Detailed Explanation of the Eight Criteria for Control Chart Anomaly Detection —— Using the Right Rules for Precise Early Warning

A control chart is the core tool of statistical process control (SPC), but many quality professionals fall into a common pitfall: they only look at whether data points exceed the control limits (UCL/LCL), and if they do, an alarm is triggered; if not, they assume "everything is normal."

In fact, anomaly detection is not limited to "exceeding limits." As early as 1956, Western Electric summarized a set of anomaly detection criteria, which were later expanded by Nelson into eight rules. These rules shift the control chart's early warning capability from "waiting for a failure to occur" to "alerting when a trend is abnormal" — identifying special causes before the process goes out of control.

This article will break down each of the eight anomaly detection criteria, along with pattern interpretation and practical tips, to help you truly use control charts effectively.

Overview of the Eight Anomaly Detection Criteria

Before delving into each rule, it's important to establish a holistic understanding. The core logic of all anomaly detection rules is: determine whether the distribution of data points follows a random pattern. If the points exhibit clear non-random characteristics (such as continuous points on one side, a trend of increasing/decreasing, or periodic fluctuations), it indicates that the process has a special cause, and an investigation is needed.

Here is a summary of the eight criteria:

Note: The A zone is the ±2σ to ±3σ interval, the B zone is the ±1σ to ±2σ interval, and the C zone is the ±1σ interval around the center line.

Interpretation and Practical Insights of Each Criterion

Rule 1: One Point Out of Limits (Exceeding A Zone)

Condition: Any 1 data point falls outside the control limits (above UCL or below LCL).

This is the most basic and intuitive anomaly detection rule. When the process is normal, the probability of a point falling outside the control limits is only 0.27% (based on the 3σ principle), which is a low-probability event. Once it occurs, it almost certainly indicates a special cause in the process.

Practical Tip: An out-of-limits point is just a signal and does not necessarily mean that the product at that point is nonconforming (control limits ≠ specification limits). Immediate investigation should be conducted from the six dimensions of personnel, machinery, materials, methods, environment, and measurement to identify and eliminate the special cause, rather than directly adjusting process parameters — that is called "over-adjustment" (tampering), which can actually increase process variability.

Rule 2: Nine Consecutive Points on One Side of the Center Line

Condition: 9 consecutive data points fall on the same side of the center line (all above or all below).

Even if all points are within the control limits, 9 consecutive points on one side is a strong signal of abnormality. The probability of points falling on either side of the center line is 50%, and the probability of 9 consecutive points falling on the same side is only 0.2% (0.5⁹ ≈ 0.002).

Common Causes: Raw material batch changes, gradual tool wear, operator changes, environmental temperature drift, etc. This rule is highly sensitive to slow mean shifts in the process and often triggers an alarm several hours or even days earlier than the out-of-limits rule.

Rule 3: Six Consecutive Points Increasing or Decreasing

Condition: 6 consecutive points show a monotonic increasing or decreasing trend.

A trend indicates that the process is drifting in one direction. The probability of 6 consecutive points moving in the same direction in a random process is low.

Typical Scenarios: Tool wear (dimensions continuously increase or decrease), chemical concentration decay, voltage/air pressure gradually decreasing, etc. Note: "Increasing" here means point-by-point rise, allowing adjacent points to have the same value but not to reverse direction.

Rule 4: Fourteen Consecutive Points Alternating Up and Down

Condition: 14 consecutive data points alternate up and down on either side of the center line (e.g., up-down-up-down... pattern).

Alternating up and down presents a "sawtooth" or "oscillating" pattern, typically indicating that some periodic factor is causing the process to switch between two states.

Common Causes: Alternating use of two machines (differences in speed or precision leading to alternating high/low results), shift changes between two operators, alternating use of two testing instruments, diurnal temperature cycles, etc.

Rule 5: Two Out of Three Points in A Zone (Same Side)

Condition: At least 2 out of 3 consecutive data points fall in the A zone (±2σ to ±3σ, and on the same side).

This indicates that process variability has significantly increased, although the points have not yet exceeded the control limits, they are very close. This rule is sensitive to increased process standard deviation.

Practical Tip: This rule is often used in conjunction with Rule 1 as a "pre-warning" signal. When Rule 5 is triggered, even without any out-of-limits points, the control level should be increased, and the frequency of sampling inspections should be raised to closely monitor the development of subsequent points.

Rule 6: Four Out of Five Points Outside B Zone (Same Side)

Condition: At least 4 out of 5 consecutive data points fall outside the B zone (i.e., in the A zone or further, and on the same side).

