📊 statistics

Improved MGLR control charts based on Klein’s sensitizing rules for joint mean and dispersion monitoring under measurement error

This paper proposes and evaluates two enhanced multivariate generalized likelihood ratio (MGLR) control charts that integrate Klein's 2-of-2 and 2-of-3 sensitizing rules to effectively monitor simultaneous shifts in process mean and dispersion while explicitly accounting for the detrimental impact of measurement error.

Original authors: Mohammad Reza Maleki, Mohammad Nikoogoftar, Hossein Eghbali, Majid Jafari

Published 2026-09-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Reza Maleki, Mohammad Nikoogoftar, Hossein Eghbali, Majid Jafari

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quiet hum of a modern factory, the quality of a product is rarely defined by a single number. Instead, it is a complex tapestry woven from many different threads: the thickness of a metal sheet, the tension of a wire, the smoothness of a surface, and the alignment of a component. These characteristics do not exist in isolation; they are deeply connected, shifting and reacting to one another as the machinery runs. To keep production running smoothly, engineers use statistical tools called control charts. These are essentially monitors that watch the process, sounding an alarm if the numbers drift away from their intended targets. For decades, the standard tools have been excellent at spotting massive, obvious problems, like a machine breaking down completely. However, they often miss the subtle, creeping changes—the slight warping of a part or a tiny increase in vibration—that happen slowly over time. These small shifts are dangerous because they can ruin a batch of products before anyone notices, yet they are too faint for the old tools to catch.

Adding to this challenge is the reality that the measurements themselves are never perfect. The instruments used to check the products, the people operating them, and the environment around them all introduce small errors. A ruler might be slightly bent, a sensor might drift with temperature, or a human operator might read a gauge a fraction of a millimeter off. These measurement errors act like static on a radio signal, blurring the true picture of what is happening on the factory floor. When this static is ignored, the monitoring systems become sluggish, taking much longer to realize that something is wrong. The question facing researchers has been how to build a monitoring system that is sharp enough to catch these tiny, simultaneous changes in multiple characteristics, even when the data being fed into it is imperfect.

A team of researchers has addressed this problem by refining a sophisticated monitoring method known as the multivariate generalized likelihood ratio chart. This tool is designed to watch several related quality features at once, looking for changes in their average values and how much they vary. While powerful, the standard version of this chart can still be slow to react to small disturbances. To fix this, the researchers introduced two new layers of logic, inspired by a set of rules developed by a statistician named Klein. These rules act like a second pair of eyes. Instead of waiting for a single, dramatic spike to trigger an alarm, the new system looks for patterns. One version, called the "2-of-2" rule, sounds the alarm if two consecutive measurements show signs of trouble. The other, the "2-of-3" rule, is slightly more nuanced, triggering an alert if two out of the last three measurements indicate a problem. By adding these pattern-recognition steps, the system becomes much more sensitive to the early warning signs of a process going off-track.

The researchers did not stop at simply making the chart smarter; they also rebuilt it to account for the inevitable noise of measurement error. In the real world, the numbers recorded by inspectors are often a mix of the true product value and the error introduced by the measurement process. The team developed a mathematical framework that treats these errors as a known factor, allowing the chart to separate the signal from the noise more effectively. They then put their new designs to the test using a vast number of computer simulations. These simulations created thousands of imaginary factory scenarios, introducing small shifts in the process mean and variability, and layering in different levels of measurement error to see how the charts would perform.

The results were clear and consistent. In every scenario tested, the presence of measurement error made the monitoring systems slower to react, increasing the time it took to detect a problem. This delay happened even when the measurement errors were quite small, confirming that ignoring this factor leads to a false sense of security. However, the charts equipped with the new pattern-recognition rules significantly outperformed the traditional version. They detected the subtle shifts much faster, cutting down the time a defective process could run undetected. Between the two new versions, the "2-of-3" rule performed slightly better than the "2-of-2" rule, offering the quickest response to changes in both the average values and the variability of the process. The study found that these improvements held true whether the problem was a shift in the average size of a part, a change in how much the parts varied, or a simultaneous change in both.

This work demonstrates that by combining smarter pattern recognition with a realistic understanding of measurement limitations, it is possible to create a much more vigilant safety net for manufacturing. The new charts do not just react to big failures; they are tuned to hear the whisper of a small change before it becomes a shout. For industries where precision is paramount, from aerospace to pharmaceuticals, this ability to catch errors early and accurately, even when the data is imperfect, translates directly into fewer defective products and more reliable quality. The researchers suggest that while their current model focuses on one specific type of measurement error, future work could expand these methods to handle even more complex sources of uncertainty, further sharpening the tools that keep our modern world running smoothly.

