Units
1. Introduction to Artificial Intelligence in Metallurgy: Boundaries, Validation, Responsibility and Ethics 2. AI in Material Selection and Decision Support Systems 3. Feature Prediction and Structure-Process-Property Relationships (Machine Learning) 4. Phase Diagrams, Thermodynamics and AI in CALPHAD Approach 5. Microstructure Image Analysis and Quantitative Metallography 6. AI in Optimization of Heat Treatment and Process Parameters 7. Manufacturing Processes: AI in Casting, Welding and Additive Manufacturing 8. Quality Control, Defect Detection and Statistical Process Control 9. AI in Damage Analysis, Root Cause and Breakage Investigation 10. Characterization Data Interpretation: XRD, SEM/EDS, DSC and Mechanical Testing 11. Standard Investigation, Compliance and Verification Discipline 12. Metallurgical Data Analysis and Frontiers of AI with Python
Unit 8 / 12

Quality Control, Defect Detection and Statistical Process Control

Gains:

  • Ability to monitor production and inspection data with AI and detect defect and deviation patterns early
  • Ability to set up and interpret statistical process control (SPC), Cpk and control charts with AI support
  • Ability to verify AI's defect classification with NDT result, acceptance criteria and authorized inspection confirmation

It is the constant duty of the metallurgical and quality engineer to notice which of the thousands of parts passing through a production line is defective and at what moment the process begins to slip. Quality control, checking whether the product meets the specified requirements; defect detection, capturing errors in the product (scratches, pores, cracks, dimensional deviations); Statistical process control (SPC) is a method of understanding whether the process is under control by monitoring production data statistically. Artificial intelligence accelerates automatic defect classification from images, pattern capture in large measurement data, and construction and interpretation of SPC graphs. But an alarm is not an automatic decision; Each AI output must be confirmed by NDT result, acceptance criteria and authorized inspection approval.

Basic concepts of SPC and the role of AI

  • Control chart: A graph showing the change of a measurement (e.g. diameter, hardness) over time, with a center line and control limits (usually ±3 standard deviations). Limit crossing is a sign that the process has changed.
  • Control limit etc. Tolerance: Control limits come from natural variation of the process; tolerance (specification limit) is what the customer wants. The two should not be confused.
  • Cp and Cpk (process capability indices): Measures how easily the process fits within tolerance. Cp evaluates the spread, and Cpk evaluates the spread + decentration together. Cpk ≥ 1.33 is generally considered sufficient.
  • Private vs. common cause: common cause is the inherent noise of the process; The specific cause (e.g. tool wear, wrong batch) is a deviation that requires intervention. The purpose of SPC is to separate the two.

AI is powerful in teaching these concepts, generating code that calculates a control chart/Cpk from a data, and interpreting chart patterns (trend, drift, periodicity). But the engineer determines the cause of the limit exceedance and the correct intervention.

Attention: Expanding the control limits "to prevent an alarm" is one of the most dangerous mistakes. The limits are calculated from the actual variation of the process; Extending them arbitrarily makes real deviations invisible and moves the defective product to the customer.

Step by step: AI-powered quality monitoring

  1. Select characteristic: Which critical metric will be monitored (CQ: critical quality characteristic)? Is the measurement system reliable (Gage R&R)?
  2. Collect data: Regular sampling in subgroups. Is the meter calibrated?
  3. Set up chart: Ask the AI ​​for the appropriate control chart (X-bar/R, individual-moving range) type and code.
  4. Comment: Are there any overshoot, trend, batch rules (e.g. Western Electric rules)?
  5. Look for root cause: Alarm is a trigger; Confirm measurement, review process.
  6. Decide: Based on acceptance criteria and NDT result; Separate the deviated batch and apply corrective action.

Flaw detection and non-destructive testing

Image-based AI (deep learning classifiers) can quickly extract surface defects (scratches, burrs, oxides, pores). Internal defects require non-destructive testing (NDT): radiography (RT), ultrasonic (UT), magnetic particle (MT), penetrant (PT). AI may pre-screen RT/UT images, but final evaluation is by certified NDT personnel and standard acceptance criteria.

Defect type

Typical method

AI contribution

final decision

Surface scratch/burr

optical image

automatic classification

Acceptance criterion + operator

Internal pore/crack

RT/UT

Image pre-screening

Certified NDT+ standard

Crack (open to surface)

MT/PT

indication marking

NDT level II/III approval

Size deviation

CMM / measurement

Trend tracking with SPC

Tolerance + decision

three mini cases

Case 1 — Border expansion trap. The upper control limit is frequently exceeded in diameter measurement on a line. The operator recommends expanding the limits to reduce alerts. The quality engineer objects to this; He has the AI ​​analyze the SPC data and finds that there is a systematic shift (specific cause) in the afternoon shift, possibly due to tool wear. When the tool change period is shortened, the process returns to within limits. Lesson: alarm is a source of information; It is necessary not to silence him, but to find the reason.

