Data-Driven Operational Excellence

Quality 4.0, predictive maintenance, digital twins, reinforcement-learning scheduling, process mining, and LSS4.0 integration.

4.10 Quality 4.0 and Machine-Learning-Augmented Process Control

4.10.1 Theoretical basis

Quality 4.0 denotes the alignment of quality management practice with Industry 4.0 capability. Sader, Husti and Daróczi (2022), reviewing 46 journal articles from 2017 to 2022, define it as an extended approach in which new technologies are combined with traditional quality activities — quality control, quality assurance and total quality management — thereby expanding the scope of quality management rather than replacing it. Their classification into four research themes (concept, implementation, quality management within Quality 4.0, and models and applications) reflects a field still consolidating its definitions. Antony and colleagues have pursued the conceptualisation from the practitioner side, and a broader review of the integration of Industry 4.0 with quality-related operational excellence methodologies situates Quality 4.0 within that wider movement.

The technical core, and the part with engineering substance, is the augmentation of statistical process control with machine learning. Classical control charts are optimal under strong assumptions — independent, identically distributed, typically normal observations, and a single characteristic per chart. Real high-volume processes violate all of these: observations are autocorrelated, characteristics are multivariate and correlated, and the relevant signal is often a pattern across many sensors rather than a shift in one. Machine learning methods relax those assumptions at the cost of interpretability and of a much heavier validation burden.

4.10.2 Mechanism

Three distinct mechanisms are commonly conflated and should be separated. Multivariate monitoring detects excursions in the joint distribution of many correlated variables that univariate charts miss. Predictive quality infers an unmeasured or expensive-to-measure quality characteristic from cheap process signals, functioning as a soft sensor and permitting 100 percent inferred inspection where physical 100 percent inspection is infeasible. Anomaly detection identifies process states unlike anything in the training distribution without requiring labelled defects — valuable precisely because rare failure modes are, by construction, under-represented in labelled data.

Only the second of these replaces an existing measurement; the first and third add observability. That distinction determines the validation burden, and it should be established before any algorithm is selected.

4.10.3 Evidence

Grade B for the concept as a research field, supported by multiple systematic reviews with defined corpora and by validated maturity-assessment frameworks. Grade C to D for specific production deployments: the applied literature is dominated by proof-of-concept studies on public or single-plant datasets, with performance reported on retrospective data rather than in prospective production use. Systematic reviews of machine-learning-based predictive quality in manufacturing confirm both the breadth of algorithmic experimentation and the scarcity of controlled prospective evaluation. Recent work examining industry perception of machine-learning-enhanced statistical process control finds practitioner acceptance to be conditional on interpretability and on the availability of evaluation criteria that quality professionals recognise — a finding that identifies the real adoption barrier as epistemic rather than technical.

The systematic reviews of Industry 4.0 and Six Sigma integration (Maia, Lizarelli and Gambi 2024, synthesising 59 articles from 2013 to 2021) and of Industry 4.0 with Lean Six Sigma more broadly (a review of 134 articles published between 2011 and mid-2022) both conclude that integration amplifies the effect of the classical methods rather than substituting for them.

4.10.4 Implementation protocol

1. Establish classical statistical process control first, with a correct rational subgrouping scheme and a demonstrated adequate measurement system. Machine learning applied to a process that is not in a state of statistical control models the special causes.

2. Build the data foundation: time-synchronised process traces, part-level or batch-level identity, and a reliable join to quality outcomes. Data engineering is typically 70 to 80 percent of the effort and should be planned as such.

3. Specify the decision the model will support and the cost asymmetry between false alarm and escape before selecting an algorithm. The threshold, not the architecture, determines operational value.

4. Prefer interpretable models where performance is comparable. In a regulated environment, a model whose behaviour can be explained to an inspector is worth a material performance concession.

5. Evaluate on a held-out future period. Random train-test splits leak information across process drift and systematically overstate performance.

6. Deploy in advisory mode and measure prospective performance against the offline estimate. Expect degradation; quantify it.

