Monitoring Multivariate Categorical Processes in Phase II: An Economic–Statistical Design of the GLT Control Chart

Technical Summary: Monitoring Multivariate Categorical Processes in Phase II: An Economic–Statistical Design of the GLT Control Chart

Problem Statement
Contemporary manufacturing systems increasingly rely on multiple interrelated categorical variables (e.g., defect classifications, product grades) to characterize product quality. While Statistical Process Control (SPC) is standard for maintaining stability, traditional control charts are often designed for univariate continuous variables and may fail to capture the dependency structures inherent in multivariate categorical data. Furthermore, existing literature on monitoring such processes, particularly those utilizing contingency tables and Generalized Linear Tests (GLT), has predominantly focused on statistical performance metrics (e.g., Average Run Length) while neglecting economic considerations. Consequently, there is a lack of integrated frameworks that simultaneously address the statistical sensitivity required to detect small, gradual shifts in dependency structures and the economic efficiency necessary for practical industrial implementation.

Methodology
The study proposes an integrated economic–statistical framework for Phase II monitoring of multivariate categorical processes represented by contingency tables. The methodology consists of three primary components:

1. Statistical Modeling and Monitoring:

   • Log-Linear Modeling: The dependency structure among categorical variables is modeled using log-linear models to establish expected cell frequencies under in-control conditions.
   • GLT Statistic: A Generalized Linear Test (GLT) statistic, denoted as $F^*$, is derived to quantify the deviation between observed cell frequencies and those predicted by the log-linear model. This statistic serves as the primary measure for detecting structural changes in the contingency table.
   • EWMA Integration: To enhance sensitivity to small and gradual shifts, the GLT statistic is incorporated into an Exponentially Weighted Moving Average (EWMA) structure. The EWMA statistic ($Z_t$) combines current and historical information, with a smoothing parameter ($L$) determining the weight assigned to past data. The control chart signals an out-of-control state when $Z_t$ exceeds a calculated Upper Control Limit (UCL).

2. Economic Formulation:

   • The study adopts an Expected Total Cost (ETC) function to evaluate the economic performance of the chart. The ETC aggregates costs associated with sampling, inspection, false alarms, searching for and removing assignable causes, and production losses during out-of-control periods.
   • The cost model accounts for the probability of false alarms (derived from the in-control Average Run Length, $ARL_0$) and the expected delay time in detecting shifts.

3. Optimization Framework:

   • An optimization procedure is developed to jointly determine the design parameters: sample size ($n$), sampling interval ($h$), control limit coefficient ($K$), smoothing parameter ($L$), and the resulting UCL.
   • The objective is to minimize the ETC subject to a constraint on the in-control Average Time to Signal ($ATS_0$), ensuring a predefined level of statistical performance (target $ARL_0 \approx 200$).
   • The optimization assumes a fixed inspection rate ($r$), where $h = n/r$, to reflect operational constraints common in industrial environments.

Key Contributions
The paper identifies four specific contributions to the field of quality engineering:

1. Extension to Categorical Data: It extends the economic–statistical design paradigm to multivariate categorical processes represented by contingency tables, an area previously underrepresented in the literature.
2. Unified Framework: It establishes a unified approach that integrates log-linear models, GLT statistics, and EWMA memory structures, rather than treating these components in isolation.
3. Simultaneous Optimization: It simultaneously considers economic and statistical objectives during the parameter optimization process to derive efficient chart configurations.
4. Practical Tool: It provides a practical monitoring strategy for industrial environments characterized by multiple correlated categorical quality characteristics.

Results
The proposed framework was evaluated using Monte Carlo simulations in MATLAB with a $4 \times 4$ contingency table and a total sample size of $N=500$. Key findings include:

• Sensitivity to Interaction Effects: The EWMA–GLT chart demonstrated significantly higher sensitivity to changes in the interaction parameters of the log-linear model (which represent the dependency structure) compared to overall or main effects. The "Effect Index" for interaction shifts was substantially higher, confirming the chart's ability to detect structural changes in variable dependencies.
• Optimal Design Parameters: The economic–statistical optimization identified the optimal configuration as a sample size of $n=50$, a sampling interval of $h=1$, a smoothing parameter of $L=0.30$, and a control limit coefficient of $K=1.50$. This configuration yielded a minimum Expected Total Cost (ETC) of 77.786 while maintaining an in-control $ARL_0$ of 201.42.
• Comparative Performance:
  • Compared to a purely statistical design, the economic–statistical design achieved lower operational costs while satisfying statistical constraints.
  • Compared to a purely economic design (which minimized cost without statistical constraints), the proposed framework avoided excessive false alarms ($ARL_0$ of 45.95 in the unconstrained economic case vs. 201.42 in the proposed design).
  • The EWMA–GLT chart outperformed both the standalone GLT chart and the classical EWMA chart in detecting small to moderate shifts, as evidenced by lower Average Time to Signal (ATS) values.
• Sensitivity Analysis: The model was found to be most sensitive to variations in the cost of out-of-control production ($C_d$) and the variable sampling cost ($C_1$). The smoothing parameter ($L$) had a more significant impact on detection speed and cost than the control limit coefficient ($K$); increasing $L$ reduced both $ARL_1$ and ETC.

Significance and Claims
The paper claims that the proposed EWMA–GLT framework offers an effective and practical strategy for monitoring multivariate categorical processes. By balancing economic and statistical objectives, the approach provides a more robust solution for industrial applications than designs relying exclusively on one criterion. The study concludes that the integrated framework successfully detects structural changes in contingency tables, particularly those involving interaction effects, while minimizing operational costs. The authors position the work as a decision-support tool for environments with correlated categorical data and suggest future research directions involving higher-dimensional tables, uncertainty modeling (fuzzy/robust approaches), and the use of real industrial datasets.