Gains:
- Ability to interpret control charts (X-bar/R, p, c) and process capability indices (Cp, Cpk)
- Ability to accelerate scanning and reporting of out-of-control signals (Western Electric rules) with AI
- Ability to validate AI's statistical interpretation with raw data and physical process reality
Quality is not "sorted" by inspection, it is "produced" by the process. Statistical Process Control (SPC) aims to do just that: monitoring the process in real time and catching signals before errors become products. The strength of SPC lies in distinguishing natural (common cause) variability from abnormal (special cause) variability. In this area, AI accelerates control chart interpretation, rule scanning, and proficiency reporting; but it is human's job to blend statistical interpretation with the reality of the physical process. In this unit, we will cover control charts, Cp/Cpk proficiency indices and the proper role of AI in SPC.
Two Faces of Variability
Every process involves variability. SPC divides this variability into two:
- Common cause: Random, predictable fluctuation inherent in the process. The process is "under control".
- Special cause: An external, assignable event (tool breakage, raw material batch change, setting shift). The process is "out of control".
The sole purpose of the control chart is to separate the two. Responding to common-cause fluctuation (over-adjustment / "tampering") actually increases variability. Deming's classic "funnel experiment" shows this: intervening at every deviation makes things worse.
Tip: Never confuse control limits (UCL/LCL) with tolerance (specification) limits. Control limits are the “voice” of the process (calculated from data); tolerance limits are the "voice" of the customer (comes from the design). AI sometimes confuses these; Make it clear which one you are talking about in each comment.
Control Chart Types
The chart is selected according to the data type.
Graphics
Data type
example
X-bar/R
Continuous (measurement), subgroup
Shaft diameter average and clearance
X-bar/S
Continuous, large subgroup
Filling weight
I-MRI
Continuous, single measurement
Single measurement per day
p/np
Discrete, defective rate/number
Faulty parts rate
c/u
Discrete, number of defects
Defects per unit
On an X-bar/R chart, control limits are calculated from subgroup averages and spans:
UCL_xbar = X-double-bar + A2 × R-barLCL_xbar =
Out of Control Signals (Western Electric Rules)
A point going out of bounds is not the only signal. Western Electric (Nelson) rules also capture patterns:
- A point is outside 3σ (out of bounds).
- Consecutive 2/3 points beyond 2σ (on the same side).
- 4/5 consecutive points beyond 1σ.
- 8 consecutive points on the same side of the center line (shift).
- 6 dots constantly increasing/decreasing (trend).
Role: You are a quality engineer experienced in SPC. Task: Scan the following X-bar values with Western Electric rules according to the given UCL/LCL and center line. For each out-of-control signal: rule number, which points, possible physical interpretation, and suggested investigation step. Center=50.2 UCL=52.1 LCL=48.3 σ=0.63Data (ordered): {{ values }}Rule: Don't generate spurious signal; Marking points that do not fully satisfy the rule. Present possible reasons as "hypothesis", do not write in precise language.
AI makes this screening fast, but remember: rule provision is a statistical alert; Field investigation finds the real cause. AI might say "probably tool wear"; It is the tool measurement that confirms this.
Process Capability: Cp and Cpk
Even if the process is under control, does it meet customer tolerances? Sufficiency indexes tell this.
Cp = (USL − LSL) / (6σ) → measures only the spreadCpk = min[ (USL − μ)/(3σ) , (μ − LSL)/(3σ) ] → includes centration
Example: USL = 10.5; LSL = 9.5; μ = 10.1; σ = 0.1.
Cp = (10.5 − 9.5) / (6 × 0.1) = 1.0 / 0.6 ≈ 1.67Cpk = min[ (10.5 − 10.1)/0.3 , (10.1 − 9.5)/0.3 ] = min[ 0.4/0.3 , 0.6/0.3 ] = min[1.33 , 2.0] = 1.33
Cp = 1.67 indicates a good spread, but Cpk = 1.33 tells us that the process is off-center (average 10.1, closer to USL than target 10.0). Cpk < Cp always indicates a centering problem. Even a process with good dispersion will produce a faulty product if it is incorrectly centered.
Weak Prompt / Strong Prompt
Weak prompt:
Is this process good?
There is no criterion for "good"; AI only makes a general guess.
Powerful prompt:
Calculate Cp and Cpk with the following measurement data (USL=10.5, LSL=9.5). Calculate σ and mean from data, show formula and each step. If Cpk < Cp, ask whether it is centering or spreading, comment. Also estimate the approximate error rate (ppm) of the process and write down your assumptions. Data: {{ ... }}
Mini Case: Good CP, Bad Product
On a filling line, most of the bottles come out of specification, but the process appears "stable" on the control chart. Industrial engineer Yağmur gives the data to the AI and has it calculate Cp and Cpk: Cp = 1.5 (good spread) but Cpk = 0.7 (insufficient). The difference is that the average has shifted significantly from the target. AI points out the centering problem; He checks the setting of the rain fill valve and indeed finds an offset. When you correct the setting, Cpk increases to 1.4 and waste decreases. Critical lesson: the process can be “stable” (under control) but still be “deficient” (not meeting specification). AI demonstrated this distinction numerically; It was Yağmur who found the root cause on the field.
Common Mistakes
- Confusing control and tolerance limit: Equating control limits with customer tolerance.
- Common cause tampering: Adjusting for each fluctuation and increasing variability.
- Ignoring Cpk: Just looking at Cp and missing the centering issue.
- Mistaking AI signals for evidence: Not verifying the cause of the Western Electric signal in the field.
- Wrong chart selection: Applying a continuous data chart (X-bar) to defect rate data.
In summary
- The essence of SPC is to separate common and special cause variability; Interfering with the common cause is harmful.
- Control limits come from the process, tolerance limits come from the design; The two should never be confused.
- Select the chart appropriate to the data type (X-bar/R, p, c...) and look for patterns with Western Electric rules.
- Cp measures spread and Cpk measures centration; Cpk < Cp indicates a centering problem.
- AI generates statistical signal fast; Field inspection confirms the reason, the decision is up to the engineer.
Application task
Take at least 20 continuous measurements from a process (real or fictional) and determine USL/LSL values. Have the AI calculate Cp and Cpk with this data; Have σ and mean calculated from the data and shown each step. Critical verification: Recalculate the σ, Cp, and Cpk values yourself in Excel or Python and compare with the AI output. Then scan the data with Western Electric rules and ask "how do I verify this in the field?" for each signal found. Answer the question. Finally, if Cpk < Cp, decide whether centering or spreading should be corrected first.