7. Establish model governance: version control, performance monitoring with defined alert thresholds, retraining triggers, change control on model updates, and a documented rollback path. Treat the model as validated software when it influences product disposition.

4.10.5 Worked example — ophthalmic

Automated optical inspection on a lens line produces a binary conforming or non-conforming call per lens. The line carries 34 process signals sampled at one hertz across dosing, cure, demould and hydration. Two applications are developed with quite different regulatory profiles.

Multivariate process monitoring (adds observability). A principal-component model on the 34 signals, with Hotelling’s T-squared and squared-prediction-error statistics charted against control limits, detects joint-distribution excursions. In one documented instance the model flags an excursion driven by a small correlated shift across three thermal signals — each individually within its univariate limits — which precedes a cosmetic reject cluster by roughly 25 minutes. Because the model triggers an engineering investigation rather than a product decision, it sits outside the validated control strategy and can be deployed with modest formality.

Predictive quality as a soft sensor (replaces a measurement). A gradient-boosted model predicting modulus from cure-profile and monomer-dosing signals is trained against destructive laboratory measurements. Deployed with disposition authority it would permit inferred 100 percent modulus verification in place of sampled destructive testing — a substantial economic gain. But it would then be part of the control strategy: software validation, model qualification against the destructive method across the design space, ongoing performance monitoring, and change control on every retraining. The recommended path is advisory deployment for a defined qualification period, with parallel destructive testing, and a formal decision on disposition authority only after prospective agreement is demonstrated.

The contrast between the two applications is the practical content of this section: the algorithm is the easy part, and the regulatory classification of the decision the algorithm supports determines the entire project shape.

4.10.6 Limitations and failure modes

• Modelling an out-of-control process. The most common technical error, and it produces models that encode the special causes as normal behaviour.

• Optimistic offline evaluation. Random splits, leakage through derived features, and evaluation on balanced samples all inflate reported performance relative to production.

• Interpretability deficit. A model that cannot be explained will not be accepted by the quality organisation and should not be, where it influences disposition.

• Silent drift. Without monitoring, degradation is invisible until a quality event exposes it.

• Technology-led project selection. Choosing the problem to fit an available algorithm rather than choosing the method to fit a quantified problem. The systematic reviews find integration amplifies classical methods — it does not replace the requirement to have a well-posed problem.

4.10.7 Primary metrics

Multivariate signal lead time (interval between model alert and conventional detection); alert precision in production; escape rate; proportion of quality decisions supported by predictive versus reactive information; model performance drift against the qualification baseline; time from process signal to operator action.

4.11 Predictive and Prescriptive Maintenance

4.11.1 Theoretical basis

Maintenance policy is a decision problem under uncertainty about remaining useful life. Run-to-failure is optimal only when failure is inconsequential and detection is immediate. Time-based preventive maintenance is optimal when the failure distribution has an increasing hazard rate and condition is unobservable. Condition-based and predictive maintenance dominate both when a measurable degradation signal exists and the cost of measurement is below the expected cost of mistimed intervention. Prescriptive maintenance extends the inference to a recommended action under operational constraints, jointly optimising intervention timing against production schedule and spares availability.

4.11.2 Mechanism

The mechanism is replacement of a population-level prior — the vendor interval, derived from a fleet failure distribution — with an asset-specific posterior conditioned on that asset’s observed degradation. The value is the reduction in two error types: intervening too early, which wastes remaining life and consumes production time, and intervening too late, which incurs failure cost. Where the failure distribution has high variance relative to its mean, the value of conditioning is large; where it is tight, time-based maintenance is nearly optimal and predictive maintenance adds cost without benefit. Computing the coefficient of variation of the failure distribution before investing is the single most useful piece of analysis in this domain, and it is almost never done.

