Reference control model for predicting screen printer quality

In electronic manufacturing, it is common practice to assure the quality of the final product by performing time-consuming and expensive 100%-tests of finished subassemblies. Integrated test strategies are becoming more important due to the complex process steps and interactions in electronics production and the many parameters that affect the final quality of the products. Today\´s Surface Mount Technology (SMT) assembly equipment is able to generate huge volumes of data for identification of machine process quality parameters, but most of this information is currently underutilized on the manufacturing floor. Each process step, e.g, component assembly, can be characterized by twelve to thirty parameters. However, it is not easy to analyze these parameters. One promising technique is machine learning based on Data Mining techniques. Using neural networks, many parameters can be compared, and a process model can be generated, even when using "raw " unfiltered process data, as demonstrated in this paper for a solder printing process, characterized by sixteen parameters. The important input parameters, as identified by a neural network model, have been collected either on-or offline to create the core database for extracting the reference model. Each of the datasets contains the x/y/theta-correction, idle time up/down and pre-/post-pressure as well as the average pad-stack coverage. Different test scenarios were run to identify the best-fit reference machine model, This paper compares the results obtained with different configurations of the machine-learning algorithm, in terms of their prediction accuracy, identified model parameters, and model structure. Reference models with up to 99% accuracy have been obtained for actual production scenarios.