Establishing a Machine-learning Based Framework for Optimising Electronics Assembly

By the spread of miniaturized components, like the 0201mm size-code (200 × 100 μm) passives, utilizing advanced optimization techniques becomes crucial in this field. A framework was established, which used machine-learning-based estimators to predict the yield of any manufacturing process in electronics technology. The framework includes using various methods, like artificial neural networks (ANN), decision trees and neuro-fuzzy inference systems. It can automatically split the input data into training and testing sets for each learning epoch to reach optimal performance and prevent possible overfitting at the same time. Besides, optimal structures and description functions are also determined automatically. To assess the prediction error, the framework calculates the MAE (Mean Absolute Error), the RMSE (Root Mean Square Error) and the MAPE (Mean Absolute Percentage Error) parameters to decide if the built estimator structure is appropriate. As an outcome, the framework can provide several parameters that the user can optionally select. Parameters like the predicted values of a process output parameter over different input process parameters are provided. Besides, KPI (Key Process Index) of the output parameters or the Desirability Function (which combines many output parameters) can be acquired. The applicability and the performance of the framework were analyzed on the stencil printing process by building an ANN structure.