4.11.3 Evidence

Grade B for effect direction and grade C for magnitude. Systematic reviews of machine-learning methods applied to predictive maintenance establish a large and methodologically diverse body of algorithmic work, with consistent reporting of reduced downtime, lower maintenance cost and improved productivity. The reviews are equally consistent about the limits of that evidence: data scarcity, class imbalance and the high cost of collecting genuine failure data are recurrent constraints, and quantitative guidance on implementation timelines and realised cost-benefit is limited, with substantial variation by organisational context. Multi-sector mapping studies confirm both the breadth of application and the heterogeneity of reported outcomes.

The practical reading is that predictive maintenance works where the preconditions hold — an observable degradation signal, sufficient failure history or a physics-based degradation model, and a failure distribution with enough variance to make conditioning worthwhile — and that vendor-quoted benefit ranges are not a substitute for asset-specific analysis.

4.11.4 Implementation protocol

1. Rank assets by criticality using the constraint analysis of 4.6 and the loss decomposition of 4.8. Predictive maintenance on a non-constraint with low failure consequence is a negative-return investment however good the model.

2. For each candidate asset, characterise the failure distribution from maintenance history. Compute mean and coefficient of variation by failure mode. Where the coefficient of variation is low, improve the time-based interval instead and stop.

3. Identify the degradation signal by failure mode: vibration, temperature, current signature, acoustic emission, pressure, or a process-derived proxy. Confirm that the signal has measurable lead time relative to functional failure — a signal that appears two minutes before failure has no operational value.

4. Instrument and collect through at least several failure cycles, or supplement scarce failure data with physics-based degradation models or accelerated testing.

5. Build remaining-useful-life models with explicit uncertainty quantification. A point estimate without a prediction interval cannot support a maintenance decision.

6. Integrate with planning: the output must be an action recommendation in the scheduling system, not an alert in a dashboard. This integration step, not the modelling, determines realised benefit.

7. In a validated environment, revise the preventive maintenance regime through change control, retaining a time-based backstop interval so that the predictive system relaxes rather than replaces the compliance baseline.

4.11.5 Worked example — ophthalmic

Blister sealing heads on a packaging line fail by platen wear, producing seal-integrity defects that are detected downstream at package inspection. Maintenance history over three years gives a mean time between platen-related events of 41 days with a standard deviation of 19 days — a coefficient of variation of 0.46, high enough that conditioning has clear value. The current preventive replacement interval of 21 days is set conservatively at roughly one standard deviation below the mean, which means most platens are replaced with substantial life remaining.

Instrumentation of seal force and platen temperature, joined to the seal-integrity inspection record, shows a monotonic rise in the ratio of peak-to-mean seal force as the platen surface degrades, with a usable lead time on the order of several days. A remaining-useful-life model on this ratio, with prediction intervals, permits replacement scheduling at a defined reliability level rather than at a fixed interval — extending mean interval materially while reducing seal-integrity escapes.

The regulatory treatment matters. Seal integrity is a validated critical characteristic under ISO 11607. The revised regime is therefore implemented as a condition-based reduction in intervention frequency within a retained maximum interval: the model may bring replacement forward, and may extend it up to a documented ceiling justified by the reliability analysis, but the ceiling remains a validated control. This structure — predictive scheduling inside a validated envelope — is the pattern that works in device manufacture, and it is worth adopting as a general design rule.

4.11.6 Limitations and failure modes

• Failure-data scarcity. Reliable equipment produces few failures, which is precisely what makes model building hard. Physics-based degradation models and accelerated testing are the standard responses; more sensors are not.

• Signal without lead time. A degradation signal must precede functional failure by enough time to act. Verify this empirically before instrumenting.

• Dashboard terminus. Predictions that do not enter the maintenance planning system produce no benefit. This is the most common organisational failure.

• Regulatory interval relaxation. Extending a validated preventive maintenance interval on the basis of a model requires documented justification; retaining a backstop ceiling is the defensible structure.

4.11.7 Primary metrics

Unplanned downtime hours on critical assets; ratio of planned to unplanned maintenance; mean time between failures by mode; prediction lead time and prediction interval coverage; remaining life consumed at replacement; maintenance cost per operating hour; failure-related quality escapes.

