Series: AI in Lean Six Sigma, post 2. This manufacturing deep dive builds on post 1, “AI in Lean Six Sigma: How Artificial Intelligence Is Rewriting the DMAIC Playbook Across Industries.”
An injection molding plant runs 14 presses across three shifts. Final inspection catches defects after parts have already consumed machine time, labour and material. Scrap is running at 2.4%, costing an estimated $412,000 a year. By the time the team begins investigating, the evidence is scattered across inspection records, machine settings and operator notes.
Now change the sequence. Sensors collect process data during each shot, and a predictive model identifies conditions associated with an emerging defect before the part is demolded. Operators can respond while the process is still producing good parts, not only after a defect appears.
That is the opportunity behind AI in manufacturing quality. The goal is not to remove human judgement or replace Lean Six Sigma. It is to give teams earlier, richer evidence so they can make better decisions through DMAIC.
The manufacturing AI toolkit, mapped to DMAIC
AI is most useful when it strengthens a disciplined improvement method. In a manufacturing project, each DMAIC phase has a distinct role: establish trustworthy data, find validated causes, test changes, and sustain the gains.
Measure: build a reliable picture of each production cycle
Predictive quality depends on knowing what happened, when it happened, and which product it affected. For an injection molding process, a useful data set may link each shot to its press, mold, cavity, resin, shift, process settings and measured outcome.
Three data streams can form a practical starting point:
- Machine vision: Calibrated cameras can inspect visually observable features at every cycle, potentially moving a plant from sample-based checks to 100% inspection for selected characteristics. Optical measurement may also automate some dimensional checks currently performed with manual calipers. It should be validated against a suitable reference method before teams rely on it for acceptance decisions.
- IIoT sensors: Capture cycle time, barrel and mold temperature, injection speed, pressure profiles and other shot-level signals. The right measures depend on the defect and the process.
- Automated gauge feeds: Connect measurement results, such as part weight or critical dimensions, to the corresponding shot and cavity, reducing manual transcription and supporting faster analysis.
Before modelling, confirm that measurements are fit for purpose. A measurement system analysis, clear operational definitions and consistent part identification help establish whether a pattern is genuine or an artefact of how data was collected.
Analyse: identify which inputs are associated with defects
A modern molding process may record 40 or more variables. Statistical and machine-learning methods can help narrow the field: which settings, sensor patterns or interactions distinguish good parts from parts with sink marks, flash, short shots or dimensional issues?
For example, a model might surface a relationship between barrel-temperature bands and sink marks. A visual explanation can show how defect frequency changes across temperature ranges, with other relevant conditions considered. This gives engineers a focused hypothesis to investigate, not automatic proof of cause.
That distinction matters. Correlation helps direct attention; controlled testing and process knowledge help establish whether changing a factor will improve the outcome. A disciplined Analyse phase uses model findings alongside stratification, process maps, cause-and-effect analysis and statistical tests.
Research illustrates both the promise and the need for care. A peer-reviewed inline AI quality-control study of injection-molded parts combined machine, mold and environmental sensing with inline quality measurement and vision. The researchers reported strong predictive performance for some characteristics, while surface-quality prediction was more challenging. In other words, the model’s usefulness can vary by quality characteristic; one impressive overall score does not mean every defect is equally predictable.
Improve: test the change before committing the line
Once the team has validated likely causes, AI can help explore improvement options more efficiently.
A digital twin or process simulation can estimate how changing injection speed, holding pressure or temperature may affect quality before engineers make a production change. It is a useful way to compare scenarios, but it is only as credible as its underlying data and assumptions. Results still need confirmation on the actual process.
Likewise, algorithm-supported design of experiments (DOE) can help select informative trials across several variables. Rather than changing multiple settings informally, the team can test planned combinations and measure their effects on defined critical-to-quality characteristics.
The Lean Six Sigma principle remains the same: use evidence to select countermeasures, run a controlled pilot, and verify the result against the baseline. AI can accelerate analysis and experiment planning; it does not remove the need for operational safeguards or engineering approval.
Control: respond while the process can still be corrected
A predictive model can monitor live sensor signals and flag a rising risk. In the example plant, an operator might see: “Scrap probability 87% in the next three shots.” That message is a warning to investigate and respond, not a guarantee that the next three parts will be defective.
An AI-enhanced statistical process control (SPC) view can complement conventional control charts by scoring unusual patterns and combining signals that may be difficult to interpret one at a time. The chart should still communicate the process measure, limits, time sequence and response plan clearly. AI anomaly scoring complements SPC; it should not obscure the evidence or replace a defined reaction plan.
A practical escalation might be:
- Alert the operator when the risk score crosses an agreed threshold.
- Check the affected cavity, material lot, machine state and sensor readings.
- Apply an approved adjustment or contain the potentially affected parts.
- Record the action and verify the next shots against the defined quality criteria.
- Escalate to engineering if the signal persists or the cause is unclear.
This is predictive quality control in operational terms: identify a developing risk early enough for a person, or a validated, bounded control system, to act before more defects are produced.
Worked example: from 2.4% scrap to 0.7%
Consider the 14-press plant introduced above. The figures below are an illustrative worked example, not a claim about a published company deployment. They show how a team could structure a project and assess the result.
