Multivariate Statistical Process Control Charts and the Problem of Interpretation: A Short Overview and Some Applications in Industry

Woodall and Montgomery [35] in a discussion paper, state that multivariate process control is one of the most rapidly developing sections of statistical process control. Nowadays, in industry, there are many situations in which the simultaneous monitoring or control, of two or more related quality - process characteristics is necessary. Process monitoring problems in which several related variables are of interest are collectively known as Multivariate Statistical Process Control (MSPC).This article has three parts. In the first part, we discuss in brief the basic procedures for the implementation of multivariate statistical process control via control charting. In the second part we present the most useful procedures for interpreting the out-of-control variable when a control charting procedure gives an out-of-control signal in a multivariate process. Finally, in the third part, we present applications of multivariate statistical process control in the area of industrial process control, informatics, and business.

💡 Research Summary

The paper provides a comprehensive overview of Multivariate Statistical Process Control (MSPC), organized into three distinct sections that together form a practical roadmap for both researchers and industry practitioners. The first part outlines the fundamental steps required to implement MSPC, beginning with the selection of relevant process variables and the necessary data preprocessing. It emphasizes the importance of normalizing variables, handling multicollinearity, and estimating a stable covariance matrix, noting that adequate sample size and regularization techniques are essential for reliable parameter estimation. The authors then introduce the most widely used multivariate control charts: Hotelling’s T² chart, the Multivariate Exponentially Weighted Moving Average (MEWMA) chart, and variants that incorporate dynamic weighting schemes. For each chart, the mathematical formulation, assumptions, and typical control‑limit calculations are presented, along with guidance on choosing the appropriate chart based on process characteristics such as the number of variables, sampling frequency, and expected shift patterns.

The second part tackles the critical issue of interpreting an out‑of‑control signal. While multivariate charts can detect a shift in the overall process, they do not directly reveal which specific variable(s) caused the alarm. To address this, the paper reviews several diagnostic tools. The contribution plot (or variable‑wise T² decomposition) visualizes each variable’s share of the overall statistic, offering an intuitive first glance. Variable‑by‑variable T² tests provide a more formal significance assessment but can suffer from inflated Type I error when variables are highly correlated. Principal Component Analysis (PCA) based methods, including score plots and loading plots, allow practitioners to see whether the shift is aligned with particular principal components, thereby highlighting groups of correlated variables. More recent approaches such as LASSO‑based variable importance, Bayesian networks, and partial least squares regression are discussed as ways to capture both main effects and interactions. The authors compare these techniques in terms of computational complexity, interpretability, and robustness to multicollinearity, concluding that a combined use of at least two complementary methods (e.g., contribution plot plus PCA) yields the most reliable diagnosis.

The third part showcases real‑world applications across three domains: manufacturing, information technology, and business analytics. In a chemical production line, temperature, pressure, flow rate, and catalyst concentration are jointly monitored using a Hotelling’s T² chart. When an out‑of‑control point occurs, a PCA‑based contribution analysis quickly identifies temperature as the primary driver, enabling rapid corrective action. In a data‑center environment, CPU utilization, memory usage, disk I/O, and network traffic are tracked with a MEWMA chart. An alarm triggers a weighted contribution analysis that isolates a memory leak on a specific server, preventing a potential service outage. Finally, in a corporate performance setting, sales, customer satisfaction, inventory turnover, and marketing spend are placed on a multivariate control chart (often called a Management‑by‑Statistical‑Control chart). The contribution plot reveals that excessive marketing expenditure is the main factor behind a dip in sales, prompting a strategic budget reallocation. Across all cases, the authors stress the necessity of validating the normality of the data (e.g., using Shapiro‑Wilk or multivariate normality tests), fine‑tuning control‑limit parameters (α levels, λ smoothing factors) through simulation, and employing complementary charts such as the Squared Prediction Error (SPE) chart to detect structural changes that T² alone might miss.

In addition to methodological guidance, the paper offers practical recommendations for reducing false alarms: adopt staged sampling schemes, adjust chart update frequencies to match process dynamics, and integrate visualization tools (heat maps, interactive dashboards) to aid frontline operators in rapid decision‑making. By systematically linking chart design, diagnostic interpretation, and industry case studies, the article bridges the gap between MSPC theory and its implementation in modern production and service environments. It serves as a valuable reference for organizations seeking to upgrade their quality‑monitoring systems from univariate to sophisticated multivariate frameworks, ensuring early detection of process shifts and facilitating timely, data‑driven corrective actions.