4.12 Digital Twin and Simulation-Based Optimisation

4.12.1 Theoretical basis

The term digital twin is used far more loosely than the literature warrants. Kritzinger, Karner, Traar, Henjes and Sihn (2018) supplied the discipline the field needed with a taxonomy based on the degree of data-integration automation: a digital model has no automated data exchange with the physical object; a digital shadow has automated one-way flow from physical to digital; and only a digital twin has automated bidirectional flow, so that a change in the digital object produces a change in the physical object and vice versa. On this criterion the overwhelming majority of industrial deployments marketed as digital twins are digital shadows.

This is not pedantry. The three levels differ in cost by roughly an order of magnitude each, and in capability qualitatively. A digital shadow supports monitoring and offline analysis. Only a true digital twin supports closed-loop control and autonomous optimisation. Specifying which level is being procured — and paying only for the level actually required — is the first and most consequential decision in any such programme.

4.12.2 Mechanism

The mechanism relevant to regulated manufacture is experimentation without physical change. Every physical process experiment in a device plant consumes material, production time and, where the process is validated, change-control effort. A sufficiently faithful simulation moves the search phase into a domain where those costs are near zero, and reserves physical experimentation for confirmation of a small number of candidate settings. This compresses the design-of-experiments programme of 4.4 and 4.5 by a large factor, and it is the single strongest argument for the investment in this sector.

A second mechanism is scenario evaluation for flow and capacity decisions — discrete-event simulation of scheduling policy, buffer sizing and capacity investment — which is mature, well-understood, and considerably cheaper than the sensor-integrated twin.

4.12.3 Evidence

Grade B for the conceptual and taxonomic literature, which is extensive and internally consistent, and grade C for operational performance effects, which are reported through case studies and structural-equation-modelling survey work rather than through controlled comparison. Systematic reviews consistently identify a gap between the conceptual richness of the field and the empirical evidence for operational benefit, and note that models should be systematically constructed from engineering and operational data and placed in a closed loop with the physical asset to realise the claimed capability. Discrete-event simulation for capacity and scheduling decisions, by contrast, has a long and solid validation record and should not be conflated with the digital twin literature.

4.12.4 Implementation protocol

1. State the decision the twin will support before specifying it. “Reduce the number of physical trials required to characterise a new lens geometry” is a specification; “create a digital twin of the plant” is not.

2. Select the Kritzinger level required by that decision and pay for no more. Most process-development use cases are satisfied by a digital model plus periodic recalibration.

3. Build from first principles where the physics is tractable — thermal, flow, mechanical — and use data-driven surrogates only where it is not. Physics-based models extrapolate; purely empirical surrogates do not, and extrapolation is precisely what design-space exploration requires.

4. Validate the model against held-out physical data across the intended operating region, and quantify prediction error explicitly. State the region of validity and refuse to use the model outside it.

5. Establish the synchronisation mechanism and its latency budget if a shadow or twin level is required.

6. Integrate into the improvement workflow: the twin must be the default first step in experimental design, or it will become an unused asset. This is an organisational change, not a technical one.

7. For any twin used to support a regulatory submission or validation rationale, apply computer software assurance principles and document the model qualification, its region of validity, and its maintenance regime.

4.12.5 Worked example — ophthalmic

A physics-based thermal and cure model of the moulding station is constructed from tool geometry, material thermal properties and measured boundary conditions, and validated against thermocouple and dimensional data across the operating region. Prediction error on final base-curve radius is characterised and reported with its region of validity.

The model is then used for in-silico screening in the design-space programme of 4.5. Where the physical central composite design over four factors would have required a substantial number of physical runs across two tool sets, in-silico screening over a much larger factor set identifies the region of interest first, and physical experimentation is reserved for confirmation at the region boundaries and centre. The reduction in physical runs translates directly into reduced material consumption, reduced production interruption, and — the dominant term in a device plant — reduced protocol and report volume.