This is an extension of Rule 5 — the criteria are relaxed from the A zone to "outside the B zone," but the number of consecutive points increases from 3 to 5. It also serves as a warning for mean shifts or increased variability in the process, though the signal strength is slightly weaker than Rule 5 but appears earlier.

Rule 7: Fifteen Consecutive Points in C Zone (Within ±1σ of the Center Line)

Condition: 15 consecutive data points fall within the C zone (±1σ of the center line).

This rule is somewhat special — "too good" data points are actually abnormal. Intuitively, points concentrated around the center seem to indicate a stable process, but excessively low variability is also an abnormal signal.

Possible Explanations: Data has been "adjusted" (operators deliberately select data, testing instruments with insufficient precision leading to excessive rounding), sampling method issues (too short sampling intervals causing high data autocorrelation), or incorrect control limit calculations (using specification limits instead of control limits). When this rule is triggered, the authenticity and rationality of the data source should be verified first.

Rule 8: Eight Consecutive Points Outside C Zone

Condition: 8 consecutive data points fall outside the C zone (all in the B or A zones, on either side).

Contrary to Rule 7, data points are too dispersed, indicating increased process variability and a structural change in the standard deviation. Although all points may still be within the control limits, the process capability has significantly decreased.

Typical Scenarios: Increased machine clearance, loose tooling, increased variability in incoming materials from a new supplier, etc. The source of variability should be investigated, and process control should be tightened.

Key Points and Common Misunderstandings in Using Anomaly Detection Rules

Point 1: Rules should be used in layers, not all at once. Not all processes require the application of all eight rules. For production lines implementing SPC for the first time, it is recommended to start with Rule 1 and Rule 2, and gradually enable Rules 3 to 8 as the team gains experience. Activating too many rules at once can significantly increase the false alarm rate.

Point 2: Manage false alarm rates. Theoretically, each additional anomaly detection rule increases the probability of false alarms. For example, using the common combination of "1/2/3/5/6" five rules, the probability of a false alarm every hour when the process is under control is about 1% to 3%. Therefore, the results of the anomaly detection rules should not be treated as "conclusive judgments" but as "investigation signals" — investigate first, then judge.

Point 3: Anomaly detection rules cannot replace the recalculation of control limits. When encountering out-of-limits points or abnormal patterns, the special cause should be eliminated, not immediately recalculating the control limits. Only after confirming that the special cause has been thoroughly eliminated and the process has returned to a controlled state should new data be used to update the control limits.

Common Misunderstanding 1: Treating all points exceeding the specification limits as anomalies in the control chart. Specification limits (USL/LSL) are customer requirements, while control limits (UCL/LCL) are the inherent variability boundaries of the process — these concepts are different and should not be confused.

Common Misunderstanding 2: Adjusting equipment parameters upon seeing an anomaly. "Over-adjustment" can actually amplify process variability. The correct approach is to first investigate and eliminate the special cause, and only after confirming that the process has returned to a controlled state, consider optimizing parameters.

Common Misunderstanding 3: Focusing only on Rule 1 (out-of-limits) and ignoring other patterns. In reality, trend-based rules like Rule 2 and Rule 3 often detect issues earlier than out-of-limits, which is key to upgrading control charts from "post-event alarms" to "pre-event warnings."

Practical Examples

The following sequence appeared on the control chart (mean chart) for a plastic injection molding production line:

From subgroup 1 to subgroup 7, there is a continuous increasing trend, meeting Rule 3 (six consecutive points increasing). Although no points have exceeded the control limits at this point, the quality engineer received a warning and conducted an on-site investigation, discovering that a partially blocked mold cooling water circuit was causing the mold temperature to rise continuously, leading to a continuous decrease in shrinkage and an increase in dimensions. After clearing the cooling water circuit, the dimensions of subgroups 8 and 9 returned to normal, with the mean back near the center line.

In this case, if only Rule 1 had been waited for, it would have taken at least until subgroup 10 for a point to exceed the control limits — by then, a large number of nonconforming products would have been produced. The early warning from Rule 3 allowed the issue to be identified and resolved early, preventing a batch of defective products.

Anomaly detection criteria are the "translators" of control charts — converting data patterns into process signals, allowing quality professionals to hear the footsteps of anomalies before a failure occurs.

Knowledge code: 6.3.1

Version: v20260724

Author: Quality Think Tank Quality Think Tank is dedicated to providing systematic professional knowledge, methodologies, and practical tools for quality management practitioners, helping companies continuously improve their quality capabilities.