Multivariate Statistical Process Control (Multivariate SPC)

Unlike univariate SPC, which monitors one variable at a time, Multivariate SPC analyzes multiple correlated quality characteristics together. This is critical because monitoring correlated variables individually can lead to Type I errors (false alarms) or Type II errors (missed signals). For example, in a chemical process, temperature and pressure are often inversely correlated; a shift in one may be normal if accompanied by a compensatory shift in the other. Multivariate SPC models these relationships to provide a single, unified control signal.

The primary tool for monitoring the process mean vector is Hotelling's T-squared (T²) statistic. It measures the standardized statistical distance of a multivariate observation from the historical process mean, accounting for correlations.

• Calculation: T² = (x - x̄)' S⁻¹ (x - x̄), where x is the observation vector, x̄ is the historical mean vector, and S⁻¹ is the inverse of the covariance matrix.
• Visualization: The T² value is plotted on a Hotelling's T² control chart. A point exceeding the upper control limit indicates a statistically significant shift in the multivariate process mean.
• Use Case: Monitoring the diameter, length, and weight of a manufactured part as a single entity.

To detect small, persistent shifts, Multivariate SPC employs advanced charts that model the covariance structure of the data:

• Multivariate Exponentially Weighted Moving Average (MEWMA) Chart: Applies exponentially decreasing weights to past observations, making it sensitive to small shifts in the mean vector. It is particularly useful when observations are autocorrelated.
• Multivariate Cumulative Sum (MCUSUM) Chart: Plots the cumulative sum of deviations from a target vector, excelling at detecting sustained drifts. It is often implemented using a generalized likelihood ratio test.

These methods are more powerful than the basic T² chart for detecting subtle process changes.

When monitoring many variables (e.g., 50+ sensor readings), multicollinearity and high dimensionality make traditional T² charts inefficient. Principal Component Analysis (PCA) is used to:

• Reduce Dimensionality: Transform the original correlated variables into a smaller set of uncorrelated principal components (PCs) that capture most of the process variance.
• Create Two Charts: A T² chart monitors variation within the PCA model (the dominant patterns), while a Q statistic (or SPE chart) monitors the residual variation not captured by the model, which can indicate new types of faults.
• Benefit: This separates common cause variation (in the model) from novel special cause variation (in the residuals).

A key challenge is diagnosis: when a Multivariate SPC chart signals an out-of-control condition, which variable(s) caused it? Contribution plots are the primary diagnostic tool.

• Function: After a T² or Q statistic alarm, a contribution plot decomposes the signal to show how much each original variable contributed to the exceedance.
• Interpretation: Variables with the largest contributions are the most likely root causes of the process shift.
• Example: In a batch reactor, a T² alarm might be diagnosed via contribution plots showing that temperature and pH contributed 60% and 30% to the signal, respectively, guiding the investigation.

Multivariate SPC is directly applicable to modern data and ML systems, moving beyond manufacturing:

• Data Quality Monitoring: Monitoring multiple correlated data quality metrics (e.g., null rate, value distribution, string length) for a table simultaneously to detect holistic data drift.
• Model Performance Monitoring: Tracking a vector of model performance metrics (accuracy, precision, recall, F1-score) together to detect degradation more reliably than any single metric.
• ML Feature Drift: Using PCA-based T² and Q charts to monitor the joint distribution of hundreds of model input features for covariate shift.
• Pipeline Health: Monitoring correlated system metrics (CPU, memory, latency, error rate) for a microservice to detect infrastructure anomalies.

Hotelling's T-squared chart is the foundational multivariate control chart for monitoring the mean vector of a process. It calculates a single composite statistic (T²) from multiple correlated variables, which is then plotted against a statistically derived control limit.

• Purpose: Detects shifts in the multivariate process mean.
• Mechanism: Combines the Mahalanobis distance of an observation from the historical mean vector.
• Key Advantage: Accounts for correlation between variables, preventing the misleading alarms that occur when monitoring each variable independently with univariate charts.
• Example: In chemical manufacturing, it monitors temperature, pressure, and pH simultaneously to detect a process upset that might not be evident in any single variable's chart.

The Multivariate Exponentially Weighted Moving Average (MEWMA) chart is a control chart designed to detect small, persistent shifts in the mean vector of a multivariate process. It applies exponentially decreasing weights to past observations, making it more sensitive to minor drifts than the Hotelling's T-squared chart.

• Purpose: Early detection of small, sustained multivariate mean shifts.
• Mechanism: Computes a weighted average of current and past T² statistics, with recent data given more influence.
• Use Case: Ideal for continuous processes in semiconductor fabrication or pharmaceutical production where early warning of a subtle process drift is critical to prevent costly scrap.

Principal Component Analysis (PCA) is a dimensionality reduction technique often used with Multivariate SPC to handle high-dimensional data. It transforms the original correlated variables into a smaller set of uncorrelated principal components that capture most of the process variation.

• Purpose: Reduces complexity and isolates dominant sources of variation.
• Application in SPC: Control charts (like T² and SPE) are then applied to the scores of the principal components rather than the original variables.
• Key Benefit: Helps distinguish between variation in the model's subspace (common cause) and variation orthogonal to it (potential special causes), aiding in fault diagnosis.

The Squared Prediction Error (SPE) Chart, also known as the Q-statistic chart, monitors the residual variation not explained by a PCA model. It works in tandem with a T² chart on the principal components to provide a complete view of multivariate process health.

• Purpose: Detects the emergence of new types of faults or variation patterns not present when the PCA model was built.
• Interpretation: A high SPE value indicates the observation does not fit the established correlation structure of the normal operating data.
• Diagnostic Pair: Used with a T² chart; a fault may be evident in one chart but not the other, helping to classify the type of process disturbance.

Multivariate Process Capability extends the univariate concept of Cp and Cpk to assess a process's ability to produce output where multiple, correlated characteristics simultaneously meet specification limits. It evaluates the proportion of output expected to fall within a multivariate specification region.

• Challenge: A process can have excellent univariate capability for each variable but still produce a high rate of overall defects due to the correlation structure.
• Common Indices: Multivariate Capability Index (MCp) and Multivariate Performance Index (MPpk).
• Visualization: Often represented using confidence ellipses plotted against rectangular specification boxes on a scatter plot of two key variables.

Canonical Variate Analysis (CVA) is a multivariate statistical method used for dynamic process monitoring and fault diagnosis. It is particularly effective for processes with serial correlation (time-series data), where it finds linear combinations of past observations that are most correlated with future states.

• Purpose: Monitoring and diagnosing faults in dynamic, time-dependent processes like batch operations or continuous chemical reactors.
• Advantage over PCA: Explicitly models the dynamic relationship between variables over time, making it more sensitive to specific fault signatures.
• Application: Widely used in the Partial Least Squares (PLS) framework for monitoring batch processes where data unfolds in three dimensions (batches x variables x time).