Separately, a discrete-event model of the moulding-to-hydration-to-inspection-to-sterilisation chain is used to evaluate the buffer sizing and campaign-scheduling decisions of 4.6 before implementation. This is a digital model in the Kritzinger sense, requires no sensor integration, costs a small fraction of the twin, and answers the question that was actually asked. Recognising when this is sufficient is worth more than the twin.

4.12.6 Limitations and failure modes

• Level inflation. Procuring a twin when a model suffices. The taxonomy exists precisely to prevent this and should be used contractually.

• Unstated validity region. A model used outside the region in which it was validated is a confident source of wrong answers, and its confidence is the danger.

• Maintenance decay. Models drift out of correspondence as the process changes. Without a revalidation cadence, trust erodes and the asset is abandoned.

• Orphaned asset. Built by a technical group, never integrated into the engineering workflow, quietly retired at the next budget cycle. Integration must be designed in from the outset.

4.12.7 Primary metrics

Physical experimental runs avoided per characterisation programme; model prediction error against held-out data, tracked over time; proportion of process changes preceded by simulation; time from question to answer for capacity and scheduling decisions; model revalidation currency.

4.13 Reinforcement Learning for Production Scheduling

4.13.1 Theoretical basis

Production scheduling in realistic settings — dynamic order arrival, machine breakdown, sequence-dependent setup, multiple objectives — is computationally intractable to solve optimally and is conventionally handled with dispatching rules or with metaheuristics that must be re-run when conditions change. Deep reinforcement learning offers a different formulation: learn a scheduling policy offline from simulated experience, then apply it in real time at negligible computational cost. The policy generalises across states, which is what dispatching rules cannot do and what re-optimisation achieves only at high computational expense.

4.13.2 Mechanism

The policy amortises the optimisation cost. Training is expensive and offline; inference is a single forward pass and is effectively free. This makes genuinely real-time rescheduling feasible in environments where order insertion, breakdown and rework make static schedules obsolete within hours. The approach also permits multi-objective policies — throughput against due-date performance against setup cost — learned from a scalarised or vector reward rather than encoded in hand-written priority rules.

4.13.3 Evidence

Grade D, and this grade should be respected. The research literature is large and growing rapidly, with demonstrated results on job-shop, flow-shop and mixed-model formulations, and with recent work addressing large-scale mixed-model production and dynamic order insertion. A comparative survey published in the Journal of Intelligent Manufacturing addresses reliability of reinforcement-learning-based production scheduling systems directly and identifies the concerns that matter for industrial adoption: policy behaviour under distribution shift, absence of feasibility guarantees, and the difficulty of verification. Scaling to large mixed-model problems remains constrained by training cost and convergence behaviour, and the field has begun to produce frameworks explicitly targeting real-world constraints — an indication that the gap between benchmark performance and deployable systems is recognised within the field itself.

Almost all published evaluation is in simulation. Industrial deployments are recent, few, and concentrated in semiconductor fabrication, where scheduling complexity is extreme and the economic stakes justify the technical risk. For a device plant this method belongs in the evaluate-and-monitor category, not the deploy category, unless scheduling complexity is genuinely exceptional.

4.13.4 Implementation protocol (evaluation-stage)

1. Establish the benchmark first: measure the performance of the current scheduling approach and of well-tuned dispatching rules. A substantial fraction of reported reinforcement-learning gains are gains over a poorly configured baseline.

2. Build a validated discrete-event simulation of the scheduling environment. This is a prerequisite, it is expensive, and it is independently valuable — the digital model of 4.12.

3. Define the reward function with explicit weights over throughput, due-date performance, setup cost and any constraint penalties. Reward specification is where most of the design effort and most of the errors are.

4. Train and evaluate against the dispatching-rule benchmark under distribution shift — demand mix changes, breakdown scenarios, order surges — not only under training-distribution conditions.

5. Deploy in advisory mode with human override, and instrument override frequency and reason. Override data is the primary evidence about policy trustworthiness.

6. Impose hard feasibility constraints outside the learned policy. Constraints that must never be violated — segregation, expiry, campaign sequencing for cross-contamination control — belong in a rule layer that the policy cannot override, not in the reward function.