Case 2 — Wrong Cpk interpretation. One report says the process is "Cpk = 1.8, perfect" but the parts are giving the customer problems. The engineer has the AI ​​examine the data; He finds that the real change is hidden because the resolution of the measurement system is insufficient (Gage R&R is bad), the high Cpk comes from measurement blindness. First, the measurement system is corrected and the real Cpk is 1.1. Lesson: any statistic based on a bad measurement system is misleading; Verify the measurement first.

Case 3 — Correct NDT elimination. One manufacturer uses AI-powered RT pre-qualifying on castings; flags suspicious parts and directs them to certified NDT personnel. The AI ​​quickly goes through most of the clean pieces, with the staff only focusing on the marked ones. However, in a safety-critical series, parts that the AI ​​calls "clean" are also sampled and checked with NDT; because AI is a screening tool, not an admissions authority. Lesson: AI improves efficiency but final acceptance comes with standard NDT and authorized approval.

Copiable prompt templates

CONTROL CHART SETUP TEMPLATE "Role: You are SPC / quality assistant. I have the following measurement data: [subgroup data]. Characteristic: [diameter/hardness]. Recommend me the appropriate type of control chart (X-bar/R or I-MR) with justification and calculate the center line + control limits with Python code. Emphasize that the control limits are different from TOLERANCE. Comment on the limit exceedance, but state that the reason should be investigated."

Cpk COMMENT TEMPLATE "Cp = [...], Cpk = [...], tolerance = [...], measured distribution = [...]. Interpret these values: is the process adequate, is there decentration? Remind me to ask FIRST whether the measurement system is reliable (Gage R&R). Explain the pitfall that high Cpk can come from measurement blindness. Relate the firm 'appropriate' decision to the experiment/inspection."

DEFECT CLASSIFICATION EVALUATION TEMPLATE "The AI ​​image model marked the following defects: [list]. For each: against which acceptance criteria (standard) should it be evaluated, which NDT method is required for confirmation, what is the risk of false-positive/negative? Specify that final acceptance should be given with certified NDT and authorized approval. Position AI output as 'pre-screening'."

SPC PATTERN ANALYSIS TEMPLATE"Examine the following control chart data: [data/time series]. Is there a trend, drift, periodicity, or series rule violation (Western Electric)? For each pattern, list the possible process cause (tool wear, batch change, temperature) as a HYPOTHESIS and write how it will be verified. Do not declare an exact cause."

Weak prompt / Strong prompt

WEAK PROMPT: "Is the quality of this data good?"

STRONG PROMPT: "Role: You are the SPC assistant. I have the following shaft diameter data (25 subgroups, n=5): [data]. Tolerance is 20.00 ± 0.05 mm. Have me: (1) calculate the center and control limits for the Compare the result with the acceptance criteria, but remind that the measurement system (Gage R&R) needs to be verified.”

The weak prompt gives an answer that is vague and without comment. Powerful prompt gives clear data, tolerance and desired outputs; It stipulates the control limit/tolerance distinction, Gage R&R control, and that the final decision remains with the engineer.

Common mistakes

  • Confusing control limits with tolerance or expanding them so that no alarm occurs.
  • Relying on Cpk without verifying the measurement system (Gage R&R).
  • Considering the limit exceedance as the automatic "root cause" and deciding to scrap without confirming the measurement.
  • consider the AI ​​defect classification as the final acceptance decision; Bypassing NDT and authorized approval.
  • Intervening in every fluctuation without distinguishing between the common cause and the specific cause (over-adjustment).
  • In a safety-critical series, fully trusting what the AI ​​says is "clean" and not performing sampling verification.

In summary

AI in quality control increases the eye and speed of the quality engineer with defect classification, pattern capture in big data and SPC graph/Cpk calculation. However, if the control limit-tolerance distinction, measurement system reliability and special/common cause distinction are not made correctly, statistics can be misleading. Treat each alarm as a trigger, confirm the measurement, investigate the root cause and link final acceptance to acceptance criteria, standard NDT and authorized approval. AI makes a preliminary selection, the engineer makes the decision.

Application task

Specify a real or fictional series of measurements (e.g. 20-25 subgroups of diameter/hardness data) and a tolerance. With the "SETUP CONTROL CHART" template, ask the AI ​​for the appropriate chart type and calculation code and have it calculate control limits. Then comment the process adequacy using the "Cpk COMMENT" template and note the Gage R&R check. Finally, create a limit exceeding scenario and write in a paragraph which steps you will decide on (measurement confirmation, root cause, acceptance criteria).

checklist

  • [ ] I have identified the critical quality attribute to be monitored and the reliability of the measurement system.
  • [ ] I selected the appropriate control chart type and calculated the limits from the process.
  • [ ] I separated the control limit and tolerance; I did not expand the boundaries arbitrarily.
  • [ ] I interpreted Cpk together with measurement system reliability.
  • [ ] I considered the alarm as a trigger, confirmed the measurement, and investigated the root cause.
  • [ ] I have based final acceptance on acceptance criteria, standard NDT and authorized approval.
Previous Next