Western Electric Rules

Applied to control charts tracking key pipeline metrics like row counts, null percentages, or data freshness. The rules automatically flag special cause variation indicating pipeline failures.

• Rule 1 (Point beyond limits): Triggers if daily ingested record count falls outside the 3-sigma control limits, signaling a complete source failure or duplication event.
• Rule 2 (Run of 9): Detects if the median value of a critical column has been trending above its historical mean for 9 consecutive batches, indicating a gradual schema drift or upstream logic change.
• Rule 4 (2 of 3 beyond 2σ): Alerts when two out of three consecutive data quality scores (e.g., completeness %) fall beyond 2 standard deviations from the mean, catching intermittent but severe corruption.

Used to monitor production machine learning model metrics (accuracy, F1-score, prediction latency) over time. Shifts signal degrading performance due to data drift or concept drift.

• Rule 1: A single day's model accuracy on a holdout set drops below the lower control limit, indicating a catastrophic failure or corrupted evaluation data.
• Rule 3 (Trend of 6): The model's inference latency shows six consecutive increasing measurements, flagging a potential infrastructure scaling issue or inefficient new code deployment.
• Rule 8 (8 points outside 1σ): Eight consecutive daily inference runs show prediction values clustered outside the 1-sigma zone on one side of the mean, suggesting the model's output distribution has permanently shifted.

Ensures the stability of automated synthetic data generation or ETL processes. Rules verify that the statistical properties of generated data remain consistent with the training distribution.

• Rule 5 (4 of 5 beyond 1σ): Four out of five consecutive batches of synthetically generated images have a mean pixel intensity beyond 1 standard deviation, indicating the generator is producing abnormally bright or dark outputs.
• Rule 6 (14 alternating points): A time-series of feature correlation coefficients from a generated tabular dataset alternates direction 14 times, highlighting an unstable or oscillating data generation process.
• Rule 7 (15 points within 1σ): Used as a positive control. Fifteen consecutive validation scores for a synthetic text generator fall within 1 sigma, confirming the process is in a state of high stability and control.

Forms the statistical core for monitoring business metrics, IoT sensor streams, and application telemetry. The rules transform raw metrics into actionable alerts for Site Reliability Engineering (SRE).

• IoT Example: Rule 2 (Run of 9) on temperature sensor readings from industrial equipment flags a sustained upward trend, enabling predictive maintenance before failure.
• Business KPI Example: Rule 1 and Rule 4 applied to a control chart of daily active users. A single point below the lower control limit (Rule 1) indicates an outage, while two of three points below 2 sigma (Rule 4) may indicate a growing retention problem.
• Infrastructure Example: Rule 3 (Trend of 6) on API 99th percentile latency detects a gradual performance degradation caused by memory leaks or database index fragmentation.

Integrates with MLOps platforms to provide deterministic, rule-based alerts alongside ML-driven anomaly detection. This creates a hybrid monitoring system with high explainability.

• Complementing ML Models: While a neural network might detect complex, non-linear anomalies, the Western Electric Rules provide simple, auditable checks for classic statistical patterns. This reduces alert fatigue and false positives.
• Benchmarking Data Quality: Rules are applied to data quality metrics (completeness, uniqueness, freshness) calculated by observability tools. A violation moves an issue from 'watch' to 'actionable' status in incident management.
• Governance and Audit: The rules provide a clear, documented rationale for alerts (e.g., "violated Rule 1"), which is critical for regulated industries requiring explainable AI and data governance.

The Western Electric Rules operate within a broader ecosystem of statistical and engineering practices for quality control.

• Cumulative Sum (CUSUM) Chart: A more sensitive alternative for detecting small, persistent shifts in a process mean, often used after Western Electric Rules for finer-grained analysis.
• Exponentially Weighted Moving Average (EWMA) Chart: Another sensitive chart that weights recent data more heavily, useful for processes where small shifts are critical.
• Statistical Process Control (SPC) Software: Platforms like JMP, Minitab, and SigmaXL implement these rules. In data stacks, they are codified in data observability platforms (e.g., Monte Carlo Data, Great Expectations) and workflow engines (Apache Airflow with custom operators).
• Average Run Length (ARL): The metric used to evaluate the performance of these rules, quantifying the average number of samples before a false alarm or before detecting a real shift.

A control chart is the primary graphical tool used in SPC where the Western Electric Rules are applied. It plots process data over time against statistically derived control limits and a center line (process mean). The chart's purpose is to visually distinguish between common cause variation (inherent, random noise) and special cause variation (assignable, non-random signals). Common types include X-bar charts for means and I-MR charts for individual observations.

Control limits are the statistically calculated boundaries on a control chart, typically set at ±3 standard deviations from the process mean. They are not specification limits but define the expected range of common cause variation. A single point outside these limits is Rule 1 of the Western Electric Rules, providing the strongest signal of an out-of-control process. Limits are calculated from historical process data when the process is stable.

Special cause variation (or assignable cause variation) is non-random, unexpected fluctuation in a process that indicates a specific, identifiable change in the system. The Western Electric Rules are explicitly designed to detect the patterns that signal special cause variation, such as:

• A point outside control limits
• A run of 7+ points on one side of the center line
• A trend of 6+ points consecutively increasing or decreasing Identifying special causes is the first step toward process improvement.

Process stability (or statistical control) is a state where a process exhibits only common cause variation and its key statistical properties (mean and variance) are constant and predictable over time. A stable process is a prerequisite for meaningful capability analysis. The Western Electric Rules are the decision criteria for assessing stability; if any rule is violated, the process is deemed unstable, and investigation into special causes is required before further analysis.

Average Run Length (ARL) is a performance metric for control charts, representing the average number of samples plotted before the chart signals an out-of-control condition. It measures the sensitivity of detection rules. For a perfectly stable process, a higher ARL is desirable (fewer false alarms). For an unstable process, a lower ARL is better (faster detection). The Western Electric Rules, used together, are designed to optimize the ARL for detecting various shift sizes and patterns.

Statistical Process Control (SPC) is the overarching methodology of using statistical techniques to monitor and control a process. It aims to achieve and maintain process stability and improve process capability. SPC encompasses:

• Data collection via rational subgrouping
• Analysis using control charts
• Decision-making using rules like the Western Electric Rules
• Corrective action for special causes It is a foundational element of Six Sigma and modern data quality monitoring frameworks.