4.13.5 Worked example — ophthalmic

A specialty lens facility schedules a surfacing and coating shop with roughly 400 active order lines, sequence-dependent setup driven by prescription geometry and tint, dynamic order insertion from clinical and rush channels, and a hard constraint that certain tint sequences cannot follow one another without an intervening clean.

The evaluation is structured as follows. A discrete-event model is built and validated against six months of historical order flow. Three approaches are compared under identical order streams: the incumbent rule set, a tuned composite dispatching rule, and a trained deep reinforcement-learning policy. The tuned dispatching rule alone recovers a meaningful share of the gap to the simulated optimum — a result that recurs across the literature and that frequently ends the business case honestly and early. The learned policy outperforms the tuned rule primarily under order-surge conditions, which is precisely the regime where the incumbent system performs worst and where the operational value is concentrated.

The tint-sequence constraint is implemented as a hard filter on the action space rather than as a reward penalty, so that the policy cannot learn to trade a cross-contamination risk against a throughput gain. In a regulated environment this architectural choice is not optional, and it should be a stated requirement in any procurement.

4.13.6 Limitations and failure modes

• Weak baseline comparison. The most common overstatement in the literature and in vendor claims. Always tune the dispatching rule first.

• Distribution shift. Policies trained on one demand regime degrade under another, often without a visible signal.

• Absence of feasibility guarantees. Learned policies offer no constraint guarantees; hard constraints must be enforced architecturally.

• Verification difficulty. Explaining to an auditor why a policy sequenced a batch as it did is an unsolved problem, which is a strong argument for the hard-constraint layer and for advisory deployment.

• Simulation dependence. Policy quality is bounded by simulation fidelity; a poor simulation yields a policy optimised for a fiction.

4.13.7 Primary metrics

Schedule adherence and due-date performance against the tuned-rule benchmark; setup time as a proportion of available time; throughput under surge scenarios; human override rate and coded reason; constraint violations (which must be identically zero); rescheduling latency.

4.14 Process Mining and Conformance Checking

4.14.1 Theoretical basis

Process mining, established as a discipline through the work of van der Aalst, extracts process knowledge from the event logs that information systems already record. Its three classical use cases are discovery (infer the actual process model from the log, with no a priori model), conformance checking (compare an executed log against a prescribed model and quantify deviation), and enhancement (extend the model with performance data such as waiting times, rework loops and bottlenecks). The field sits at the intersection of data science and process science and has matured from workflow analysis into a general instrument for operational diagnosis.

Its distinctive epistemic property is that it measures what the process did, not what participants believe it does. Value stream mapping, by contrast, records a socially negotiated account of the process, and the two frequently differ in ways that matter. Where event data exist, process mining should precede interview-based mapping.

4.14.2 Mechanism

Two mechanisms. First, rework and loop discovery: event logs expose repeat visits to the same activity — deviation reopened, batch record returned for correction, inspection re-run — that are structurally invisible in a value stream map drawn as a linear chain. In administrative processes these loops routinely account for the majority of elapsed time. Second, variant analysis: the distribution of distinct execution paths, which typically shows that a small number of variants covers most cases while a long tail of exception paths consumes a disproportionate share of time and attention. Targeting the tail is usually the highest-yield intervention and is not identifiable without the log.

4.14.3 Evidence

Grade B. The methodological foundation is well established, and application in manufacturing is supported by a growing set of studies, including systematic reviews of process mining applied to Industry 4.0 contexts and documented industrial applications such as productivity improvement in make-to-stock manufacturing published in the International Journal of Production Research. There is empirical evidence that process mining identifies improvement potential in high-volume, high-mix production, and practical accounts describe manufacturers retrofitting automated data capture onto existing lines to enable it. The field acknowledges a genuine technical limitation in manufacturing: analysing flows in which multiple components converge into an assembly requires merging case identifiers, which is non-trivial and remains an active research problem.

