Predictive Quality and Early Warning — From "Post-SPC" to "Pre-Prevention"

Abstract: Addressing control chart out-of-control conditions only after they occur often results in an entire batch of defects. A gradual mean shift can go undetected by the 3σ rule for weeks. Predictive Quality (PQ) builds upon traditional SPC by using EWMA/CUSUM, multivariate monitoring, and (optionally) machine learning to shift the alarm point earlier. This article provides a three-tier early warning framework, a template for rule design, two complete calculation examples, and methods for controlling false alarms and organizational closed-loop processes — helping quality and digital teams transform their "alarm systems" into "improvement systems."

1. Case Study: Cpk 1.4 Still Draws Customer Complaints

A precision machining supplier processes the diameter of a center shaft with a Cpk of 1.45 for a long time, and the X̄-R control chart shows "no out-of-control points." In the 7th week, the customer's sampling inspection found that the mean had shifted 0.012 mm towards the upper specification limit (USL), with a cumulative 4000 pieces within tolerance but close to the USL, causing assembly interference.

Retrospective:

• No single value exceeded the UCL, but the 12th point on the EWMA chart exceeded the warning line.
• Tool wear showed a linear trend, with a daily compensation of 0.002 mm, but the operator did not record it.
• No trend rules (such as 7 consecutive points rising) were enabled.

If a yellow warning had been triggered in the 3rd week and the tool was replaced or the compensation verified, the batch risk could have been avoided. This is the gap that Predictive Quality aims to fill — not to replace SPC, but to detect trends and combined anomalies earlier.

2. Concept Boundaries

2.1 Predictive Quality vs. Predictive Maintenance (PdM)

The data platform can be shared (MES/SCADA), but metrics and response processes must be separated.

2.2 Relationship with 6.3 SPC and 12.2 Digitalization

• 6.3.1 Control charts are the foundational layer.
• 6.3.3 SPC digitalization involves data collection and deployment.
• 12.2.2 This article — early warning logic and operations.
• 12.2.3 Visual + SPC integration — online 100% data for early warning.

3. Three-Tier Early Warning Architecture (Recommended Implementation Order)

L1 — Classic SPC + Western Electric Rules

Applicability: All critical CTQs, mandatory for audits.

Common 8 rules (excerpt):

1. 1 point exceeds 3σ
2. 9 consecutive points on the same side of the center
3. 6 consecutive points monotonically increasing/decreasing
4. 14 consecutive points alternating above and below the center
5. ...

Advantages: Explainable, trainable, customer-recognized. Limitations: Insensitive to small shifts, slow drifts; primarily single-variable.

L2 — Statistical Early Warning (Strongly Recommended)

Intuition of EWMA: Recent data has more weight — "gradual deterioration" alarms earlier than "sudden failure."

L3 — Machine Learning (Pilot)

• Unlabeled: Isolation Forest, Autoencoder anomaly detection
• Labeled: Nonconforming batch characteristics → classification/regression risk assessment

Red Line:

• No black-box automatic line stop
• Must run in shadow mode for 1-3 months to compare with L1/L2
• Model changes must be validated and version-managed

4. Calculation Case 1: EWMA Detects Drift Early

Target diameter μ₀ = 10.000 mm, σ estimated at 0.004 mm (during control period).

Recent 10 points (mm): 10.000, 10.001, 10.002, 10.003, 10.004, 10.005, 10.006, 10.007, 10.008, 10.009

• X Chart: All within UCL/LCL
• Rule 3 (6 points monotonically increasing): Triggered
• EWMA (λ = 0.2): Exceeds 2σ warning from the 8th point

SOP for Handling (Yellow):

1. Process engineer confirms tool/compensation on-site within 2 hours.
2. Increase inspection of the next batch's 100% critical dimensions.
3. Root cause and corrective action (tool replacement cycle? compensation SOP?) within 24 hours.

Without Early Warning: Production might continue for another 2-3 days until the customer discovers the issue.

5. Calculation Case 2: Multivariate — Injection Molding Dimensions + Weight

A plastic part shows decreased weight + decreased dimensions simultaneously, but the dimension chart still appears within tolerance — this is an early sign of underfill.

Establish a T² control chart (2 variables), starting from the 6th mold, T² exceeds the boundary → check material temperature, injection speed, and holding pressure — find that the material pipe temperature is 5°C lower.

Insight: Critical processes should identify physically related parameter groups and perform multivariate monitoring, rather than using 20 single-variable charts in isolation.

6. Early Warning Tiers and Response SLA

Fields for Each Early Warning Work Order: Trigger rule, CTQ, batch, responsible person, handling, closure verification, whether to update FMEA/CP.

7. Management of False Alarms and Missed Alarms

Alarm fatigue is the primary cause of predictive quality failure — turning off notifications on-site = system abandonment.

Controlling False Alarms:

• Multiple rules AND combination (e.g., EWMA exceeds boundary and 3 consecutive points on the same side)
• Hysteresis (dead zone): Recovery requires N consecutive normal points
• Segmented control charts for changeover/startup — startup data not included in steady-state rules
• Monthly review: False alarm rate = proportion of closed "no action" cases, target <30%

Controlling Missed Alarms:

• Retain the full set of L1 rules
• Post-incident review of customer complaint batches — which rule should have triggered but did not?
• Closed-loop with 10.2.1 warranty data

8. Implementation Roadmap (90 Days)

Month 1 — Foundation

• Select 3 CTQs, MSA qualified
• Implement X̄-R or I-MR + 8 Western Electric rules
• Automatic data collection via MES (eliminate manual data entry)

Month 2 — L2

• Implement EWMA or CUSUM for the same CTQs
• Define yellow/orange/red and SLA
• Run early warning work orders in a closed loop

Month 3 — Evaluation

• Compare: number of line stops after early warning, scrap tonnage, customer complaints
• Decide whether to expand or pilot L3

Organization: Quality owns the rules; IT owns the pipeline; Process owns the handling.

9. Integration with OEE and FPY

• Early warning → short line stop for adjustment → counted as OEE adjustment loss (7.3.1) — but total loss is still less than batch scrap
• Steady-state FPY improvement → 3.3.2 first pass yield improvement can be quantified

In the OEE analysis meeting (Issue 3 of the Equipment Series), a separate column should be listed for quality early warning-induced planned minor adjustments vs unplanned quality line stops.

10. Common Misconceptions

11. Summary

Predictive Quality = extension of SPC + trend sensitivity + tiered response + closed-loop improvement.

A good system: few and accurate, explainable, capable of closed-loop, and able to review. Start with L1+L2, then discuss AI — this is the replicable path for manufacturing companies.

Knowledge code: 12.2.2

Version: v20260528

Author: Quality Think Tank