The Define phase sets the problem: scrap is 2.4%, the annualized scrap cost is $412,000, and defects are generally discovered at final inspection. The team agrees on operational definitions for each defect family and identifies critical characteristics for the product.
In Measure, it links each part’s inspection result to its press, mold cavity, shot and process data. Machine vision is introduced for selected visual and dimensional characteristics, while sensor data captures key molding conditions. The team checks measurement reliability and establishes the baseline.
In Analyse, the team examines 40+ process variables. It uses model outputs to prioritize factors for review, then tests those hypotheses with engineering knowledge and controlled analysis. The team finds a repeatable relationship between a particular barrel-temperature band and sink marks under specified material and mold conditions. It checks whether the pattern holds across shifts and cavities before treating it as actionable.
In Improve, engineers pilot adjusted settings and monitoring on a limited set of presses. They use planned trials to assess quality and cycle performance together, rather than improving one metric at the expense of another. Only after verification are the changes introduced more widely.
In Control, alerts are paired with a clear operator response, containment rules and a review of model performance. Machine vision checks parts in-line, and the team monitors the new scrap rate, OEE and customer escapes.
After six weeks, the illustrative results are:
| Measure | Baseline | After pilot |
|---|---|---|
| Scrap rate | 2.4% | 0.7% |
| Annualized scrap cost | $412,000 | $120,000 |
| OEE | 71% | 79% |
| Previously escaping defects detected by machine vision | Baseline inspection approach | 96% |
The scrap-rate change is a 1.7 percentage-point reduction, or about 71% relative to baseline. The annualized scrap-cost difference is $292,000. OEE improves by 8 percentage points. The vision result means the system identified 96% of the defined defects that had previously escaped to customers in the validated test set; teams should state the denominator and test conditions when reporting this metric.
There is also a real investment. At $18,000 per press for sensors, instrumentation across 14 presses totals $252,000, before integration, vision equipment or other implementation costs. Model training takes six weeks, and the plant hires one data engineer to support the work. The modeled annual scrap-cost reduction is encouraging, but it is not a complete ROI calculation: the business should include installation, support, labour, software and ongoing model validation before making an investment decision.
Designing a trustworthy predictive-quality system
Predictive models can provide useful signals, but the process and product context determine whether those signals can be trusted. Start with a bounded pilot and monitor performance by material, cavity, machine and shift, not just at plant level.
Three risks deserve particular attention:
- Models that do not generalize across resin types: A relationship learned on one resin or product may not hold for another. Label the material context and validate performance on the intended products.
- Data drift after maintenance or process changes: A mold overhaul, sensor replacement or machine update can change the data pattern. Track changes, check model performance, and revalidate when the process changes.
- Overfitting to one machine: Randomly splitting highly similar shots between training and test data can make a model appear more capable than it is. Validate on separate time periods, cavities or machines where appropriate.
Machine vision also needs representative examples, stable lighting, suitable optics and a clear plan for false rejects and missed defects. Agree on what happens when the model is uncertain. For consequential decisions, keep people in the loop until performance has been demonstrated under real production conditions.
From plant insight to daily improvement work with Ci Flow
The models exist in the plant; Ci Flow is where the insights land in your improvement workflow. Predictive alerts only create lasting value when teams can connect them to an accountable investigation, actions and verified results.
Ci Flow is a separate continuous improvement platform designed to support structured improvement work. Its project and portfolio workspace can help teams organise AI-assisted DMAIC work, record root-cause suggestions for practitioner review, track CI project pipelines and monitor project KPIs. Daily boards can keep emerging issues, owners and follow-up visible alongside routine management.
For a manufacturing project, the workflow could look like this:
- A sensor or vision alert highlights a potential quality issue.
- The team logs the problem and evidence in its improvement workflow.
- Practitioners use AI-assisted root-cause suggestions to generate hypotheses, then validate them with process data and the appropriate Lean Six Sigma tools.
- Actions, owners, milestones and project KPIs remain visible to the team and portfolio leaders.
- A daily board supports review of the response, verification and next steps.
Ci Flow supports project tracking and improvement governance; it is not a substitute for the plant’s machine-control or safety systems. Teams should confirm how their manufacturing data is securely shared, captured or referenced in their chosen workflow. Ci Flow’s security information explains its stated approach to protecting project information.
Build the capability behind the technology
AI can expand what manufacturing teams can see, but experienced practitioners turn signals into effective changes. They define the problem, check measurement quality, distinguish correlation from cause, design tests and establish controls that operators can sustain.
That is why the AI amplifies the Belt; it does not replace it. A CSSC-accredited Lean Six Sigma Green Belt course builds practical capability for leading data-driven improvement projects. For professionals leading complex initiatives and mentoring project teams, the Black Belt course develops advanced project leadership skills. Both support application through Lean Six Sigma methods and manufacturing-relevant case work.
Explore Ci Flow for AI-augmented continuous improvement in your plant, and build your capability with Lean 6 Sigma Hub’s CSSC-accredited Green Belt or Black Belt certification.
Kaizen. Kai-Care. Kai-Done. Lean Six Sigma
