Charles Chen, Master Black Belt, Continuous Improvement Expert, Applied Materials
Mason Chen, Black Belt Student, Stanford University OHS
Patrick Giuliano, JMP Technical Support Engineer, JMP
During the machine warming-up period, the manufacturing process can be unstable due to assignable causes. This presentation discusses how several platforms and Poisson statistics found in JMP 16 were used to detect the higher defect counts.
Both the continuous normal fit and discrete Poisson fit were utilized and compared to differentiate the unstable and stable process periods. It’s critical to search the rational subgroupings to separate common and special variations. Three platforms in JMP 16 were utilized to search the rational subgroupings:
- Time series analysis and lag/autocorrelation plots to detect the recurring defect count cycles.
- Hierarchical clustering analysis and constellation plots to group the homogeneous periods.
- T-Square principal component SPC and the change point detection to define the rational subgrouping.
All three techniques found the best rational subgrouping of every hour to detect the special machine warming-up patterns. Seven different control chart types were conducted on the defect count data. Only the Shewhart Xbar-R and Poisson U control charts (based on one-hour rational subgroupings) could detect the first hour assignable causes accurately. This one-hour rational subgrouping was utilized in the process capability to differentiate short-term Cp and long-term Pp. Poisson statistics and normal approximation were used to further estimate the process capability Cp. Process capability summary reports, goal plot and performance plot were utilized to help visualize the process capability across different rational subgroups. After eliminating this warming-up cause, the defect count upper spec limit was reduced by 20%.
