Study on intelligent identification technology of solder joints defects based on LMBP neural network

Abstract

Using Surface Evolver to predict solder joint three dimensional shape of chip components, selecting solder joints characteristic parameters as well the possibility of feature combination corresponding to solder joints defects as training sample of LMBP neural network and simulating excessive solder and other defects to form training samples of solder joints defects based on solder joints virtual technology. We can achieve the table of solder joints quality characteristic parameters and training samples of solder joints intelligent identification by choosing the four key factors of characterizing solder joint defects: solder joint center cross-sectional area S, solder center wetting height H, solder center contact angle 91 and 92. Characteristic information of solder joints quality as network input sample and possibility of solder joints defects as network desired output. We construct the training model of LM network to conduct network training. When error e <0.01, we stop training and save network structure, weight and other parameters. Then neural network model of intelligent identification actual solder joints defects can be obtained. We can obtain 2-D image of actual chip solder joints by machine vision technology and achieve 3-D image of actual solder joints after solder joint 3D reconstruction. Solder joints cross-sectional shape can be obtained by display of sectional profiles. Utilizing the model of trained BP network to calculate the possibility of solder joints defects and compared with production statistic data. If the result does not beyond error range, we think the test is successful. So the network can be used defects intelligent identification of actual solder joints. Or we should choose samples again, adjust parameters of network structure and conduct new network training until success. The results show that we can see that possibility identification output of solder joints defects are close to the actual possibility. That shows it can meet basic requirements of intelligent identification precision and can use intelligent identification of actual solder joints defect. At the same, the results of intelligent identification will be as the input variable of intelligent analysis of solder joints defects.