Research On Performance Prediction And Control Of Semiconductor Chip Packaging And Testing Production Line

In order to adapt to the ever-changing market,it is essential for the semiconductor industry to analyze,optimize,and control production line performance.At present,there are few researches on intelligent,comprehensive and dynamic control of production line performance,most of which are limited to a certain aspect of production line variability;the performance prediction model of the semiconductor series and parallel production line established in the current research has a certain deviation from the actual production situation,and the accuracy is lacking;traditional performance control optimization methods are difficult to perform real-time control on the changes of the production line variability factors,and the flexibility is insufficient.This thesis took the semiconductor chip packaging and testing series and parallel production lines as the research object,and improved on the defects of incomplete consideration of the variability factors in the existing research,insufficient accuracy of the performance prediction model,lack of real-time and flexibility of performance control,and incomplete control strategy A method for performance control based on performance prediction and sensitivity analysis with the help of reinforcement learning algorithms was proposed.This thesis focused on the following aspects:First,based on the queuing theory,comparatively analyzed the shortcomings of the current commonly used queuing models,and innovatively abstracted the series and parallel production line for packaging and testing semiconductor chips into G / G / m / b production line queuing models;based on the variability theory,comprehensively considered various variability factors such as processing variability,flow variability,and defective product rate and quantify them.and the production rate TH,production cycle CT and WIP level were used as performance prediction indicators to establish a more accurate production line performance and benefit prediction model.Secondly,Morris screening legality analysis and Arena simulation quantitative analysis were used to carry out sensitivity analysis.Used Morris screening method to find some key factors that had the greatest impact on the performance of semiconductor chip packaging and testing production lines,such as the buffer capacity,and established Arena simulation model to quantitatively analyze its influence law.Thirdly,this thesis used the performance prediction model of semiconductor chip packaging and testing production line as the external environment of reinforcement learning method,established a semiconductor production line performance control model based on Q-learning algorithm,studied and set the key parameters of reinforcement learning algorithm.After the occurrence of the volatile event,solved iteratively with the minimization of the production line efficiency index Bf as the performance control objective,and the global optimal performance control strategy for different factors such as buffer capacity,processing batch size,and number of parallel devices was obtained.Finally,a production scheme containing two types of products was designed,a performance control model based on the Q-learning algorithm was verified and analyzed,and a "two-step" production line performance control strategy was proposed.