Control Charts for Autocorrelated Data

Autocorrelated process data can break ordinary Shewhart interpretation.
Nearby observations are not independent.
Time-series methods account for internal structure such as autocorrelation, trend, or seasonality.^[1]

Prerequisites

Prerequisites: I-MR charts and basic time-series autocorrelation.

Process Context

Autocorrelation is common in continuous chemical processes, sensor streams, temperature profiles, financial operations, and automated measurements sampled faster than the process can physically change.

Definition

For an autocorrelated series $y_t$ , a robust SPC approach models the predictable structure first and charts residuals:

The residual chart asks whether the unexplained part of the process changed.

Assumptions / Requirements

• Observations are time ordered and equally spaced, or the model explicitly handles spacing.
• Autocorrelation is diagnosed before ordinary control limits are trusted.
• The residual model is validated; residuals should be approximately stable and uncorrelated.
• Process knowledge supports the model and sampling interval.

Control Limits / Formula

For residuals with mean near zero and estimated residual standard deviation ${\hat{\sigma }}_e$ :

This residual chart formula is not universal for all autocorrelated processes.

Interpretation Rules

• A naive I-MR chart may false alarm or miss shifts when autocorrelation is strong.
• Chart the residuals only after the time-series model is checked.
• A residual signal means the process deviated from its expected dynamic behavior.
• Trends and seasonality should be modeled or removed before Shewhart-style limits are interpreted.

Worked Example

One furnace temperature series follows the forecast ${\hat{y}}_{t|t-1}=0.8y_{t-1}+20$ .
If one observation is $y_{t-1}=100$ , the forecast is $100$ .
If the next observed value is $104$ , the residual is $e_t=4$ .
With ${\hat{\sigma }}_e=1.2$ , the residual UCL is $3.6$ , so this residual is a signal.

Common Mistakes

• Applying I-MR limits to highly autocorrelated sensor data without diagnosis.
• Treating residual charting as valid without checking residual autocorrelation.
• Sampling too frequently and then blaming the chart for false alarms.
• Removing trend without preserving the operational meaning of the signal.

Connections

References