Predicting Component Self-alignment by Machine Learning Technique

In this paper, a method was developed for predicting the component self-alignment within the liquid state of reflow soldering. The method is based on utilizing neural network as a predictor. The neural network was built containing three input and three output parameters, which are the x, y and θ offset before and after soldering, respectively. Experimental was carried out for analysing the prediction error of our method. In the experiment, an FR4 based test board was designed, which allows the measurement of component placement offsets by utilizing registration marks (fiducial points). 0603 size (1.5× 0.75 mm) chip resistors were placed onto the test board with intentional component misplacement in x (0-450 μm) and in y (0-350 μm) direction and with intentional rotational offset (0-20°). The position of the components was measured both before and after soldering, and the extent of component movement during reflow soldering was calculated. One part of the samples was used for training (training set) the neural network, whereas the other part was used for testing (testing set) and for determining the prediction error. The results proved that the self-alignment ability of components can be effectively predicted by using neural networks.