The SNN-Based Predictive Model for HGA Manufacturing

Product types in the hard disk drive (HDD) industry have different specifications depending on the customer orders.  These specifications along with the machine parameters have a direct impact on the production yield.  The problems on the manufacturing line are called the  croot causee. By accurately identifying the root cause, the engineers can suggest yield improvement solutions.  Our research focus on a design of the effective prediction technique required at the end of the analysis steps in order to validate the suggested solution by simulation. Initially, we developed a cCe program based on a multiple regression technique and tested it's validity toward prediction accuracy. From the experimental results, we could conclude that the multiple regression method (up to 10 polynomial degrees) did not produce a sufficiently good result. The method gave a high error rate for fault prediction.  The Hard Gimbal Assembly (HGA) yield prediction is proved to be non-linear. Therefore, we adapted the Stochastic Neural Networks (SNNs) for use with the yield prediction. The inputs of our model consist of several machine parameters and specification attributes. Our version of SNNs can approximate a complex non-linear system. The genetic algorithm is used as a learning algorithm instead of the backpropagation method in order to handle the non-linear and stochastic relationships between input parameters. Our prediction model can then be used to validate and revise the yield improvement plan. The output of the prediction model is the yield rate. From SNNs' results, we can conclude that our initial version of SNNs gave a favorable prediction results with very low error rates.  The model can thus be used as a simulation tool for yield improvement without having to actually implement the solution on the production line. Keywords : Stochastic Neural Networks / Multiple Regression /Yield Prediction Technique