4.14.4 Implementation protocol

1. Select a process with existing event data. In device plants the highest-yield candidates are administrative rather than physical: deviation and non-conformance management, change control, complaint handling, batch record review, supplier corrective action. These are the processes that constrain improvement throughput.

2. Extract the log in case-activity-timestamp form and assess data quality explicitly — missing timestamps, activity granularity mismatch, case identifier integrity.

3. Discover the as-is model and compare it with the documented procedure. The gap between the two is, in a regulated environment, immediately actionable in both directions: it identifies inefficiency and it identifies procedure-adherence exposure.

4. Run variant analysis. Quantify what proportion of cases follow the dominant variants and what proportion of total elapsed time is consumed by the tail.

5. Enhance with performance data: waiting time by activity and by handover, rework loop frequency, resource-dependent throughput.

6. Target the loops and the handovers, which is where administrative time accumulates, rather than the activity durations, which are usually small.

7. Institutionalise conformance checking as a periodic control. Persistent, quantified conformance monitoring is a stronger position at inspection than periodic internal audit sampling.

4.14.5 Worked example — ophthalmic

The deviation management process is mined from twenty-two months of quality-system event data covering 2,847 deviations. The documented procedure describes eight activities in sequence. Discovery finds 214 distinct execution variants; the four most common cover 61 percent of cases, and the remaining 210 variants cover 39 percent of cases but 68 percent of total elapsed time.

Loop analysis identifies the dominant pathology: 34 percent of deviations traverse the investigation-review loop more than once, with a mean of 2.7 traversals among those that loop, and a mean added elapsed time of 11.4 days per additional traversal. Conformance checking attributes the majority of these loops to a single cause — investigation packages returned for insufficient root-cause depth. The intervention is therefore not a workflow redesign but a competence and template intervention at the investigation-authoring step, together with a pre-submission checklist derived from the actual return reasons.

Two consequences follow, and both matter. Deviation cycle time falls, which directly increases the plant’s change-control throughput and therefore raises the ceiling on improvement velocity established in Section 2.2. And the conformance analysis itself becomes standing evidence of process control over a quality-system process — an asset at inspection under the QMSR inspection programme. Very few methods in this review produce a benefit that is simultaneously operational and inspectional.

4.14.6 Limitations and failure modes

• Log quality. Timestamps recorded at data entry rather than at event occurrence produce systematically wrong waiting times. Assess this before analysing.

• Activity granularity mismatch. Logs record system transactions, which may not correspond to the activities of interest.

• Convergent flows. Assembly and kitting processes require case-identifier merging, which is a known open problem.

• Analysis without intervention. Process mining produces compelling visualisations very quickly, which makes it unusually prone to becoming an analytical exercise that terminates in a presentation.

4.14.7 Primary metrics

Case cycle time distribution (median and 90th percentile, not mean); variant count and concentration; rework loop frequency and cost in elapsed days; conformance rate against the prescribed model; waiting time by handover; deviation and change-control throughput.

4.15 Integration Frameworks: LSS4.0 and DMAIC 4.0

4.15.1 Theoretical basis

The integration of Industry 4.0 technologies with Lean Six Sigma has been the most active research theme in operational excellence over the past five years, and until recently it lacked a structured implementation guideline. Skalli, Cherrafi, Charkaoui, Chiarini, Shokri, Antony, Garza-Reyes and Foster (2025) address this directly, developing an LSS4.0 framework through design science research with an associated fourteen-step implementation process, evaluated through action-research-based interviews and demonstrations with case studies drawn from the automotive and mining industries. A parallel line of work applies the same design-science approach at the level of the improvement cycle itself, producing a DMAIC 4.0 framework and roadmap that maps specific Industry 4.0 technologies onto each DMAIC phase.

The underlying empirical justification comes from the systematic review literature. Maia, Lizarelli and Gambi (2024), synthesising 59 articles from 2013 to 2021, examine the relationships between Six Sigma and Industry 4.0 technologies and the benefits of their integration. A separate review of 134 articles published between 2011 and mid-2022 examines the impact of Industry 4.0 technologies on Lean Six Sigma and the implications for operational excellence. Both converge on the same conclusion: the technologies amplify the classical methods rather than displacing them, and the integration is most valuable in the measure and analyse phases, where data availability is the historical constraint.