Technical Summary: Improved MGLR Control Charts Based on Klein's Sensitizing Rules for Joint Mean and Dispersion Monitoring Under Measurement Error

Problem Statement
In modern manufacturing and service systems, product quality is often defined by multiple correlated characteristics. While multivariate control charts are essential for simultaneous monitoring, existing methods face three primary limitations. First, traditional Shewhart-type charts, though simple, lack sensitivity to small process shifts. Second, many assignable causes simultaneously affect both the process location (mean vector) and dispersion (covariance matrix), yet many existing schemes monitor these parameters separately. Third, despite advancements in data collection, measurement errors (due to instrument calibration, operator inconsistency, or environmental factors) remain inevitable. These errors introduce additional variability that contaminates observations, often leading to a significant reduction in the performance of statistical process control charts, a factor frequently overlooked in chart design.

Methodology
The paper proposes a unified framework to address these challenges by developing enhanced Multivariate Generalized Likelihood Ratio (MGLR) control charts. The methodology proceeds in three stages:

Development of Enhanced MGLR Charts: The authors integrate Klein's sensitizing rules (specifically the 2-of-2 and 2-of-3 rules) into the standard MGLR statistic. These rules are designed to detect non-random patterns and small shifts more rapidly without increasing the structural complexity of the chart statistic.
- 2-of-2 Rule: Signals an out-of-control condition if two consecutive points exceed the control limit.
- 2-of-3 Rule: Signals an out-of-control condition if two out of three consecutive points exceed the control limit.
- The charts are designed to monitor the mean vector ( $\boldsymbol{\mu}$ ) and covariance matrix ( $\boldsymbol{\Sigma}$ ) simultaneously.
Incorporation of Measurement Error: To reflect realistic industrial conditions, the improved charts are redesigned under an additive covariate model. In this model, the observed value ( $\boldsymbol{y}_{tj}$ ) is expressed as a linear function of the true quality characteristic ( $\boldsymbol{x}_{tj}$ ) plus a random error vector ( $\boldsymbol{\epsilon}_{tj}$ ):
$\boldsymbol{y}_{tj} = \boldsymbol{A} + \boldsymbol{B}\boldsymbol{x}_{tj} + \boldsymbol{\epsilon}_{tj}$
The error vector is assumed to follow a multivariate normal distribution with mean zero and covariance matrix $\boldsymbol{\Sigma}_e$ . The proposed charts utilize standardized observations derived from this model to compute the MGLR statistic, thereby explicitly accounting for the variance introduced by measurement systems.
Performance Evaluation: The effectiveness of the proposed schemes (Classical MGLR, MGLR-2-of-2, and MGLR-2-of-3) is evaluated using extensive Monte Carlo simulations. The study considers:
- Scenarios: Separate shifts in the mean vector, separate shifts in the covariance matrix, and simultaneous shifts in both.
- Error Levels: Four levels of measurement error variance ( $\sigma^2_\epsilon \in \{0, 0.10, 0.15, 0.20\}$ ).
- Metrics: Average Run Length (ARL), Standard Deviation of Run Length (SDRL), and Median Run Length (MRL). The control limits for all charts are calibrated to achieve an in-control ARL ( $ARL_0$ ) of approximately 200.

Key Results
The simulation results yield the following findings:

Impact of Measurement Error: Measurement error consistently degrades the performance of all control charts. Even at low error variance levels, the presence of error increases the ARL values, indicating a delay in detecting out-of-control signals. This negative impact intensifies as error variance increases and, in the case of the classical chart, as the magnitude of the shift increases.
Superiority of Sensitizing Rules: The improved charts incorporating Klein's rules (2-of-2 and 2-of-3) substantially outperform the classical MGLR chart across all shift scenarios and error variance levels. They achieve shorter run lengths, signifying faster detection of process disturbances.
Comparison of Rules: While both improved schemes are superior to the classical chart, the 2-of-3 scheme demonstrates slightly better overall performance than the 2-of-2 scheme in terms of detection speed.
Behavior Under Error: Interestingly, the behavior of the improved charts differs slightly from the classical chart regarding the interaction between shift magnitude and error. For the improved charts, the percentage increase in ARL due to measurement error tends to increase for small-to-moderate shifts but may decrease for very large shifts, whereas the classical chart shows a consistent increase in ARL percentage as shift magnitude grows.

Significance and Claims
The paper claims that no previous study has integrated multivariate joint monitoring of mean and dispersion, Klein's sensitizing rules, and measurement error effects within a single framework. The significance of this work lies in:

Practical Realism: By explicitly modeling measurement error via an additive covariate model, the proposed charts offer a more realistic representation of industrial monitoring conditions where data is rarely error-free.
Enhanced Sensitivity: The incorporation of Klein's rules successfully accelerates the detection of small and moderate shifts in both process location and dispersion, addressing a known weakness of standard Shewhart-type multivariate charts.
Robustness: The study demonstrates that while measurement error inevitably reduces sensitivity, the proposed enhanced schemes maintain a significant performance advantage over classical methods, making them more robust choices for environments with imperfect measurement systems.

The authors conclude that neglecting measurement error can lead to misleading assessments of control chart performance and that the proposed 2-of-3 MGLR chart offers a particularly effective solution for simultaneous monitoring in the presence of measurement noise. Future research directions suggested by the authors include extending the model to a four-component error structure and investigating the economic and economic-statistical design of these charts.

More like this