TL;DR
This paper introduces a novel Modified One-Sided EWMA (MOEWMA) control chart for monitoring the mean of a normal process, offering improved detection of shifts compared to existing one-sided EWMA charts. By incorporating Variable Sampling Intervals (VSI), the MOEWMA chart further enhances its efficiency, especially for detecting small shifts, and is validated with a real-world manufacturing dataset.
What Problem Does This Paper Address?
Statistical Process Monitoring (SPM) is crucial for quality control, aiming to detect deviations from a stable process. While two-sided EWMA charts are widely studied, one-sided EWMA charts, which are more suitable when the direction of a process shift is known, have received less attention. Existing one-sided EWMA methods often involve resetting the EWMA statistic to the target or truncating observations, which can lead to a loss of valuable historical data. This paper addresses the need for a more effective one-sided EWMA chart that retains historical information and improves shift detection capabilities, particularly for normally distributed processes.
Methods
The core of this research is the development and evaluation of the Modified One-Sided EWMA (MOEWMA) X-bar chart. Unlike traditional EWMA charts that might discard past information upon resetting, the MOEWMA chart's statistic is designed to incorporate all previously collected sample information. This is achieved through a specific resetting mechanism that aims to preserve data continuity.
The performance of the MOEWMA chart is rigorously assessed using Monte-Carlo simulations to calculate Average Run Length (ARL) and Standard Deviation of Run Length (SDRL) for both "zero-state" (shift occurs at the start) and "steady-state" (shift occurs after a period of in-control operation) scenarios. These results are then compared against established one-sided EWMA charts: SEWMA (Simple EWMA), REWMA (Resetting EWMA), and IEWMA (Information-preserving EWMA).
Furthermore, the paper extends the MOEWMA chart by integrating Variable Sampling Intervals (VSI). This involves defining upper and lower warning limits within the control limits. If the charting statistic falls within the warning region, a shorter sampling interval is employed, while it remains longer in the central region. This VSI approach aims to increase the chart's sensitivity to shifts by allowing for more frequent monitoring when the process is suspected to be at risk. The performance of the VSI MOEWMA chart is evaluated using Average Time to Signal (ATS) and Average Adjusted Time to Signal (AATS), comparing it against its Fixed Sampling Interval (FSI) counterpart.
Datasets
The study primarily relies on simulated data generated through Monte-Carlo simulations to evaluate the performance metrics (ARL, SDRL, ATS, AATS) of the proposed control charts under various shift scenarios.
A real-world dataset from a semiconductor manufacturing process, specifically monitoring the flow width of photoresist during the hard-baking process, is used to demonstrate the practical implementation and effectiveness of the proposed MOEWMA and VSI MOEWMA charts.
Key Insights and Ideas
- Preserving Information is Key: The MOEWMA chart's novel resetting mechanism effectively retains historical data, leading to better shift detection compared to charts that discard information. This is a crucial improvement over some existing one-sided EWMA methods.
- VSI Boosts Sensitivity: Integrating Variable Sampling Intervals (VSI) into the MOEWMA chart significantly enhances its ability to detect shifts, especially small ones, by allowing for more frequent sampling when the process shows signs of instability.
- Optimal Design Matters: While the MOEWMA chart generally performs well across a range of shifts, optimizing its parameters (λ and control limits) for a specific shift size can further improve its performance. However, the paper suggests that the MOEWMA chart's robustness across various shifts makes it a practical choice even without precise shift size knowledge.
- One-Sided Charts for Known Shift Directions: The research reinforces the value of one-sided control charts when the expected direction of a process shift is known, allowing for more targeted and efficient monitoring.
Key Results
- MOEWMA vs. Competitors: Extensive simulations show that the MOEWMA X-bar chart generally outperforms SEWMA, REWMA, and IEWMA charts, particularly in detecting small shifts. For instance, in zero-state scenarios, MOEWMA often yields lower ARL1 and SDRL1 values.
- VSI MOEWMA Superiority: The VSI MOEWMA chart consistently demonstrates better performance (lower ATS1 and AATS1) than its FSI counterpart, especially for small shifts. For a shift of δ=0.1, the VSI MOEWMA chart with λ=0.05 achieved an ATS1 of 32.52, significantly lower than the FSI MOEWMA's ATS1 of 60.12.
- Real-World Application: In a semiconductor manufacturing example, the MOEWMA and VSI MOEWMA charts successfully detected an assignable cause (a shift in photoresist flow width) more quickly than some traditional EWMA charts. The VSI MOEWMA chart, in particular, detected the shift in 3.8 time units.
Limitations and Future Work
- Assumption of Known Parameters: The current research assumes that the process parameters (mean and standard deviation) are known. The authors suggest that future work should investigate how the proposed charts perform when these parameters are estimated from data.
- Focus on Process Mean: This study exclusively focuses on monitoring the process mean. Future research could extend the methodology to monitor process variance, ratios of distributions, or other quality characteristics.
- Parameter Selection: While the paper provides guidance on selecting parameters like λ and the VSI parameters (p1, p2, hs, hL), further research could explore more automated or adaptive methods for parameter selection in real-time applications.
Summary updated