4.15.2 Mechanism

The integration attacks the phase-level bottlenecks of the classical method. In define, process mining and connected quality data replace opinion-based problem selection. In measure, automated data capture removes the manual data-collection phase that historically consumed a large share of project elapsed time, and makes population data available in place of samples. In analyse, machine learning handles the high-dimensional, nonlinear and interaction-rich structures that classical stratification handles poorly. In improve, digital twins permit in-silico experimentation. In control, real-time monitoring and automated response replace periodic charting.

The mechanism is therefore reduction of project cycle time and increase of analytical reach — not replacement of the DMAIC logic, which the frameworks retain deliberately.

4.15.3 Evidence

Grade B. Design-science artefacts are evaluated by expert assessment and demonstration rather than by controlled trial, which bounds the strength of the claim; the underlying systematic reviews, however, rest on substantial and independently assembled corpora. The evidence supports the proposition that firms combining the approaches report better outcomes than firms adopting either alone, while noting the standard confound: firms that adopt both are systematically different from firms that adopt one.

4.15.4 Implementation protocol

1. Assess classical Lean Six Sigma maturity honestly before integrating. Organisations without a functioning DMAIC capability do not obtain a functioning LSS4.0 capability by adding technology; they obtain expensive dashboards.

2. Map the phase-level bottlenecks in the existing project portfolio. Measure actual elapsed time by DMAIC phase across the last twenty projects. Direct technology investment at the measured bottleneck, which in device plants is frequently the improve and control phases owing to validation load, not the measure phase that the generic frameworks assume.

3. Select technology per phase against that measurement, not against a reference architecture.

4. Build the data layer once, as shared infrastructure across projects, rather than per project.

5. Retrain practitioners in the augmented method: the analytical skill set for machine-learning-assisted analysis differs materially from classical Six Sigma training, and the gap is usually underestimated.

6. Preserve the control-phase discipline. The most frequent degradation in technology-augmented improvement is the substitution of monitoring for control — a dashboard that displays a deviation is not a mechanism that prevents it.

4.15.5 Worked example — ophthalmic

The plant measures elapsed time by phase across its last twenty Lean Six Sigma projects: define 3.1 weeks, measure 6.8, analyse 4.2, improve 11.9, control 5.4 — a mean of 31.4 weeks, consistent with the nine-month figure McGrane and colleagues (2022) report for a medical device setting. The improve phase dominates, and decomposition attributes 7.1 of its 11.9 weeks to validation protocol authoring, execution and approval.

A generic LSS4.0 deployment would target the measure phase with automated data capture, saving perhaps three weeks of a 31.4-week cycle. The measured bottleneck argues for a different allocation: digital-twin-based in-silico experimentation to reduce the number of physical confirmation runs, and process mining on the change-control process itself to attack the approval cycle. The plant deploys automated data capture as well, because it is cheap and it is a prerequisite for the other two, but it does so knowing the expected contribution.

The generalisable point is the protocol step: measure your own phase-level bottleneck before adopting a reference framework built on someone else’s.

4.15.6 Limitations and failure modes

• Integration without foundation. Adding Industry 4.0 technology to an organisation that cannot execute DMAIC produces neither capability.

• Reference-architecture adoption. Frameworks derived from automotive and mining case studies embody those sectors’ bottlenecks. Device plants have a different one.

• Monitoring substituted for control. The characteristic failure of digitally augmented improvement.

• Skill gap. Practitioners trained in classical statistical methods require substantial additional development to work competently with machine-learning-assisted analysis, and the intermediate state — using methods one does not understand — is worse than either endpoint.

4.15.7 Primary metrics

Project cycle time by DMAIC phase; projects completed per practitioner per year; benefit per project; proportion of projects using automated rather than manual data collection; data-layer reuse across projects; practitioner capability assessment against the augmented method.