LAUSR

Abstract Technologies for inserting electronic components are necessary within the electronics industry. Previously this was done by manual assembly, but todays customized machines have been specially designed for automatic assembly. A number of these machines even employ robot arms to insert nonconventional components. However, because special-purpose machines are unable to insert transformers with six manually soldered pins onto printed circuit boards, this study proposed a learning system for such machines that incorporates image characteristics into the insertion motions performed by a robot arm to solve problems related to transformer insertion. The proposed system operates in three layers: vision, motion, and decision. The vision layer involves preprocessing image data, extracting pin image features by locally linear embedding (LLE), and setting parameters for teaching insertion motions to the robot arm. In the motion layer, motions qualified for inserting the transformers were collected and the weighted Fuzzy C-means was used to converge the insertion motions and create target markers for the decision layer. The decision layer uses one-against-rest support vector machines (SVMs) to establish classifiers for applying the collected image characteristics to the calculation of insertion motions. Experiments were performed to verify the various research methods by using 300 transformers as training samples and 200 transformers as test samples. By imposing a number of rules to limit image characteristics, this study applied three classifiers (SVMs, Bayes, and a neural network) to the test samples and compared their accuracy. The experimental results indicated an accuracy rate of 88%, an average area under the receiver operating characteristic curves of 0.88, and that the employed SVM classifiers were more accurate than the other two classifiers.