of the most prominent. These networks have a functioning inspired by the learning observed in living beings. To determine the ideal ANN architecture, including layers, neurons, activation function and learning rate, many authors resort to trial-and-error selection. This work aims to develop a methodology to optimize the parameters of an ANN architecture without causing overfitting, in addition to promoting comparison with regression techniques. To address data limitation, advanced data generation and cross-validation techniques are explored, such as Leave One Out (LOO), Kriging and Synthetic Data Vault (SDV). The objective of the ANN is to predict values such as: voltage, gas flow, feed speed and stickout, based on geometric information resulting from the MAG welding process on SAE 1020 steel. The data are divided into training, validation and test sets to ensure that the ANN architecture is defined in the validation phase and kept constant during testing, avoiding overfitting. In the optimization process, differential evolution (DE) was chosen compared to simplicial homology (SHGO) because it presents better results. A study of the number of additional data generation for the SDV and Kriging techniques was proposed, increasing the standard dataset by 50%, 100% and 150%. The best results were found when increasing the data by 50% by SDV for the ANN through 1 hidden layer, 2 neurons and logistic activation function. Artificially augmenting the dataset improved the predictive accuracy of the ANN, with SDV showing a 10.1% improvement over the standard condition with the original dataset and ED, and LOO achieving an error reduction of 55.3%. The logistic activation function was found in most of the best results. All results presented by regression techniques are inferior to those presented by the ANN, however, there is also an improvement in the result from data augmentation through SDV. The work shows that using alternative techniques to improve performance when dealing with reduced datasets improved the prediction results.