Título: Parametric Optimization of Artificial Neural Networks and Machine Learning Techniques Applied to Small Welding Datasets
Autores: ROCHA, V. R.; LOBATO, F. S.; ASSIS, P. A. Q.; RIBEIRO, C. R.; CUNHA JR, S. S.; VILARINHO, L. O.; ANDRADE, J. R.; SILVA, L. R. R.; PAES, L. E. S.
Revista: Processes, 13, 9 (1-38), 2025.
Resumo: Establishing precise welding parameters is crucial to attain the desired bead geometry, adhering to stringent quality standards. However, configuring these settings can pose challenges in the manufacturing process. Addressing this issue, various methods are utilized to derive parameters from specified geometry, with artificial neural networks (ANNs) emerging as a prominent approach. Determining the optimal ANN architecture involves defining layers, units, activation functions, and crucial hyperparameters, typi-cally employing a trial-and-error method. Generating a sufficiently large dataset for effective ANN training remains a challenge, which complicates architecture definition. In this context, the present study proposes leveraging differential evolution (DE) to op-timize the ANN architecture. Furthermore, cross-validation techniques like leave-one-out (LOO) and data augmentation methods (such as kriging and synthetic data vault (SDV)) are introduced to enhance ANN performance. Specifically, the pro-posed ANN predicts welding parameters—voltage, gas flow, feed speed, and stickout—based on desired bead geometry (plate thickness, bevel type, gap, height, and width) for metal active gas (MAG) welding on SAE 1020 steel. The original dataset contains 44 sets of data, and the new datasets were augmented at 1.5x, 2x and 2.5x rates for each technique. Comparative analyses between DE and simplicial homology global optimization (SHGO) demonstrate DE's superior performance with a mean average percentage error of 0.394 for DE and 0.408 for SHGO. Additionally, augmenting the data set size through artificial means improves the ANN's predictive accuracy on previously unseen data. The best result from SDV presented 10.1% and the LOO 55.3% of im-provement compared to the standard DE condition in a 1.5x data augmentation condition.

Acessar a Publicação.
Voltar para a seção Revistas.